Prediction system based on point cloud data and spatiotemporal frequency domain joint attention mechanism

Abstract

This invention relates to the field of short-term meteorological forecasting technology, specifically to a forecasting model architecture based on point cloud data and a spatiotemporal frequency domain joint attention mechanism. This forecasting model architecture includes a Z-coordinate system three-dimensional point cloud visualization data model and a P-coordinate system four-dimensional spatiotemporal frequency domain joint attention model. The three-dimensional point cloud visualization data model directly represents the three-dimensional distribution of meteorological elements. Combined with the DEM trust propagation completion algorithm, the observation coverage of a 1km resolution area is increased by 40%, significantly improving the forecast accuracy in complex terrain areas (such as mountainous regions). Then, a decision tree model and satellite observation data are used to fuse precipitation data, filling in the gaps in precipitation observation data for areas without monitoring stations. Finally, based on the Earth's rotation period and P-coordinate system four-dimensional spatiotemporal time domain data, a Fourier transform + phase modulation module enhances the frequency domain capture capability of precipitation cycle signals, significantly improving the TS score for severe convective weather.

Description

Technical Field

[0001] This invention relates to the field of short-term meteorological forecasting technology, and more specifically, to a forecasting system based on point cloud data and a spatiotemporal frequency domain joint attention mechanism. Background Technology

[0002] There are three major bottlenecks in existing short-term meteorological forecasting technology: (1) Data representation limitations: Traditional gridded models are difficult to accurately depict the spatial distribution of three-dimensional meteorological fields under complex terrain, and blind spots in station observations lead to insufficient fusion accuracy; (2) Spatiotemporal modeling defects: Mainstream CNN / RNN models have weak ability to extract frequency domain features of periodic weather (such as precipitation wave groups), and have not effectively integrated the correlation between geographical static features and meteorological dynamic elements; (3) Insufficient real-time performance: High-precision models (such as WRF) have large computational overhead, while lightweight models (such as radar extrapolation) are difficult to support high-resolution operational deployment at the 1km / 10-minute level.

[0003] Although AI models (such as NowcastNet) attempt to fuse multi-source data, they are still limited by the singular representation of grid data structures and the lack of physically constrained frequency domain optimization mechanisms. Therefore, there is an urgent need to develop new forecasting systems to overcome the challenge of balancing spatiotemporal resolution and computational efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a prediction system based on point cloud data and a spatiotemporal frequency domain joint attention mechanism to improve the above-mentioned defects.

[0005] To achieve the above objectives, the embodiments of this application provide the following technical solutions:

[0006] This application provides a prediction system based on point cloud data and a spatiotemporal frequency domain joint attention mechanism, the system comprising:

[0007] The point cloud data processing module is used to convert multi-source meteorological data from ground automatic stations and satellite remote sensing into point cloud format, fill in observation blind spots through spatial interpolation compensation algorithms, and generate station-by-station time series.

[0008] The spatiotemporal frequency domain joint attention module is used to decompose the spatiotemporal data of meteorological elements into the frequency domain through Fourier transform. Then, the phase modulation module is used to enhance the frequency domain attention of periodic weather characteristics. Combined with the multi-head self-attention mechanism, the spatiotemporal dependence of geographic information and meteorological elements is integrated.

[0009] The lightweight inference module reduces the model parameter precision from Float32 to Float16, thereby reducing storage overhead and dynamically releasing GPU intermediate tensor memory to avoid resource overflow.

[0010] Optionally, the point cloud data processing module fills in the observation blind spots using a spatial interpolation compensation algorithm, including:

[0011] Based on the DEM digital elevation and KNN analysis results from neighboring stations, the trust propagation algorithm is used to iteratively correct missing values.

[0012] By aligning satellite grid data through a time sliding window, the brightness temperature difference in the infrared band and the RGB channel features are extracted.

