Weather prediction method and system based on AI multi-mode integration, computer and storage medium
By using an AI-based multi-model fusion method to dynamically evaluate and fuse multi-source forecast data, the bias problem of single-model forecasts is solved, and more accurate and stable weather forecasts are achieved.
Patent Information
- Application Number
- CN202611273040.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-06
AI Technical Summary
In existing technologies, single numerical weather prediction models have inherent biases and uncertainties, and multi-model integration methods are difficult to adapt to complex and ever-changing actual weather processes, resulting in insufficient forecast accuracy and stability.
An AI-based multi-mode fusion approach is adopted. Heterogeneous meteorological forecast data is acquired, standardized preprocessed, statistical features and deep learning features are extracted, fusion weights are dynamically allocated, and weighted fusion is performed to generate the optimal forecast field. The fusion process is then optimized through forecast evaluation indicators.
It significantly improves the accuracy and stability of weather forecasts, and enhances the robustness and long-term applicability of the forecasting system.
Smart Images
Figure CN123047032A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical and health data processing technology, and in particular to a weather forecasting method, system, computer, and storage medium based on AI multi-mode integration. Background Technology
[0002] Currently, weather forecasting primarily relies on forecast results generated by global or regional numerical weather prediction models (such as ECMWF and GFS). However, any single numerical model inherently suffers from systematic biases and uncertainties due to initial field errors, imperfect physical parameterization schemes, and other factors. Furthermore, different models perform differently under varying weather conditions, geographical regions, and forecast lead times; no single model remains optimal in all situations. To improve forecast reliability, existing technologies often employ multi-model ensemble forecasting, such as simply averaging the outputs of multiple models or assigning fixed weights for integration. However, this static integration method struggles to adapt to complex and ever-changing real-world weather processes and cannot dynamically adjust based on specific forecast objects (such as different meteorological elements) and real-time forecast performance. Consequently, the integrated results often fail to significantly outperform or may even be inferior to the best single-model forecast, limiting further improvements in forecast accuracy. Therefore, a method capable of intelligently evaluating, dynamically optimizing, and adaptively fusing multi-source forecast information is urgently needed. Summary of the Invention
[0003] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to propose a weather forecasting method, system, computer, and storage medium based on AI multi-mode integration, which can realize intelligent evaluation, dynamic optimization, and adaptive fusion of multi-source numerical weather forecast results, thereby effectively improving the accuracy and stability of the final weather forecast product.
[0004] One technical solution adopted in this invention is: a weather forecasting method based on AI multi-model integration, which includes the following steps: S1: Acquire heterogeneous weather forecast data from at least two independent data sources, and perform standardized preprocessing on the data to form a standardized dataset with unified spatiotemporal structure and consistent format; S2: Extract the statistical and deep learning features of the forecast data from each data source in the standardized dataset, and initially generate a set of feature weights that reflect the differences in the current forecast quality of each data source. S3: For the target meteorological element, a dynamic integrated weight is assigned to the forecast results of each data source based on the historical forecast performance of each data source. S4: Using the integrated weights, the forecast results from all data sources are weighted and fused to generate the final forecast field of the target meteorological element.
[0005] Furthermore, the method also includes the following steps: S5: Compare the final forecast field with the actual observation field during the same period, calculate the forecast evaluation index, and optimize the integrated weight calculation strategy in step S3 or the deep learning feature extraction model parameters in step S2 based on the evaluation index.
[0006] Furthermore, step S1 includes the following sub-steps: S11: Obtain gridded forecast data for the target area at the same start time and forecast lead time from at least two independent global or regional numerical weather prediction centers; S12: The acquired gridded forecast data is parsed, resampled to a uniform spatial grid resolution, and interpolated and aligned to a uniform time series. At the same time, the format is converted to form the standardized dataset.
