Meteorological numerical prediction correction method and device based on epfr multi-branch framework

Abstract

This application discloses a method and apparatus for correcting meteorological numerical forecasts based on the EPFR multi-branch framework. The method includes: acquiring initial meteorological data corresponding to each time granularity within a preset time period to obtain an initial meteorological dataset; preprocessing the initial meteorological dataset to generate a meteorological dataset; partitioning the meteorological dataset to generate a meteorological test dataset and a meteorological training sample set; training an EPFR multi-branch model based on the meteorological training sample set; and inputting the meteorological test dataset into the EPFR multi-branch model to obtain a meteorological numerical forecast correction information set. This implementation method relates to the field of meteorological forecast artificial intelligence application technology and significantly improves the accuracy of multivariate forecast correction.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence application technology in meteorological forecasting, specifically to a method and apparatus for correcting meteorological numerical forecasts based on the EPFR multi-branch framework. Background Technology

[0002] The core evolution of weather forecasting technology revolves around "improving forecast accuracy, expanding the time range, and enhancing scenario adaptability." From traditional numerical forecasting to ensemble forecasting, and then to various branch technologies derived from ensemble forecasting, a complete technological development path has been formed. This paper provides a systematic summary of existing technological research.

[0003] Traditional numerical weather prediction is a forecasting technique based on the fundamental equations of atmospheric dynamics and thermodynamics, using numerical calculations to solve for atmospheric motion states. It is a core technology of meteorological forecasting. However, traditional numerical forecasting is deterministic, outputting only a single forecast result. It cannot quantify the uncertainties in the forecasting process, making it difficult to assess forecast reliability. Furthermore, the atmospheric system is a highly complex nonlinear system, and small errors in the initial field accumulate exponentially with the forecast lead time. It is also computationally expensive, heavily reliant on supercomputing resources, and makes it difficult to balance computational efficiency with forecast accuracy.

[0004] To address the insufficient characterization of uncertainty in traditional deterministic numerical weather prediction, ensemble forecasting techniques have emerged. The core idea is to construct a forecast ensemble containing forecast members with different initial fields, parameterization schemes, or model structures. The uncertainty is quantified by utilizing the discreteness of the ensemble members, and more reliable forecast results are obtained through ensemble statistics. This represents a significant breakthrough in meteorological forecasting technology and has led to three core parallel optimization directions within the industry:

[0005] (1) Data-driven meteorological big model is based on artificial intelligence technologies such as deep learning. By mining the spatiotemporal correlation features in massive historical meteorological observation data and numerical forecast data, it directly constructs the mapping relationship from input to output, realizes the prediction and correction of meteorological field, does not rely on complex atmospheric physical equations, and has strong nonlinear fitting ability and efficient reasoning advantages. Representative products include the Pangu-Weather model, the FuXi model, the GraphCast model, and the FourCastNet model. However, these models have significant bottlenecks: First, they are heavily reliant on computing power. For example, training FourCastNet requires 64 Nvidia A100 GPUs (A100 graphics processors) to run for 16 hours, while Pangu-Weather trains networks independently for different forecast steps, with each network requiring 192 V100 GPUs (V100 graphics processors) for up to 16 days of training. Second, they have poor physical interpretability. The models are "black box" structures, lacking explicit constraints on atmospheric physical laws, and their reliability is questionable in extreme scenarios. Third, they do not fully utilize ensemble information. Most models are deterministic prediction architectures and do not fully integrate the uncertainty information of ensemble forecasts.

[0006] (2) Hybrid models aim to combine the physical rationality of traditional numerical models with the strong fitting ability of artificial intelligence technology. By optimizing key aspects of traditional numerical models (such as initial field, parameterization scheme, and forecast result correction) through AI (Artificial Intelligence) modules, they achieve forecast optimization driven by both physics and data. This is an important technical direction for balancing physical interpretability and forecast accuracy, with the AIFS (Artificial Intelligence Forecasting System) model proposed by ECMWF (European Centre for Medium-Range Weather Forecasts) as a representative. However, these models have poor fusion and synergy; the computational costs are superimposed, retaining the high computational cost of traditional numerical models while increasing the training and inference costs of AI modules.

[0007] (3) Ensemble forecast post-processing technology addresses the systematic bias and uncertainty quantification issues of ensemble forecast members. It is a technique that uses statistical or machine learning methods to calibrate and optimize ensemble forecast results. The core objective is to reduce systematic bias, improve forecast accuracy, and optimize the uncertainty quantification effect. It is a key supporting technology for ensemble forecasts to move towards operational applications. Traditional statistical post-processing methods: represented by Ensemble Model Output Statistics (EMOS). Traditional methods such as EMOS rely on linear or simple nonlinear regression models, which are difficult to characterize the complex nonlinear bias of meteorological fields at multivariable, global grid scales. They also lack spatiotemporal consistency constraints, which can easily lead to spatial discontinuities and abnormal time series fluctuations.