[0013] Optionally, the phase modulation module enhances the frequency domain attention to periodic weather features, including:

[0014] The phase modulation module applies dynamic weights to the frequency domain components after Fourier transform, with the formula: Output = FFT(X)⋅σ(W⋅Phase(X)+b), where X is the input spatiotemporal data tensor, FFT(X) is the Fast Fourier Transform performed on the spatiotemporal dimension of X, Phase(X) is the phase component extracted from FFT(X), W is the weight matrix, b is the bias vector, (W⋅Phase(X)+b) is the linear transformation of the phase component, σ(W⋅Phase(X)+b) is the activation function, and Output is the weighted frequency domain feature.

[0015] Optionally, the lightweight inference module deploys a multi-level fault switching mechanism: when the main inference path fails, it switches to RBF kernel function interpolation forecast; in case of extreme anomalies, it falls back to historical climate state prediction values.

[0016] The missing values ​​are iteratively corrected using the trust propagation algorithm based on the DEM digital elevation and neighboring station KNN analysis results, including:

[0017] Perform time-dimensional exponential sliding imputation on missing data from ground-based automatic weather stations, using the following formula:

[0018]

[0019] in, The padding value at time t, These are the observations from the previous 10 minutes. ;

[0020] For spatially missing data, perform trust propagation iterative imputation with DEM constraints. The imputation method is as follows:

[0021] ;

[0022] in The difference in altitude between stations This represents the vertical temperature lapse rate. The KNN weights represent the spatial weight of station j relative to station i, and are calculated using the inverse of the distance between the two points or a Gaussian kernel function. This is the fill-in value for station i in the (K+1)th iteration, used to fill in the missing meteorological elements (such as temperature and precipitation) in the final estimate, and thus correct the reasonable forecast value after terrain correction. Let i be the set of neighboring sites. The normalized denominator, representing the weights, ensures that the output value remains within a reasonable physical range. Let j be the value of the neighboring station j in the kth iteration.

[0023] The beneficial effects of this invention are as follows:

[0024] The forecasting system based on point cloud data and spatiotemporal frequency domain joint attention mechanism described in this invention directly represents the three-dimensional distribution of meteorological elements through point cloud data. Combined with the DEM trust propagation completion algorithm, it increases the observation coverage of a 1km resolution area by 40%, significantly improving the forecast accuracy of complex terrain areas (such as mountainous areas).

[0025] Secondly, by enhancing the frequency domain capture capability of precipitation cycle signals through Fourier transform and phase modulation modules, the system's TS score for severe convective weather is significantly improved. Furthermore, the multi-head attention mechanism dynamically weights geographic static features (elevation, land use) and meteorological dynamic fields, which significantly reduces the RMSE of temperature forecasts in this system.

[0026] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and claims. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention shown herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0028] Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0029] Example 1:

[0030] This embodiment provides a prediction system based on point cloud data and a spatiotemporal frequency domain joint attention mechanism. The system includes:

[0031] The point cloud data processing module is used to convert multi-source meteorological data from ground automatic stations and satellite remote sensing into point cloud format, fill in observation blind spots through spatial interpolation compensation algorithms, and generate station-by-station time series.

[0032] The point cloud data processing module fills in the observation blind spots using a spatial interpolation compensation algorithm, including:

[0033] Step S1: Based on the DEM digital elevation and the KNN analysis results of neighboring stations, the missing values ​​are iteratively corrected using the trust propagation algorithm;

[0034] Step S2: Align satellite grid data through a time sliding window and extract infrared band brightness temperature difference and RGB channel features.

[0035] The spatiotemporal frequency domain joint attention module is used to decompose the spatiotemporal data of meteorological elements into the frequency domain through Fourier transform. Then, a phase modulation module is used to enhance the frequency domain attention of periodic weather characteristics. Combined with a multi-head self-attention mechanism, the spatiotemporal dependency between geographic information and meteorological elements is integrated. Secondly, the lightweight inference module deploys a multi-level fault switching mechanism: when the main inference path fails, it switches to RBF kernel function interpolation forecast; in case of extreme anomalies, it falls back to historical climate state prediction values.