[0007] Furthermore, step S2 includes the following sub-steps: S21: For each data source, calculate the statistical characteristics of its forecast data for the target meteorological element within a preset historical period. The statistical characteristics include the climate average, the variance of the forecast error, and the correlation coefficient of the forecast trend. S22: Input the spatial-temporal forecast field data of each data source in the target area into a pre-trained feature extraction network to extract deep learning features that can characterize the multi-scale structure and evolution of the weather system. S23: Based on the statistical features and deep learning features, generate feature weights for each data source using a preset weight estimation function: ; in, This represents the feature weights of the i-th data source. This represents the initial weight estimation function. This represents the climate average. The variance represents the forecast error. The correlation coefficient represents the predicted trend. This represents a deep learning feature vector.
[0008] Furthermore, step S3 includes the following sub-steps: S31: For the target meteorological element, calculate the average absolute error of the historical forecast for each data source within the preset historical period; S32: Calculate the final dynamic ensemble weights for each data source based on the feature weights and the mean absolute error of historical forecasts. ; in, This represents the integrated weight of the i-th data source. This represents the mean absolute error of historical forecasts from the i-th data source. This represents a positive value to prevent the denominator from being zero, and N represents the total number of data sources participating in the integration.
[0009] Furthermore, step S4 includes the following sub-steps: S41: For each grid point on the standard grid, the fused forecast value for that grid point is obtained by multiplying the forecast values of each data source for the target meteorological element at that grid point with their corresponding final dynamic ensemble weights and summing the results. The calculation formula is as follows: ; in, Indicates spatial location The final fusion forecast value at the location, Indicates the spatial location of the i-th data source. Forecast values for target meteorological elements; S42: Traverse all standard grid points and repeat the calculation in step S41 to generate the final forecast field of the target meteorological elements covering the entire target area.
[0010] Furthermore, step S5 includes the following sub-steps: S51: Compare the final forecast field with the actual observation field at the same spatiotemporal grid points, and calculate at least one forecast evaluation index, wherein the forecast evaluation index includes mean absolute error, root mean square error, temporal correlation coefficient or spatial correlation coefficient. S52: Based on the forecast evaluation index, adjust the parameters of the initial weight estimation function, and / or the dynamic weight fusion calculation formula, and / or the model parameters of the feature extraction network through an optimization algorithm to minimize the evaluation index of subsequent forecasts.
[0011] The second technical solution adopted in this invention is a weather forecasting system based on AI multi-mode integration, which includes the following modules: The preprocessing module is used to acquire heterogeneous weather forecast data from at least two independent data sources and perform standardized preprocessing on the data to form a standardized dataset that is spatiotemporally unified and formatted. The feature statistics module is used to extract statistical and deep learning features from the forecast data of each data source in the standardized dataset, and initially generate a set of feature weights that reflect the differences in the current forecast quality of each data source. The weight calculation module is used to assign a dynamic integrated weight to the forecast results of each data source based on the historical forecast performance of each data source for the target meteorological element. The forecast integration module is used to use the integration weights to perform weighted fusion of forecast results from all data sources to generate the final forecast field of the target meteorological element.
[0012] The third technical solution adopted by the present invention is a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the weather forecasting method based on AI multi-mode integration as described above.
[0013] The fourth technical solution adopted in this invention is a computer-readable storage medium storing a computer program thereon, characterized in that the computer program, when executed by a processor, implements the steps of the weather forecasting method based on AI multi-mode integration as described above.
[0014] The weather forecasting method, system, computer, and storage medium based on AI multi-mode integration of the present invention have at least the following beneficial effects: Based on the specific implementation methods described above, the beneficial effects of this technical solution are mainly reflected in the following three aspects: Firstly, by fusing multi-source heterogeneous forecast data and utilizing statistical methods and deep learning techniques to collaboratively extract multi-level features, this solution effectively overcomes the inherent biases and uncertainties of a single numerical forecast model. Secondly, by dynamically allocating integrated weights based on the historical performance of the data source and current forecast features, intelligent optimization and adaptive weighted fusion of multi-source forecast results are achieved, significantly improving the accuracy and stability of the final weather forecast product. Thirdly, this method establishes an evaluation feedback optimization mechanism, enabling continuous self-improvement and enhancing the robustness and long-term applicability of the forecast system. Attached Figure Description
[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a weather forecasting method based on AI multi-mode integration according to the present invention.