[0008] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not form prior art known to those skilled in the art. Summary of the Invention

[0009] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0010] Some embodiments of this disclosure propose a method, apparatus, electronic device, and computer-readable medium for correcting meteorological numerical forecasts based on the EPFR multi-branch framework to address the technical problems mentioned in the background section above.

[0011] In a first aspect, some embodiments of this disclosure provide a meteorological numerical forecast correction method based on the EPFR multi-branch framework. The method includes: acquiring initial meteorological data corresponding to each time granularity within a preset time period to obtain an initial meteorological dataset; performing data preprocessing on the initial meteorological dataset to generate a meteorological dataset; dividing the meteorological dataset to generate a meteorological test dataset and a meteorological training sample set; training an EPFR multi-branch model based on the meteorological training sample set; and inputting the meteorological test dataset into the EPFR multi-branch model to obtain a meteorological numerical forecast correction information set.

[0012] Secondly, some embodiments of this disclosure provide a meteorological numerical forecast correction apparatus based on the EPFR multi-branch framework. The apparatus includes: an acquisition unit configured to acquire initial meteorological data corresponding to each time granularity within a preset time period to obtain an initial meteorological dataset; a data preprocessing unit configured to preprocess the initial meteorological dataset to generate a meteorological dataset; a partitioning unit configured to partition the meteorological dataset to generate a meteorological test dataset and a meteorological training sample set; a training unit configured to train an EPFR multi-branch model based on the meteorological training sample set; and an input unit configured to input the meteorological test dataset into the EPFR multi-branch model to obtain a meteorological numerical forecast correction information set.

[0013] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation of the first aspect above.

[0014] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0015] The above-described embodiments of this disclosure have the following beneficial effects: the meteorological numerical forecast correction method based on the EPFR multi-branch framework of some embodiments of this disclosure significantly improves the accuracy of multivariate forecast correction. Specifically, the reasons for the insufficient accuracy of multivariate forecast correction are: lack of multi-branch collaboration, rigid fusion mechanism, and imbalance of training objectives in existing meteorological numerical forecast correction technologies. Based on this, the meteorological numerical forecast correction method based on the EPFR multi-branch framework of some embodiments of this disclosure, through the core innovations of the original EPFR multi-branch framework, adaptive four-branch fusion mechanism, and custom weighted joint loss function, performs excellently in the full-time (6-360h) forecast correction of core meteorological variables such as 2m air temperature (T2m), mean sea level pressure (MSLP), 500hPa geopotential height (Z500), and 500hPa wind speed (WS500). Compared with existing technologies (such as Pangu-Weather, ECMWF IFS (Integrated Forecast System) post-processing methods, etc.), it achieves the following significant beneficial effects:

[0016] 1. Significantly Improved Multivariate Forecast Correction Accuracy: The EPFR (Ensemble Probabilistic-Forecasting-Residual) four-branch collaborative architecture achieves full-link modeling of "probabilistic information mining - deterministic correction - adaptive fusion - residual optimization". It fully utilizes the probabilistic uncertainty information of ensemble members through the PEW (Probabilistic Ensemble Weighting) branch, and captures nonlinear bias characteristics through the DDF (Deterministic Direct Forecasting) branch. In addition, it incorporates an adaptive fusion branch (AB) to learn variable fusion coefficients to adaptively weight the PEW and DDF predictions, so as to take into account the complementary advantages of ensemble information and regression correction. The two branches are fused first, and then the RCF (Residual-Corrected Fusion) branch is used for correction. Finally, the RCF branch is used to further compensate for detail errors. The synergistic utilization of multi-dimensional information enabled EPFR to demonstrate a sustained and significant performance advantage in the long-term and ultra-long-term stages of forecasting for the four core variables: T2m, MSLP, Z500, and WS500, as forecast lead time increases. Compared to HRES (High Resolution Forecast), EPFR achieved stable improvements in all four variables in the long-term (144–240h) and ultra-long-term (240–360h) stages, with improvement rates ranging from 11.56%–17.58% and 9.86%–17.31%, respectively, and average improvement rates of approximately 13.74% and 13.03%. This represents an average improvement of 4% compared to the IFS ENS mean (Integrated Forecasting System-Ensemble mean).

[0017] 2. Improved Training Stability and Robustness: A custom weighted joint loss function is designed with differentiated loss terms for different modeling objectives in the four branches. By adaptively balancing the training priorities of each variable and branch, the overfitting problem caused by a single loss function is effectively avoided. Combined with label smoothing strategy and early stopping mechanism, the model's generalization ability to all four types of variables is improved by 30% in small sample scenarios, the loss convergence speed is accelerated by 20%-30% during training, and the consistency of the correction effect of each variable under different batches of data training is significantly improved.