[0036] The lightweight inference module is used to reduce the model parameter precision from Float32 to Float16, thereby reducing storage overhead and dynamically releasing GPU intermediate tensor memory to avoid resource overflow.

[0037] Secondly, the lightweight inference module deploys a multi-level fault-to-connection mechanism:

[0038] When the main inference path fails, the system switches to RBF kernel function interpolation prediction. The specific implementation can be as follows:

[0039] Step S31: Monitor the operation status and output quality of the main weather forecast inference path in real time. The monitoring includes, but is not limited to, the survival status of computing nodes, the timeliness of input data, the confidence index of output results, the spatial / temporal self-consistency of forecast results, and the computing resource consumption index. Based on the preset multi-dimensional fault judgment rule set, dynamically evaluate the fault level of the main inference path. The fault level includes at least normal, warning, partial failure, and complete failure.

[0040] The confidence index can be calculated as follows:

[0041] Step S311: Run multiple lightweight simplified models in parallel to quickly predict the same input data, calculate the average similarity between the main inference path output and the simplified model output, and trigger a low confidence alarm when the similarity is lower than the threshold.

[0042] Step S312: Check whether the physical quantities (such as temperature, air pressure, and humidity) output by the main inference path meet the preset atmospheric physical constraint equations, calculate the number and degree of variables that violate the constraints, and mark them as low confidence when they exceed the tolerance range.

[0043] Step S313: Perform similarity matching between the current input data and the historical database. If the prediction error of the main model in similar historical cases is significantly higher than that of the backup model, then reduce the confidence score of the current output.

[0044] Secondly, the specific calculation method for the aforementioned spatial / temporal self-consistency can be as follows:

[0045] Step S314: Divide the forecast area into grids and calculate the gradient values ​​of the same type of meteorological variables (such as precipitation) in adjacent grid cells;

[0046] Step S315: If there are more than 5 consecutive grid cells whose gradient values ​​exceed 200% of the maximum gradient threshold in climatological history, then it is determined to be spatially inconsistent.

[0047] The method for calculating the time self-consistency is as follows:

[0048] Step S316: Perform sliding window analysis on the results of three consecutive forecast times for the same geographical location, and calculate the variable difference between the later time and the earlier time.

[0049] If the direction of the difference value (e.g., heating / cooling) is opposite to the trend of the historical observation data of the previous 24 hours for three consecutive times, it is determined to be time inconsistency.

[0050] Step S32: During the period when the main inference path is in a normal state, continuously collect the input data samples of the main inference path and their corresponding forecast output results; based on the collected data samples and output results, dynamically train or optimize the key parameters of the backup RBF kernel function interpolation forecast model, the key parameters including the shape parameters of the radial basis function, the center point selection strategy, and the interpolation weight coefficients.

[0051] Step S33: When step S31 determines that the main inference path has reached the preset "partial failure" or "complete failure" fault level, the weather forecast calculation task is immediately and seamlessly switched to the standby RBF kernel function interpolation forecast path configured with dynamic generation parameters in step S32.

[0052] Step S34: After the fault level of the main inference path is restored to "normal", the main inference path is not immediately switched back. Instead, the main inference path and the backup RBF kernel function interpolation prediction path are run in parallel within a preset verification time window. The prediction results and operational stability of the two are compared within the verification window. Only when the prediction results of the main inference path meet the preset accuracy and stability requirements and its operational status remains stable, the computation task is gradually or all at once switched back to the main inference path. This step introduces a parallel verification period and stability / accuracy requirements to avoid secondary faults or prediction quality fluctuations caused by immediately switching back when the main path may not be stable after recovery.