[0016] Figure 2 for Figure 1 The flowchart of sub-process S1.
[0017] Figure 3 for Figure 1 The sub-flowchart of S2.
[0018] Figure 4 for Figure 1 The sub-flowchart of S3.
[0019] Figure 5 This is a block diagram of a weather forecasting system based on AI multi-mode integration according to the present invention. Detailed Implementation
[0020] The invention will now be further described with reference to the accompanying drawings.
[0021] Please see Figure 1 The above is a flowchart of a weather forecasting method based on AI multi-mode integration according to the present invention. The method may include the following steps: S1: Acquire heterogeneous weather forecast data from at least two independent data sources and perform standardized preprocessing on the data to form a standardized dataset with unified spatiotemporal structure and consistent format. This S1 step can select two independent data sources: the European Centre for Medium-Range Weather Forecasts (ECMWF) and the National Center for Environmental Prediction-Global Forecasting System (NCEP-GFS) of the United States. Obtain gridded 2-meter temperature forecast data for the same target area (e.g., 115°–123°E, 27°–35°N) at the same start time (e.g., 00:00 UTC on October 1, 2025), with 3-hour intervals, for the next 24 hours. The acquired raw data formats may differ (e.g., GRIB or NetCDF), and the spatial resolution may also differ (e.g., 0.1°×0.1° and 0.25°×0.25°). Through subsequent standardized preprocessing, the two sets of data are resampled onto a unified 0.25°×0.25° latitude and longitude grid, and the forecast time interpolation is aligned to a unified 00, 03, 06...24-hour time series. Finally, it is converted into a unified NetCDF format, thus forming a standardized dataset.
[0022] This S1 step may also include the following sub-steps: Step S1 includes the following sub-steps: S11: Obtain gridded forecast data for the target area at the same start time and forecast lead time from at least two independent global or regional numerical weather prediction centers. This S11 step can be achieved by downloading the High Resolution Forecasts (HRES) product from the European Centre for Medium-Range Weather Prediction (ECMWF) public data interface, and simultaneously downloading the Global Forecast System (GFS) product from the server of the National Center for Environmental Prediction (NCEP) in the United States. The acquired data must be gridded forecast data, meaning the data is defined on a regular geographic grid (latitude and longitude grid), rather than scattered data. It is essential to ensure that the data covers the same target area (e.g., the entire country of China), the same meteorological elements (e.g., sea level pressure), the same forecast start time (e.g., today's 00Z), and the same forecast lead time (e.g., the next 72 hours).
[0023] S12: The acquired gridded forecast data is parsed, resampled to a uniform spatial grid resolution, and interpolated and aligned to a uniform time series. Simultaneously, format conversion is performed to form the standardized dataset. This S12 step can utilize computer programs (such as Python's xarray or cfgrib library) to read the downloaded data files and identify their dimensions (longitude, latitude, time, forecast lead time) and variables (temperature, air pressure, etc.). Resampling is necessary because the grids of different data sources are inconsistent; for example, it is necessary to interpolate ECMWF 0.1° grid data to a 0.25° grid consistent with GFS. Specifically, bilinear interpolation can be used, with the formula being a distance-weighted average of the values of four known source grid points surrounding the target point. Interpolation and alignment to a uniform time series means, for example, calculating the forecast values at uniform times such as 00, 06, 12, and 18 using linear interpolation for both GFS 3-hour interval data and ECMWF 6-hour interval data. Format conversion typically involves saving all processed data in a unified format commonly used in scientific data: NetCDF. This format is highly standardized and easy for subsequent programs to read.