[0018] 3. Lightweight design lowers the deployment threshold: The EPFR framework uses U-Net (image segmentation network) as the backbone for spatial feature extraction. Compared with complex architectures such as Transformer (transformer model) and graph neural networks, the number of parameters is reduced by more than 100%. It does not rely on supercomputing clusters and is easy to deploy and apply quickly.

[0019] 4. Optimization of information utilization and modeling flexibility: The four branches achieve feature sharing through the U-Net backbone, fully exploring the ensemble probability information and multivariate nonlinear bias characteristics; the framework structure is scalable, and the backbone network (such as Transformer) can be flexibly replaced or new branches can be added to adapt to the forecast correction needs of more meteorological variables (such as precipitation and humidity) and different data sources, and has broad business promotion value. Attached Figure Description

[0020] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0021] Figure 1 This is a flowchart of some embodiments of the meteorological numerical forecast correction method based on the EPFR multi-branch framework according to this disclosure;

[0022] Figure 2 This is a technical roadmap of some embodiments of the meteorological numerical forecast correction method based on the EPFR multi-branch framework of this disclosure;

[0023] Figure 3 This is a schematic diagram of the structure of an EPFR multi-branch model according to some embodiments of the meteorological numerical forecast correction method based on the EPFR multi-branch framework disclosed herein;

[0024] Figure 4 This is a schematic diagram of the structure of some embodiments of the meteorological numerical forecast correction device based on the EPFR multi-branch framework according to the present disclosure;

[0025] Figure 5 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0026] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0027] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0028] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0029] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0030] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0031] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] Figure 1 A flowchart 100 is shown, illustrating some embodiments of a meteorological numerical forecast correction method based on the EPFR multibranch framework of this disclosure. This meteorological numerical forecast correction method based on the EPFR multibranch framework includes the following steps:

[0033] Step 101: Obtain the initial meteorological data corresponding to each time granularity within the preset time period to obtain the initial meteorological dataset.

[0034] In some embodiments, the implementing entity of the meteorological numerical forecast correction method based on the EPFR (Ensemble Probabilistic-Forecasting-Residual) multi-branch framework can acquire initial meteorological data corresponding to each time granularity within a preset time period to obtain an initial meteorological dataset. For example, the preset time period can be from January 1, 2018 to December 31, 2022. The time granularity can be 6 hours. The spatial resolution of the initial meteorological data in the initial meteorological dataset can be 5.625° (32×64 grid points). As an example, the technical roadmap of the meteorological numerical forecast correction method based on the EPFR (Ensemble Probabilistic-Forecasting-Residual) multi-branch framework is as follows: Figure 2 As shown, Figure 2 A technical roadmap is shown for some embodiments of the meteorological numerical forecast correction method based on the EPFR multi-branch framework according to this disclosure. Figure 2 The ECMWF IFS-ENS (Integrated Forecasting System-Ensemble) ensemble data can be the initial meteorological dataset. The ERA5 (Fifth Generation Atmospheric Reanalysis Dataset) observational ground truth data can be the sample labels of the EPFR multi-branch model. WeatherBench2 can be the weather baseline 2.

[0035] In practice, the aforementioned implementing entities can obtain initial meteorological data corresponding to each time granularity within a preset time period through the following steps:

[0036] The first step is to determine the initial field and the initial perturbation field set. In practice, firstly, the aforementioned implementing entity can obtain the ECMWF (Early Weather Forecasting Center) high-precision atmospheric analysis field from a terminal device (e.g., the WeatherBench2 platform) via wired or wireless connection. Secondly, the implementing entity can determine the aforementioned ECMWF high-precision atmospheric analysis field as the initial field. Then, the initial perturbation field set is generated using a preset perturbation algorithm. For example, the preset perturbation algorithm can be Singular Vector (SV) or Ensemble Data Assimilation (EDA). The initial perturbation field set can include 50 initial perturbation fields. Here, the initial field can include, but is not limited to, at least one of the following: 2m (meter) air temperature, mean sea level pressure, geopotential height, wind speed, and specific humidity.

[0037] Optionally, the aforementioned execution entity can perform time integration processing on the initial field using a preset time integration algorithm to generate a control forecast. For example, the preset time integration algorithm could be an IFS (Integrated Forecast System). As an example, the IFS has a time resolution of 6 hours, forecasts up to 360 hours, and a forecast lead time covering 0-360 hours.

[0038] The second step is to superimpose the aforementioned initial perturbation field set onto the aforementioned initial field to obtain the perturbation initial field set. In practice, the aforementioned executing entity can superimpose each initial perturbation field in the aforementioned initial perturbation field set onto the aforementioned initial field to generate the perturbation initial field, thus obtaining the perturbation initial field set.