[0053] Step S35: Record information for each fault switching event, including the fault level that triggered the switching, specific indicators, switching time, predicted performance data of the backup path, and the verification results of the main path recovery; periodically analyze the recorded event information, and automatically adjust the threshold parameters in the multi-dimensional fault judgment rule set, the dynamic parameter generation strategy of the backup RBF model, or the verification conditions for intelligent back-switch based on the analysis results.

[0054] Among them, the phase modulation module enhances the frequency domain attention of periodic weather features, including:

[0055] The phase modulation module applies dynamic weights to the frequency domain components after Fourier transform, with the formula: Output = FFT(X)⋅σ(W⋅Phase(X)+b);

[0056] Where X is the input spatiotemporal data tensor, specifically: time step, number of spatial grids / stations, and raw data of meteorological elements (such as temperature field, precipitation field, etc.); FFT(X) performs a Fast Fourier Transform on the spatiotemporal dimension of X to convert meteorological data from the time domain to the frequency domain, separating weather signals of different frequencies (such as 2-hour periodic precipitation waves); Phase(X) extracts the phase component of FFT(X), W is the weight matrix, used to dynamically adjust the importance of different frequency bands (such as assigning higher weights to key frequency bands of strong convection), and adaptively enhances the focus on frequency bands sensitive to severe weather through training; b is the bias vector, which compensates for baseline deviations in frequency domain features (such as background noise caused by terrain); (W⋅Phase(X)+b) is a linear transformation of the phase component, generating frequency domain attention weights (e.g., amplifying phase change signals related to rainstorms); σ(W⋅Phase(X)+b) is the activation function, which compresses the weights to [0,1]. The range represents the retention probability of frequency domain components, enabling nonlinear filtering and suppressing noise frequency bands; the output is the weighted frequency domain features. The enhanced frequency domain representation can significantly strengthen the 30-60 minute precipitation cycle signal and can be directly input into downstream networks (such as CNN / LSTM) for forecasting.

[0057] The forecasting system based on point cloud data and a spatiotemporal frequency domain joint attention mechanism described in this embodiment transforms meteorological spatiotemporal data (such as precipitation fields) from the time domain to the frequency domain, separating weather signals of different frequencies (such as precipitation wave groups with a 30-minute cycle). At the same time, a phase modulation module applies dynamic weights to the frequency domain components, enhancing the attention weights on key frequency bands of strong convection (such as the frequencies corresponding to the life cycle of mesoscale convective cloud clusters). This frequency domain enhancement enables the system to accurately capture periodic convective signals (such as the thunderstorm formation and dissipation cycle), thereby identifying the initial location of strong convection in advance. In tests conducted by the provincial meteorological bureau, the hit rate (POD) for 100 historical strong convective events increased from 68% to 83%, the false alarm rate (FAR) decreased from 35% to 24%, and the TS score directly increased from 0.52 to 0.63.

[0058] Secondly, in this embodiment, a multi-head self-attention mechanism generates a dynamic weight matrix, which associates three types of features, including static geographical features (altitude, land use type), dynamic meteorological elements (temperature time series changes, wind field of nearby stations), and spatiotemporal dependencies (such as temperature inversion layer in mountainous areas, urban heat island effect). By calculating the correlation between features through cross-feature attention layers (e.g., the transmission weight of temperature changes at high-altitude stations to low-altitude areas), traditional models in complex terrain areas suffer from an RMSE of 2.5℃ due to neglecting the altitude-temperature lapse rate (0.6℃ / 100m). However, the forecasting system described in this embodiment dynamically adjusts the weights through the attention mechanism (e.g., increasing the weight of mountainous stations by 30%), reducing the temperature forecast error of Nanjing Zijin Mountain (altitude 448m) from 2.2℃ to 1.4℃ (average RMSE of the test dataset: from 1.8℃ to 1.0℃).