[0024] S2: For the forecast data from each data source in the standardized dataset, extract its statistical and deep learning features, and initially generate a set of feature weights reflecting the differences in the current forecast quality of each data source. This S2 step extracts features from the forecast data of ECMWF and GFS in the standardized dataset separately. For ECMWF forecast data, on the one hand, calculate its statistical characteristics for forecasts with the same lead time over the past 30 days, such as the 30-day average temperature (climate average), the variance of the forecast error, and the correlation coefficient between the forecast change trend and the climate trend. On the other hand, input the spatial field (e.g., an 8×8 grid area) and time series (8 time intervals) data of this data source's current forecast into a pre-trained hybrid model of a convolutional neural network (CNN) and a long short-term memory network (LSTM) to extract a 128-dimensional deep learning feature vector. This vector encodes the spatial structure and temporal evolution pattern of the weather system in this forecast. Based on these extracted features, the feature weights of the ECMWF data source in this round of forecasts are initially calculated using a preset function. The same processing is performed on the GFS data source to obtain its feature weights. The weights of the two features mentioned above preliminarily reflect the relative quality of the two data sources in this forecast.
[0025] Please see Figure 2 This S2 step may include the following sub-steps: S21: For each data source, calculate the statistical characteristics of its forecast data for the target meteorological element within a preset historical period. These statistical characteristics include the climate average, the variance of the forecast error, and the correlation coefficient of the forecast trend. Specifically, the climate average in step S21 calculates the average forecast value at the target grid point for the same forecast lead time (e.g., 24-hour forecast) each day over the past 30 days (preset historical period), reflecting the forecast climate state of the data source. The variance of the forecast error is calculated by comparing the daily forecast value with the corresponding actual value over the past 30 days, and then calculating the variance of these 30 error values. A smaller variance indicates higher historical forecast stability of the data source. The correlation coefficient of the forecast trend is calculated by comparing the difference sequence (representing the forecast trend) between the data source's "24-hour forecast" and "12-hour forecast" over the past 30 days with the corresponding actual trend sequence. This coefficient reflects the data source's ability to capture the evolution trend of weather systems.
[0026] S22: Input the spatial-temporal forecast field data of each data source in the target area into a pre-trained feature extraction network to extract deep learning features that can characterize the multi-scale structure and evolution of the weather system. In this S22 step, the spatial field of the current forecast from data source i (e.g., a multi-channel two-dimensional field containing temperature, humidity, and air pressure) is input into a convolutional neural network (CNN) to extract spatial features. Then, the CNN outputs from multiple forecast lead times are arranged chronologically and input into a long short-term memory network (LSTM) to extract spatiotemporal evolution features. The hidden state of the last time step of the LSTM constitutes a fixed-length deep learning feature vector, which encodes the complex nonlinear patterns of the current forecast field.
[0027] S23: Based on the statistical features and deep learning features, generate feature weights for each data source using a preset weight estimation function: ; in, This represents the feature weights of the i-th data source. This represents the initial weight estimation function. This represents the climate average. The variance represents the forecast error. The correlation coefficient represents the predicted trend. This represents a deep learning feature vector.
[0028] The specific implementation process of step S23 is as follows: Feature normalization and preprocessing: reducing variance Convert to stability score This allows the smaller the variance to be, the closer the score is to 1. The correlation coefficient... pass Mapping to the interval [0, 1], denoted as Deep learning feature vectors Mapped to a scalar through a fully connected layer Where W and b are learnable parameters, This can be considered a score for physical plausibility.
[0029] Calculate the overall score: Perform a linear weighted summation on the features processed above to obtain the overall score for data source i. ,in, , , All are preset weighting coefficients, and .
[0030] Normalization to weights: Finally, the overall score is normalized to feature weights using the softmax function, ensuring that the sum of all weights is 1. ; in, The temperature parameter controls the smoothness of the weight distribution. Thus, the function... The mapping from features to weights has been completed.
[0031] S3: For the target meteorological element, a dynamic integrated weight is assigned to the forecast results of each data source based on the historical forecast performance of each data source. This S3 step takes the core meteorological element of "2-meter temperature" as an example, querying the historical performance of two data sources over the past 30 days. The calculated historical average absolute error of ECMWF is 1.2°C, and that of GFS is 1.5°C. This indicates that ECMWF's historical performance is more stable. The feature weights obtained in step S2 are combined with the historical errors, and the final dynamic integrated weight is calculated using an optimization formula. This integrated weight considers both the immediate characteristics of this forecast and the long-term historical reliability.