[0039] Therefore, a set of perturbation initial fields can be obtained by applying reasonable, small perturbations that conform to atmospheric error characteristics to the high-precision atmospheric analysis field (three-dimensional atmospheric states such as temperature, wind field, humidity, air pressure, and geopotential height) used in control forecasting. Its function is to characterize the forecast uncertainty caused by initial observation and analysis errors, and it forms the basis for generating ensemble members. In practice, the aforementioned implementing entity can introduce parameterized scheme perturbations into the physical processes of numerical models to characterize the uncertainties of the model's physical processes.

[0040] The third step is to perform time integration on the aforementioned initial perturbation field set to generate a set of members. In practice, the executing entity can use the aforementioned preset time integration algorithm to perform time integration on each initial perturbation field in the initial perturbation field set to generate set members and obtain the set of members.

[0041] Therefore, the ensemble members in the ensemble membership set can be composed of the forecast start time and the forecast lead time. The forecast start time can be the start time of the ECMWF daily ensemble forecasts at 0:00 and 12:00 UTC (Coordinated Universal Time); the forecast lead time is the forecast length of 6 to 360 hours after each start time. The ensemble members in the ensemble membership set include 61 forecast steps (from 0h to 360h). Thus, 50 ensemble members for a forecast time can be used as a training sample to subsequently train the EPFR multi-branch model to predict data at 0:00 and 12:00 daily, with 61 steps for each time point.

[0042] The fourth step is to determine the above set of members as the initial meteorological data.

[0043] It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other currently known or future wireless connection methods.

[0044] Step 102: Perform data preprocessing on the initial meteorological dataset to generate a meteorological dataset.

[0045] In some embodiments, the aforementioned executing entity may perform data preprocessing on the aforementioned initial meteorological dataset to generate a meteorological dataset.

[0046] In practice, the aforementioned implementing entity can perform data preprocessing on the initial meteorological dataset using the following steps to generate a meteorological dataset:

[0047] The first step is to perform data cleaning on each initial meteorological data point in the initial meteorological dataset to generate cleaned meteorological data, thus obtaining a cleaned meteorological dataset. In practice, the executing entity can use a preset data cleaning algorithm to perform data cleaning on each initial meteorological data point in the initial meteorological dataset to generate cleaned meteorological data, thus obtaining a cleaned meteorological dataset. For example, the preset data cleaning algorithm could be: deleting data representing empty data.

[0048] The second step involves normalizing each meteorological cleansing data point in the aforementioned meteorological cleansing dataset to generate meteorological data, thus obtaining the meteorological dataset. In practice, the executing entity can use a preset normalization algorithm to normalize each meteorological cleansing data point in the aforementioned meteorological cleansing dataset to generate meteorological data. For example, the preset normalization algorithm could be the Z-score (standard score) standardization algorithm.

[0049] Step 103: Divide the meteorological dataset into meteorological test dataset and meteorological training sample set.

[0050] In some embodiments, the executing entity may partition the meteorological dataset to generate a meteorological test dataset and a meteorological training sample set. In practice, firstly, the executing entity may determine each meteorological dataset representing a preset test time period within the aforementioned meteorological dataset as the meteorological test dataset. Secondly, the executing entity may determine each meteorological data in the aforementioned meteorological dataset, excluding the meteorological test data included in the aforementioned meteorological test dataset, as the meteorological training sample set. For example, the preset test time period may be from January 1, 2020 to December 31, 2020. The meteorological training samples in the meteorological training sample set may include a set of membership data.

[0051] Step 104: Train the EPFR multi-branch model based on the meteorological training sample set.

[0052] In some embodiments, the aforementioned execution entity can train an EPFR multi-branch model based on the aforementioned meteorological training sample set. The EPFR (Ensemble Probabilistic-Forecasting-Residual) multi-branch model can be a pre-trained neural network model that takes meteorological data as input and forecast correction information as output. As an example, the network structure of the EPFR multi-branch model is as follows: Figure 3 As shown, Figure 3The diagram illustrates the structure of an EPFR multi-branch model according to some embodiments of the meteorological numerical forecast correction method based on the EPFR multi-branch framework of this disclosure. As an example, in EPFR, E corresponds to the ensemble forecast basic attribute (Ensemble, core anchor point, supporting ensemble member information mining and fusion). P in EPFR corresponds to the probabilistic ensemble weighting (PEW) branch. F in EPFR corresponds to the deterministic direct forecasting (i.e., the original DDF branch, deterministic direct forecasting). R in EPFR corresponds to the residual-corrected fusion (RCF) branch. Therefore, the core idea of ​​this application is to use a meteorological training sample set as a basis, and through the complementary and collaborative modeling of the four branches (E, P, F, and R), to connect the entire chain of "ensemble information mining - probabilistic fusion optimization - deterministic forecast correction - accurate residual compensation," while incorporating an adaptive blending (AB) mechanism to balance the advantages of probabilistic and deterministic results, achieving efficient utilization of both ensemble probabilistic information and nonlinear bias characteristics, and achieving accurate optimization of full-time forecasts of meteorological elements.