[0059] Secondly, based on the DEM digital elevation and neighboring station KNN analysis results, the missing values ​​are iteratively corrected using the trust propagation algorithm, including:

[0060] Perform time-dimensional exponential sliding imputation on missing data from ground-based automatic weather stations, using the following formula:

[0061] ;

[0062] in, The padding value at time t, These are the observations from the previous 10 minutes. ;

[0063] For spatially missing data, perform trust propagation iterative imputation with DEM constraints. The imputation method is as follows:

[0064] ;

[0065] in The difference in altitude between stations This represents the vertical temperature lapse rate. The KNN weights represent the spatial weight of station j relative to station i, and are calculated using the inverse of the distance between the two points or a Gaussian kernel function. This is the fill-in value for station i in the (K+1)th iteration, used to fill in the missing meteorological elements (such as temperature and precipitation) in the final estimate, and thus correct the reasonable forecast value after terrain correction. Let i be the set of neighboring sites. The normalized denominator, representing the weights, ensures that the output value remains within a reasonable physical range. Let j be the value of the neighboring station j in the kth iteration.

[0066] Example 2: Based on Example 1, this example uses severe convective weather forecasting in mountainous areas as an example to briefly illustrate the linkage between the point cloud data processing module and the spatiotemporal frequency domain joint attention module:

[0067] Point cloud data processing module:

[0068] Missing precipitation data for stations at an altitude of 1200m were supplemented based on DEM (trust propagation algorithm).

[0069] Construct the spatial topology of the stations (Station A → Station B: elevation difference 300m, distance 5km).

[0070] Spatiotemporal frequency domain joint attention module:

[0071] Fast Fourier Transform (FFT(X)) revealed a 45-minute periodic wave in the precipitation sequence;

[0072] The phase modulation module enhances this frequency band (e.g., weight = 0.92) because the phase difference indicates that the cloud cluster is moving towards the valley;

[0073] The attention mechanism integrates altitude difference (static) and precipitation wave (dynamic) to output heavy rain warnings for valley areas.

[0074] The point cloud data processing module is the physical representation basis of the three-dimensional meteorological field, and the spatiotemporal frequency domain joint attention module is the intelligent extraction engine for multi-scale features.

[0075] The point cloud data processing module provides structured input + spatial constraints (terrain / proximity relationship) → frequency domain attention to achieve physically interpretable feature enhancement; the two are seamlessly connected through data format standardization (NetCDF point cloud → spatiotemporal tensor) and feature sharing mechanism (geographic static feature injection attention) to jointly solve the core problem of "high-resolution forecasting under complex terrain".

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A prediction system based on point cloud data and a spatiotemporal frequency domain joint attention mechanism, characterized in that, The system includes: The point cloud data processing module is used to convert multi-source meteorological data from ground automatic stations and satellite remote sensing into point cloud format, fill in observation blind spots through spatial interpolation compensation algorithms, and generate station-by-station time series. The spatiotemporal frequency domain joint attention module is used to decompose the spatiotemporal data of meteorological elements into the frequency domain through Fourier transform. Then, the phase modulation module is used to enhance the frequency domain attention of periodic weather characteristics. Combined with the multi-head self-attention mechanism, the spatiotemporal dependence of geographic information and meteorological elements is integrated. The lightweight inference module is used to reduce the model parameter precision from Float32 to Float16, thereby reducing storage overhead and dynamically releasing GPU intermediate tensor memory to avoid resource overflow. The lightweight inference module employs a multi-level fault-switching mechanism: when the main inference path fails, it switches to RBF kernel function interpolation forecast; in case of extreme anomalies, it reverts to historical climate state predictions. When the main inference path fails, the system switches to RBF kernel function interpolation prediction. The specific implementation methods include: The system monitors the operation status and output quality of the main weather forecast inference path in real time. The monitoring includes, but is not limited to, the survival status of computing nodes, the timeliness of input data, the confidence index of output results, the spatial / temporal self-consistency of forecast results, and the computing resource consumption index. Based on a preset multi-dimensional fault judgment rule set, the system dynamically evaluates the fault level of the main inference path. The fault level includes at least normal, warning, partial failure, and complete failure. The confidence index is calculated as follows: By running multiple lightweight simplified models in parallel to quickly predict the same input data, the average similarity between the output of the main inference path and the output of the simplified model is calculated, and a low confidence alarm is triggered when the similarity is lower than the threshold. Check whether the physical quantities output by the main inference path meet the preset atmospheric physical constraint equations, calculate the number and degree of variables that violate the constraints, and mark them as low confidence when they exceed the tolerance range. The current input data is matched with the historical database for similarity. If the prediction error of the main model is significantly higher than that of the backup model in similar historical cases, the confidence score of the current output is reduced. Secondly, the specific calculation method for the aforementioned spatial / temporal self-consistency is as follows: The forecast area is divided into grids, and the gradient values ​​of the same type of meteorological variables in adjacent grid cells are calculated. If there are more than 5 consecutive grid cells whose gradient values ​​exceed 200% of the maximum gradient threshold in climatological history, it is considered spatially inconsistent. The method for calculating the time self-consistency is as follows: A sliding window analysis was performed on the results of three consecutive forecast times for the same geographical location to calculate the variable differences between the later and earlier times. If the direction of the difference value is opposite to the trend of the historical observation data of the previous 24 hours for three consecutive times, it is judged as time inconsistency.