[0032] Please see Figure 3 This S3 step may also include the following sub-steps: S31: For the target meteorological element, calculate the mean absolute error of the historical forecast for each data source within a preset historical period; the calculation formula for the mean absolute error of the historical forecast in step S31 is as follows, taking data source i as an example: ; Where K represents the number of days in the preset historical period (e.g., K=30 days), and L represents the number of grid points used for verification within the target area. This represents the forecast value of data source i at the l-th grid point on the k-th day. This represents the corresponding real-time observation value. The calculation runs automatically in the background periodically, and the results are stored in a database for real-time retrieval.
[0033] S32: Calculate the final dynamic ensemble weights for each data source based on the feature weights and the mean absolute error of historical forecasts. ; in, This represents the integrated weight of the i-th data source. This represents the mean absolute error of historical forecasts from the i-th data source. This represents a positive value to prevent the denominator from being zero, and N represents the total number of data sources participating in the integration.
[0034] S4: Using the integrated weights, the forecast results from all data sources are weighted and fused to generate the final forecast field for the target meteorological element. In step S4, for each grid point on the standardized grid within the target area (e.g., 116°E, 28°N), the ECMWF temperature forecast value at that point is multiplied by its corresponding weight, and the GFS forecast value at that point is multiplied by its corresponding weight. The two products are then added together to obtain the final fused forecast value for that grid point. This weighted calculation is repeated for all grid points within the area, ultimately generating a complete final forecast field covering the entire East China region with 2-meter temperatures for the next 24 hours.
[0035] This S4 step may also include the following sub-steps: S41: For each grid point on the standard grid, the fused forecast value for that grid point is obtained by multiplying the forecast values of each data source for the target meteorological element at that grid point with their corresponding final dynamic ensemble weights and summing the results. The calculation formula is as follows: ; in, Indicates spatial location The final fusion forecast value at the location, Indicates the spatial location of the i-th data source. This provides the forecast values for the target meteorological elements; step S41 assumes that the target area has a standard grid of 100×100 points. For any given grid point, its latitude and longitude coordinates are... The program reads the raw forecast value of data source i (such as ECMWF) at that grid point from a standardized dataset. Obtain the final dynamic integration weights corresponding to this data source. The summation operation after multiplication is performed on a per-data source basis: first calculate... , then calculate Continue this process until the Nth data source is reached, then sum all N products to obtain the grid point. Final fusion forecast value .
[0036] S42: Traverse all standard grid points and repeat the calculation in step S41 to generate the final forecast field of the target meteorological elements covering the entire target area. This step S42 uses nested loops or efficient array operations in a computer program to traverse all 100×100 grid points described in step S41. For each grid point, perform the weighted fusion calculation described in S41 once. When all grid points... After all calculations are completed, these values are rearranged according to their original geographical locations (longitude and latitude) to form a new two-dimensional data matrix with the same spatial dimensions as the input data (100×100). This data matrix is the final forecast field for the target meteorological elements, covering the entire target area.
[0037] S5: Compare the final forecast field with the concurrent observed field, calculate the forecast evaluation index, and optimize the ensemble weight calculation strategy in step S3 or the deep learning feature extraction model parameters in step S2 based on the evaluation index. This S5 step compares the final forecast field generated in step S4 with the actual observed field (e.g., from the interpolation analysis field of the meteorological station). Calculate the forecast evaluation index, such as the mean absolute error (MAE) for all grid points in the entire region, assuming the MAE of this fusion forecast is 1.0°C. Simultaneously, calculate the MAE of this forecast from a single reference data source (e.g., ECMWF), which is 1.1°C. The comparison shows that the fusion forecast is more skillful. Based on this evaluation result, perform optimization. For example, if the skill score of the fusion forecast has been declining over a period of time, the system can automatically initiate an optimization process. This process aims to minimize the MAE of the fusion prediction during the validation period. It utilizes optimization algorithms such as gradient descent to adjust the parameters of the CNN-LSTM hybrid model used to extract deep learning features in step S2, or to adjust a coefficient in the formula for combining feature weights and historical errors in step S3. Through this periodic feedback optimization, the performance of the entire ensemble system can adaptively and continuously improve.