[0053] In practice, the aforementioned implementing entities can train the EPFR multi-branch model based on the aforementioned meteorological training sample set through the following steps:

[0054] The first step is to obtain the sample label corresponding to each meteorological training sample in the meteorological training sample set, thus obtaining the sample label set. In practice, the aforementioned execution entity can obtain the ERA5 reanalysis data corresponding to each meteorological training sample in the meteorological training sample set from the terminal device via wired or wireless connection, thus obtaining the sample label set.

[0055] The second step is to determine the initial EPFR multi-branch model. This initial EPFR multi-branch model may include: an initial backbone network, an initial probability-weighted fusion branch network, an initial direct regression prediction branch network, an initial adaptive fusion branch network, and an initial residual correction branch network. For example, the initial learning rate of the initial EPFR multi-branch model can be set to 1×10⁻. 4 The batch size was set to 32, the number of training rounds to 200, and the label smoothing coefficient to 0.05. Parameter initialization was based on deep learning training experience. Furthermore, while ensuring the meteorological test dataset remained completely independent, the meteorological training sample set was divided into a training set and a validation set in a 7:3 ratio for parameter learning and hyperparameter selection.

[0056] The initial backbone network can be a neural network that takes target meteorological training samples as input and an initial set of spatial feature information as output. For example, the initial backbone network can use U-Net (image segmentation network) as the spatial feature extraction backbone, extract high-level semantic features through an encoder (convolution + max pooling), and fuse shallow detail features through a decoder (transposed convolution + skip connections). Therefore, the initial backbone network can provide unified feature support for the four branches.

[0057] An Initial Probability Weighted Fusion Branch Network (PEW) can be a network that takes an initial set of spatial feature information as input and outputs initial weighted fusion information. For example, an PEW can take an initial set of spatial feature information as input, pass it through a softmax (normalized exponential function) activation function, and output the dynamic weights of each set member (reflecting the reliability of the member in the current scene). The initial weighted fusion information is obtained by weighted summation along the member dimension. In practice, a label smoothing strategy using one-hot encoding is introduced into the PEW. This can mitigate the overfitting risk caused by hard labels and improve the stability of weight learning.

[0058] An initial direct regression prediction branch network (DDF) can be a neural network that takes an initial set of spatial feature information as input and outputs initial correction information. For example, an initial direct regression prediction branch network can include convolutional and linear regression heads. Therefore, the initial direct regression prediction branch model can directly learn the nonlinear deviation correction term of meteorological elements through convolutional and linear regression heads, and output deterministic correction results.

[0059] An initial adaptive fusion branch network (AB) can be a neural network that takes initial weighted fusion information and initial correction information as input and initial adaptive fusion information as output. For example, the AB can learn variable fusion coefficients α (0≤α≤1) through a gated layer (sigmoid (activation) function) to dynamically weight the PEW probabilistic fusion result (initial weighted fusion information) and the DDF direct regression result (initial correction information) (α×PEW+(1-α)×DDF). This balances the complementary advantages of the PEW and DDF results.

[0060] The initial residual correction branch network (RCF) can be a neural network that takes initial adaptive fusion information as input and initial prediction correction information as output. For example, the initial residual correction branch network can be a single convolutional layer. Therefore, based on the AB output, by learning the residual terms through a single convolutional layer (4→4 channels), it is possible to compensate for the systematic biases and local detail errors that were not eliminated by the preceding branch.

[0061] The third step is to select target meteorological training samples from the meteorological training sample set. In practice, the implementing entity can randomly select meteorological training samples from the meteorological training sample set as target meteorological training samples.

[0062] The fourth step is to input the selected target meteorological training samples into the initial backbone network to obtain the initial spatial feature information set.

[0063] The fifth step is to input the initial spatial feature information set into the initial probability weighted fusion branch network to obtain the initial weighted fusion information.

[0064] The sixth step is to input the initial spatial feature information set into the initial direct regression prediction branch network to obtain the initial correction information.

[0065] Step 7: Input the initial weighted fusion information and the initial correction information into the initial adaptive fusion branch network to obtain the initial adaptive fusion information.

[0066] The eighth step is to input the initial adaptive fusion information into the initial residual correction branch network to obtain the initial prediction correction information.