2. The prediction system based on point cloud data and spatiotemporal frequency domain joint attention mechanism according to claim 1, characterized in that, The point cloud data processing module fills in the observation blind spots through spatial interpolation compensation algorithms, including: Based on the DEM digital elevation and KNN analysis results from neighboring stations, the trust propagation algorithm is used to iteratively correct missing values. By aligning satellite grid data through a time sliding window, the brightness temperature difference in the infrared band and the RGB channel features are extracted.

3. The prediction system based on point cloud data and spatiotemporal frequency domain joint attention mechanism according to claim 1, characterized in that, Phase modulation modules enhance the frequency domain attention to periodic weather features, including: The phase modulation module applies dynamic weights to the frequency domain components after Fourier transform, with the formula: Output = FFT(X)⋅σ(W⋅Phase(X)+b), where X is the input spatiotemporal data tensor, FFT(X) is the Fast Fourier Transform performed on the spatiotemporal dimension of X, Phase(X) is the phase component extracted from FFT(X), W is the weight matrix, b is the bias vector, (W⋅Phase(X)+b) is the linear transformation of the phase component, σ(W⋅Phase(X)+b) is the activation function, and Output is the weighted frequency domain feature.

4. The prediction system based on point cloud data and spatiotemporal frequency domain joint attention mechanism according to claim 2, characterized in that, The missing values ​​are iteratively corrected using the trust propagation algorithm based on the DEM digital elevation and neighboring station KNN analysis results, including: Perform time-dimensional exponential sliding imputation on missing data from ground-based automatic weather stations, using the following formula: ； in, The padding value at time t, The padding value at time t-1 These are the observations from the previous 10 minutes. ; For spatially missing data, perform trust propagation iterative imputation with DEM constraints. The imputation method is as follows: Calculate the elevation difference between the target station i and the neighboring station j. ; The neighboring station observations were corrected based on the vertical gradient β of meteorological elements. The corrected neighboring station observations are: ; Based on spatial weights The weighted average correction value is then used to output the iterative estimate of the target site: ； in The difference in altitude between stations This represents the vertical temperature lapse rate. The KNN weights represent the spatial weight of station j relative to station i, and are calculated using the inverse of the distance between the two points or a Gaussian kernel function. This is the supplementary value for station i in the (K+1)th iteration, used to fill in the missing meteorological elements in the final estimate, and thus correct the reasonable forecast value after terrain correction. Let i be the set of neighboring sites. The normalized denominator, representing the weights, ensures that the output value remains within a reasonable physical range. Let j be the value of the neighboring station j in the kth iteration.

Images