[0038] This S5 step may also include the following sub-steps: S51: Compare the final forecast field with the actual observation field at the same spatiotemporal grid points, and calculate at least one forecast evaluation index, including mean absolute error, root mean square error, temporal correlation coefficient, or spatial correlation coefficient. In calculating the forecast evaluation index, step S51 requires ensuring the fairness of the comparison, i.e., "comparing at the same spatiotemporal grid points." This means that the actual observation field (which may be a grid field interpolated from station observations) needs to be interpolated onto a standard grid that is completely consistent with the final forecast field. Then, the error is calculated grid-by-grid.
[0039] S52: Based on the forecast evaluation index, adjust the parameters of the initial weight estimation function, and / or the dynamic weight fusion calculation formula, and / or the model parameters of the feature extraction network through an optimization algorithm to minimize the evaluation index of subsequent forecasts. This S52 step is typically performed offline and periodically. One implementation is to construct a training task aimed at minimizing the evaluation index (such as MAE). Data from a historical period (e.g., the past 90 days) is used as the training set, and the parameters of the initial weight estimation function, the variant parameters of the dynamic weight formula, and the parameters of the feature extraction network are all set as trainable variables. Then, iterative training is performed using an optimization algorithm (such as Adam gradient descent). In each iteration, steps S2-S4 are simulated on the entire training set with the current parameters to obtain a simulated fused forecast field and calculate its MAE loss. Then, the gradient of the loss with respect to each trainable parameter is calculated using the backpropagation algorithm, and these parameters are updated. Through multiple iterations, the parameters of the entire ensemble system are optimized, thereby improving the accuracy of future forecasts.
[0040] Please see Figure 5 The diagram above illustrates the structural block of a weather forecasting system based on AI multi-mode integration according to the present invention. The present invention also provides a weather forecasting system based on AI multi-mode integration, which implements the aforementioned weather forecasting method based on AI multi-mode integration. Specifically, the system may include a preprocessing module 100, a feature statistics module 200, a weight calculation module 300, a forecast integration module 400, and an early warning module 500. More specifically: The preprocessing module 100 is used to acquire heterogeneous weather forecast data from at least two independent data sources and perform standardized preprocessing on the data to form a standardized dataset with unified spatiotemporal structure and consistent format. The variational mode decomposition module 200 is used to extract statistical and deep learning features from the forecast data of each data source in the standardized dataset, and initially generate a set of feature weights that reflect the differences in the current forecast quality of each data source. The parameter prediction module 300 is used to assign a dynamic integrated weight to the forecast results of each data source based on the historical forecast performance of each data source for the target meteorological element. The parameter prediction module 400 is used to use the integrated weights to perform weighted fusion of the forecast results from all data sources to generate the final forecast field of the target meteorological element.
[0041] In one embodiment of this application, a computer device is provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the AI-based multi-mode integrated weather forecasting method as described in any of the above embodiments.
[0042] In one embodiment of this application, a computer-readable storage medium is provided, which stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the AI-based multi-mode integrated weather forecasting method as described in any of the above embodiments.
[0043] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.
[0044] The above description merely illustrates preferred embodiments of the present invention and is quite specific and detailed; however, it should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the appended claims.