[0067] Step 9: Based on a preset loss function, determine the difference between the initial forecast correction information and the sample labels corresponding to the target meteorological training samples. The preset loss function can be:

[0068] ,

[0069] in, Indicates the difference value. , , , Indicate weight (e.g., It can be 1, It can be 1, It can be 0.5. It can be 0.5). Indicates PEW loss, Indicates DDF loss. Indicates the loss of AB. Represents the RCF loss, where,

[0070] ,

[0071] in, , Indicate weight (e.g., It can be 0.1. It can be 0.9). This represents the cross-entropy loss for multi-class classification. This represents the regression loss due to the average error. Indicates the sample number. , Represents the total number of samples. Indicates the member sequence number. , This represents the number of set members (the set of members included in the target meteorological training samples, and the number of members in each set). It can be 50). Indicates the first In the nth sample The set members (the first in the meteorological training sample set) The meteorological training samples include the first set of members. Smoothed labels for (each set of members) Indicates the first In the nth sample The weighted probabilities of each set member This represents the initial weighted fused information. This indicates the sample label (ERA5 reanalysis data). Indicates one-hot tag, Represents the smoothing coefficient (e.g., It can be 0.05). Indicates the serial number. Indicates the first In the nth sample The logits values ​​(raw output scores) of each set member. Indicates the first In the nth sample The logits value of each set member. For example, the first... In the nth sample The logits value of each set member can be the raw, unnormalized real-valued vector output from the last layer of the model (usually a fully connected layer). It represents the model's raw prediction score for each class (or each set member here), before it has been converted into probabilities by activation functions such as Softmax and Sigmoid; it is used to calculate the weight probabilities. The input is eventually normalized by Softmax to the fusion weights of each set member. Used to perform weighted merging of set members.

[0072] Step 10: In response to the determination that the difference value meets the preset adjustment condition, adjust the network parameters of the initial EPFR multi-branch model. The preset adjustment condition can be: no improvement in the difference value after 5 consecutive training rounds (i.e., the difference value no longer decreases after 5 consecutive training rounds).

[0073] Optionally, the aforementioned executing entity may also, in response to determining that the difference value does not meet the aforementioned preset adjustment conditions, determine the initial EPFR multi-branch model as an EPFR multi-branch model.

[0074] Step 105: Input the meteorological test dataset into the EPFR multi-branch model to obtain the meteorological numerical forecast correction information set.

[0075] In some embodiments, the aforementioned implementing entity can input the aforementioned meteorological test dataset into the aforementioned EPFR multi-branch model to obtain a meteorological numerical forecast correction information set.

[0076] In practice, the aforementioned implementing entity can input the meteorological test dataset into the aforementioned EPFR multi-branch model through the following steps to obtain the meteorological numerical forecast correction information set:

[0077] The first step is to input each meteorological test data in the above meteorological test dataset into the EPFR multi-branch model to obtain meteorological numerical forecast correction information.

[0078] The second step is to determine the obtained meteorological numerical forecast correction information into a meteorological numerical forecast correction information set.

[0079] Therefore, the EPFR multi-branch model can be used to predict the meteorological numerical forecast correction information for four time points: 0:00, 6:00, 12:00, and 18:00 every day.

[0080] Therefore, conventional evaluation metrics such as RMSE (root mean square error) and ACC (abnormal correlation coefficient) are used to compare the correction effect of the method of this invention with existing baseline models (such as ECMWF HRES (high resolution forecast system) and Pangu-Weather), and output the final forecast correction result that meets the accuracy requirements.

[0081] Therefore, firstly, a dynamic weight and variable fusion coefficient design was implemented: the PEW branch achieves dynamic weight allocation of ensemble members through convolution + softmax, and the AB branch learns the variable fusion coefficient α through a gating mechanism. Compared with existing fixed-weight fusion methods, this approach can adaptively adapt to different forecast scenarios and improve fusion performance. Secondly, an innovative EPFR multi-branch ensemble framework was proposed: this framework, for the first time, collaboratively models four branches: probabilistic weighted fusion, direct nonlinear regression, adaptive fusion, and residual correction, achieving dual utilization of ensemble probability information and nonlinear bias characteristics, thus solving the problem that single-branch models cannot simultaneously handle probabilistic fusion and deterministic correction. Next, a lightweight backbone and efficient training strategy were implemented: U-Net is used as the spatial feature extraction backbone, which reduces the parameter size compared to the Transformer / graph neural network architecture, making it suitable for conventional business computing environments; combined with a weighted joint loss function and a label smoothing-early stopping mechanism, training stability is improved and the risk of overfitting is reduced. Furthermore, it achieves full-process standardization and multi-timeframe adaptation: A complete process solution of "data preprocessing - multi-branch modeling - dynamic optimization - multi-timeframe evaluation" is constructed, adaptable to a full timeframe range of 6-360 hours, covering core meteorological elements, and compatible with the WeatherBench2 platform, making it easy to integrate into existing business processes. Then, a custom weighted joint loss function is designed: differentiated loss terms are designed for different modeling objectives in the four branches, and adaptive weight allocation achieves a synergistic balance of training objectives for each branch, effectively solving the problems of imbalanced training objectives and overfitting in existing models, and improving training stability and correction accuracy.

[0082] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a meteorological numerical forecast correction device based on the EPFR multi-branch framework. These device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, this meteorological numerical forecast correction device based on the EPFR multi-branch framework can be specifically applied to various electronic devices.

[0083] like Figure 4As shown, a meteorological numerical forecast correction device 400 based on the EPFR multi-branch framework in some embodiments includes: an acquisition unit 401, a data preprocessing unit 402, a partitioning unit 403, a training unit 404, and an input unit 405. The acquisition unit 401 is configured to acquire initial meteorological data corresponding to each time granularity within a preset time period to obtain an initial meteorological dataset; the data preprocessing unit 402 is configured to preprocess the initial meteorological dataset to generate a meteorological dataset; the partitioning unit 403 is configured to partition the meteorological dataset to generate a meteorological test dataset and a meteorological training sample set; the training unit 404 is configured to train an EPFR multi-branch model based on the meteorological training sample set; and the input unit 405 is configured to input the meteorological test dataset into the EPFR multi-branch model to obtain a meteorological numerical forecast correction information set.

[0084] It is understandable that the units and references recorded in the EPFR-based multi-branch framework-based meteorological numerical forecast correction device 400 are related. Figure 1 The steps described in the method correspond to each other. Therefore, the operations, characteristics, and beneficial effects described above for the method are also applicable to the meteorological numerical forecast correction device 400 based on the EPFR multi-branch framework and the units contained therein, and will not be repeated here.

[0085] The following is for reference. Figure 5 It shows a schematic diagram of the structure of an electronic device (e.g., a computing device) 500 suitable for implementing some embodiments of the present disclosure. Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0086] like Figure 5 As shown, the electronic device 500 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory 502 or a program loaded from a storage device 508 into a random access memory 503. The random access memory 503 also stores various programs and data required for the operation of the electronic device 500. The processing unit 501, the read-only memory 502, and the random access memory 503 are interconnected via a bus 504. An input / output interface 505 is also connected to the bus 504.

[0087] Typically, the following devices can be connected to the input / output interface 505: input devices 506 including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 507 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 508 including, for example, magnetic tape, hard disk, etc.; and communication devices 509. Communication device 509 allows electronic device 500 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 An electronic device 500 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 5 Each box shown can represent a device or multiple devices as needed.

[0088] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 509, or installed from a storage device 508, or installed from a read-only memory 502. When the computer program is executed by the processing device 501, it performs the functions defined above in the methods of some embodiments of this disclosure.

[0089] It should be noted that, in some embodiments of this disclosure, the computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0090] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0091] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: acquire initial meteorological data corresponding to each time granularity within a preset time period to obtain an initial meteorological dataset; perform data preprocessing on the initial meteorological dataset to generate a meteorological dataset; partition the meteorological dataset to generate a meteorological test dataset and a meteorological training sample set; train an EPFR multi-branch model based on the meteorological training sample set; and input the aforementioned meteorological test dataset into the aforementioned EPFR multi-branch model to obtain a meteorological numerical forecast correction information set.

[0092] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0093] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0094] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0095] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for correcting meteorological numerical forecasts based on the EPFR multi-branch framework, characterized in that, include: Obtain the initial meteorological data corresponding to each time granularity within a preset time period to obtain the initial meteorological dataset; The initial meteorological dataset is preprocessed to generate a new meteorological dataset. The meteorological dataset is divided to generate a meteorological test dataset and a meteorological training sample set; Based on the meteorological training sample set, train the EPFR multi-branch model; The meteorological test dataset is input into the EPFR multi-branch model to obtain a meteorological numerical forecast correction information set; The step of training the EPFR multi-branch model based on the meteorological training sample set includes: Obtain the sample label corresponding to each meteorological training sample in the meteorological training sample set to obtain the sample label set; The initial EPFR multi-branch model is determined, which includes: an initial backbone network, an initial probability-weighted fusion branch network, an initial direct regression prediction branch network, an initial adaptive fusion branch network, and an initial residual correction branch network. Select target meteorological training samples from the meteorological training sample set; The selected target meteorological training samples are input into the initial backbone network to obtain the initial spatial feature information set; The initial set of spatial feature information is input into the initial probability weighted fusion branch network to obtain the initial weighted fusion information; The initial spatial feature information set is input into the initial direct regression prediction branch network to obtain the initial correction information; The initial weighted fusion information and the initial correction information are input into the initial adaptive fusion branch network to obtain the initial adaptive fusion information; The initial adaptive fusion information is input into the initial residual correction branch network to obtain the initial prediction correction information; Based on a preset loss function, the difference between the initial forecast correction information and the sample labels corresponding to the target meteorological training samples is determined. In response to the determination that the difference value meets the preset adjustment conditions, the network parameters of the initial EPFR multi-branch model are adjusted.