Claims
1. An AI-based multi-modal integrated weather prediction method, characterized by, The method includes the following steps: S1: Acquire heterogeneous weather forecast data from at least two independent data sources, and perform standardized preprocessing on the data to form a standardized dataset with unified spatiotemporal structure and consistent format; S2: Extract the statistical and deep learning features of the forecast data from each data source in the standardized dataset, and initially generate a set of feature weights that reflect the differences in the current forecast quality of each data source. S3: For the target meteorological element, a dynamic integrated weight is assigned to the forecast results of each data source based on the historical forecast performance of each data source. S4: Using the integrated weights, the forecast results from all data sources are weighted and fused to generate the final forecast field of the target meteorological element. 2.The AI-based multi-modal integrated weather prediction method of claim 1, wherein The method also includes the following steps: S5: Compare the final forecast field with the actual observation field during the same period, calculate the forecast evaluation index, and optimize the integrated weight calculation strategy in step S3 or the deep learning feature extraction model parameters in step S2 based on the evaluation index. 3.The AI-based multi-modal integrated weather prediction method of claim 1, wherein Step S1 includes the following sub-steps: S11: Obtain gridded forecast data for the target area at the same start time and forecast lead time from at least two independent global or regional numerical weather prediction centers; S12: The acquired gridded forecast data is parsed, resampled to a uniform spatial grid resolution, and interpolated and aligned to a uniform time series. At the same time, the format is converted to form the standardized dataset. 4.The AI-based multi-modal integrated weather prediction method of claim 1, wherein Step S2 includes the following sub-steps: S21: For each data source, calculate the statistical characteristics of its forecast data for the target meteorological element within a preset historical period. The statistical characteristics include the climate average, the variance of the forecast error, and the correlation coefficient of the forecast trend. S22: Input the spatial-temporal forecast field data of each data source in the target area into a pre-trained feature extraction network to extract deep learning features that can characterize the multi-scale structure and evolution of the weather system. S23: Based on the statistical features and deep learning features, generate feature weights for each data source using a preset weight estimation function: ； in, This represents the feature weights of the i-th data source. This represents the initial weight estimation function. This represents the climate average. The variance represents the forecast error. The correlation coefficient represents the predicted trend. This represents a deep learning feature vector.
5. The weather forecasting method based on AI multi-mode integration as described in claim 1, characterized in that, Step S3 includes the following sub-steps: S31: For the target meteorological element, calculate the average absolute error of the historical forecast for each data source within the preset historical period; S32: Calculate the final dynamic ensemble weights for each data source based on the feature weights and the mean absolute error of historical forecasts. ； in, This represents the integrated weight of the i-th data source. This represents the mean absolute error of historical forecasts from the i-th data source. This represents a positive value to prevent the denominator from being zero, and N represents the total number of data sources participating in the integration.
6. The weather forecasting method, system, computer, and storage medium based on AI multi-mode integration as described in claim 1, characterized in that, Step S4 includes the following sub-steps: S41: For each grid point on the standard grid, the fused forecast value for that grid point is obtained by multiplying the forecast values of each data source for the target meteorological element at that grid point with their corresponding final dynamic ensemble weights and summing the results. The calculation formula is as follows: ； in, Indicates spatial location The final fusion forecast value at the location, Indicates the spatial location of the i-th data source. Forecast values for target meteorological elements; S42: Traverse all standard grid points and repeat the calculation in step S41 to generate the final forecast field of the target meteorological elements covering the entire target area.
7. The weather forecasting method based on AI multi-mode integration as described in claim 2, characterized in that, Step S5 includes the following sub-steps: S51: Compare the final forecast field with the actual observation field at the same spatiotemporal grid points, and calculate at least one forecast evaluation index, wherein the forecast evaluation index includes mean absolute error, root mean square error, temporal correlation coefficient or spatial correlation coefficient. S52: Based on the forecast evaluation index, adjust the parameters of the initial weight estimation function, and / or the dynamic weight fusion calculation formula, and / or the model parameters of the feature extraction network through an optimization algorithm to minimize the evaluation index of subsequent forecasts.
8. A weather forecasting system based on AI multi-mode integration, characterized in that, The system includes the following modules: The preprocessing module is used to acquire heterogeneous weather forecast data from at least two independent data sources and perform standardized preprocessing on the data to form a standardized dataset that is spatiotemporally unified and formatted. The feature statistics module is used to extract statistical and deep learning features from the forecast data of each data source in the standardized dataset, and initially generate a set of feature weights that reflect the differences in the current forecast quality of each data source. The weight calculation module is used to assign a dynamic integrated weight to the forecast results of each data source based on the historical forecast performance of each data source for the target meteorological element. The forecast integration module is used to use the integration weights to perform weighted fusion of forecast results from all data sources to generate the final forecast field of the target meteorological element.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the weather forecasting method based on AI multi-mode integration as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the weather forecasting method based on AI multi-mode integration as described in any one of claims 1-7.