2. The meteorological numerical forecast correction method based on the EPFR multi-branch framework according to claim 1, characterized in that, The step of acquiring the initial meteorological data corresponding to each time granularity within a preset time period includes: Determine the initial field and the initial perturbation field set; The initial perturbation field set is superimposed on the initial field to obtain the initial perturbation field set; The initial set of disturbance fields is subjected to time integration to generate a set of members; The set of members of the set is determined as the initial meteorological data.

3. The meteorological numerical forecast correction method based on the EPFR multi-branch framework according to claim 1, characterized in that, The step of preprocessing the initial meteorological dataset to generate a meteorological dataset includes: Each initial meteorological data in the initial meteorological dataset is cleaned to generate cleaned meteorological data, thus obtaining a cleaned meteorological dataset. Each meteorological cleansing data point in the meteorological cleansing dataset is normalized to generate meteorological data, thus obtaining the meteorological dataset.

4. The meteorological numerical forecast correction method based on the EPFR multi-branch framework according to claim 1, characterized in that, The method further includes: In response to the determination that the difference value does not meet the preset adjustment conditions, the initial EPFR multi-branch model is determined to be an EPFR multi-branch model.

5. The meteorological numerical forecast correction method based on the EPFR multi-branch framework according to claim 1, characterized in that, The preset loss function is: ， in, Indicates the difference value. , , , Indicates weight, Indicates PEW loss, Indicates DDF loss. Indicates the loss of AB. Represents the RCF loss, where, ， in, , Indicates weight, This represents the cross-entropy loss for multi-class classification. This represents the regression loss due to the average error. Indicates the sample number. , Represents the total number of samples. Indicates the member sequence number. , Indicates the number of members in the set. Indicates the first In the nth sample Smoothed labels of set members Indicates the first In the nth sample The weighted probabilities of each set member This represents the initial weighted fused information. Indicates sample label, Indicates one-hot tag, Represents the smoothing coefficient. Indicates the serial number. Indicates the first In the nth sample The logits value of each set member. Indicates the first In the nth sample The logits value of each set member.

6. The meteorological numerical forecast correction method based on the EPFR multi-branch framework according to claim 1, characterized in that, The step of inputting the meteorological test dataset into the EPFR multi-branch model to obtain meteorological numerical forecast correction information includes: For each meteorological test data in the meteorological test dataset, the meteorological test data is input into the EPFR multi-branch model to obtain meteorological numerical forecast correction information; The obtained meteorological numerical forecast correction information is determined as the meteorological numerical forecast correction information set.

7. A meteorological numerical forecast correction device based on the EPFR multi-branch framework, characterized in that, include: The acquisition unit is configured to acquire the initial meteorological data corresponding to each time granularity within a preset time period to obtain the initial meteorological dataset. A data preprocessing unit is configured to preprocess the initial meteorological dataset to generate a meteorological dataset; The partitioning unit is configured to partition the meteorological dataset to generate a meteorological test dataset and a meteorological training sample set. The training unit is configured to train the EPFR multi-branch model based on the meteorological training sample set; The training unit is further configured as follows: Obtain the sample label corresponding to each meteorological training sample in the meteorological training sample set to obtain the sample label set; The initial EPFR multi-branch model is determined, which includes: an initial backbone network, an initial probability-weighted fusion branch network, an initial direct regression prediction branch network, an initial adaptive fusion branch network, and an initial residual correction branch network. Select target meteorological training samples from the meteorological training sample set; The selected target meteorological training samples are input into the initial backbone network to obtain the initial spatial feature information set; The initial set of spatial feature information is input into the initial probability weighted fusion branch network to obtain the initial weighted fusion information; The initial spatial feature information set is input into the initial direct regression prediction branch network to obtain the initial correction information; The initial weighted fusion information and the initial correction information are input into the initial adaptive fusion branch network to obtain the initial adaptive fusion information; The initial adaptive fusion information is input into the initial residual correction branch network to obtain the initial prediction correction information; Based on a preset loss function, the difference between the initial forecast correction information and the sample labels corresponding to the target meteorological training samples is determined. In response to the determination that the difference value meets the preset adjustment conditions, the network parameters of the initial EPFR multi-branch model are adjusted. The input unit is configured to input the meteorological test dataset into the EPFR multi-branch model to obtain a meteorological numerical forecast correction information set.

8. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 6.

9. A computer-readable medium, characterized in that, It stores a computer program thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images