A multi-model ensemble post-processing method and system for Chinese regional sub-seasonal prediction
By using a multi-model ensemble post-processing system and leveraging FUNet deep learning and dynamic weighted fusion strategies, the resolution and stability issues of sub-seasonal forecasts in China's regional applications were resolved, achieving high-precision, high-resolution weather forecasts.
Patent Information
- Application Number
- CN202610797738.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-04
- Publication Date
- 2026-08-25
AI Technical Summary
Existing subseasonal forecasting methods suffer from low spatial resolution, large systematic errors, and insufficient local forecasting capabilities in complex terrain areas when applied in China. Single models exhibit significant differences in performance across different regions, lead times, and weather and climate backgrounds, and there is a lack of a unified multi-model collaborative training and dynamic fusion system.
A multi-model integrated post-processing system is adopted, including a data layer, a model layer, a decision layer, and an evaluation layer. The system uses the FUNet deep learning-led approach, combined with multi-mode multi-set feature fusion, terrain-aided modeling, and a dynamic weighted fusion strategy based on validation error, to output high-resolution post-processed forecast results.
It improves the accuracy and spatial resolution of forecast results, enhances the model's adaptability to different regions and weather and climate backgrounds, reduces the risk of single model failure, provides stability and robustness, and supports extreme weather response and refined meteorological services.
Smart Images
Figure CN122634553A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of meteorological forecast post-processing and artificial intelligence technology, specifically involving a multi-model integrated post-processing method and system for sub-seasonal forecasting in China. Background Technology
[0002] Subseasonal forecasts typically cover the period from the second week to the second month of the future, serving as a crucial timescale connecting weather forecasts and seasonal predictions. They have significant applications in agricultural production, energy dispatch, disaster prevention and mitigation, and public climate services. Existing subseasonal numerical model ensemble forecasts can provide some large-scale predictable signals, but due to limitations in model resolution, accumulated systematic errors, and the influence of complex terrain, their raw outputs are often insufficient to directly meet the needs of refined meteorological services in China.
[0003] Currently, commonly used post-processing methods mainly include interpolation methods, traditional statistical methods, machine learning methods, and deep learning methods. Interpolation methods can only achieve spatial resampling and cannot effectively correct model systematic errors; traditional statistical methods have a certain degree of interpretability, but their ability to express multi-source high-dimensional grid fields and complex nonlinear relationships is limited; although machine learning methods can handle some nonlinear mappings, they still have shortcomings in multivariable grid fields, spatial structure preservation, and high-resolution detail recovery; deep learning methods can establish complex mapping relationships between coarse-resolution forecast fields and high-resolution observation ground truth fields end-to-end, but single deep learning models often have inconsistent stability in different regions, different lead times, and different weather and climate backgrounds, making it difficult to fully adapt to the complex scenarios in China.
[0004] On the other hand, in practical applications of sub-seasonal forecast post-processing in China, data sources are characterized by multiple models, multiple ensembles, multiple resolutions, and multiple types. These include raw coarse-resolution sub-seasonal model ensemble forecast data, as well as dynamic downscaling results, high-resolution gridded observation data, station observations, and topographic-aided data. The key challenge in achieving a high-precision, stable, and scalable post-processing system lies in how to accomplish multi-source data acquisition, unified sample construction, multi-model training, multi-model multi-ensemble prediction, dynamic result fusion, and operational product output within a unified framework.
[0005] In summary, the existing technology has the following problems:
[0006] First, the original sub-seasonal model ensemble forecast products have low spatial resolution, large systematic errors, and insufficient local forecasting capabilities in complex terrain areas.
[0007] Second, single statistical learning, machine learning or deep learning models show significant differences in performance across different regions, different lead times and different meteorological backgrounds, making it difficult to balance accuracy and stability.
[0008] Third, existing methods lack a unified post-processing system that integrates data preprocessing, multi-model collaborative training, dynamic fusion, and product output. Summary of the Invention
[0009] Purpose of the invention: This invention proposes a multi-model integrated post-processing method and system for sub-seasonal forecasting in China, which solves the problems of insufficient forecasting capability and difficulty in balancing accuracy and stability of existing post-processing results. Using FUNet as the deep learning-led method, it outputs high-resolution post-processing forecast results through multi-model multi-set feature fusion, terrain-assisted modeling, and dynamic weighted fusion strategy based on validation error.
[0010] Technical solution: A multi-model integrated post-processing system for sub-seasonal forecasting in China, comprising:
[0011] The data layer is responsible for collecting data, processing the collected data in a unified manner, and building training and validation sets.
[0012] The model layer uses training and validation sets to train and predict various models in the model library, which includes statistical learning models, machine learning models, and deep learning models.
[0013] The decision-making layer integrates the output of the FUNet post-processing model with the outputs of the statistical learning model, the machine learning model, and the deep learning model through a dynamic weighted fusion module.
[0014] The evaluation layer is used to calculate the evaluation index of each model in the model library; the evaluation results are used as the basis for the decision layer to set dynamic weighting parameters, and are used to comprehensively evaluate the overall accuracy, system bias and anomaly characterization ability of the correction results of the dynamic weighted fusion module.
[0015] The application output layer, after determining the optimal weighting parameters for each region and lead time, outputs corresponding post-processing forecast results according to user needs; the output content includes high-resolution grid correction products, regional statistical results, and visualization products for business applications.
[0016] A multi-model ensemble post-processing method for sub-seasonal forecasts in China, based on the aforementioned multi-model ensemble post-processing system, generates high-resolution gridded forecast products, including the following steps:
[0017] S1, acquire multi-model multi-ensemble data, including raw sub-seasonal ensemble re-forecast data, dynamic downscaling results data, high-resolution observation true data, topographic auxiliary data and station observation supplementary data;
[0018] S2 processes multi-mode, multi-set data to obtain coarse-resolution input samples, high-resolution input samples, and corresponding target output samples that can be used by the model layer.
[0019] S3, after training the constructed FUNet post-processing model, constructs a dynamic weighted fusion module consisting of a statistical learning model, a machine learning model, and a deep learning model;
[0020] S4, the dynamic weighted fusion module uses the comprehensive verification error to calculate the comprehensive error of each model at the partition position; then it updates the fusion parameters based on the comprehensive error, and determines the fusion weights based on the adaptive inverse error weighting method based on the comprehensive verification error, and outputs the final fusion prediction result.
[0021] Furthermore, in step S2, after completing time alignment and spatial processing for the multi-mode, multi-set data, using the reporting start date, lead time, and set members as indices, a coarse-resolution input sample is constructed for the branch with coarse resolution as input:
[0022] ,
[0023] in, This indicates that the China Meteorological Administration's sub-seasonal ensemble reforecast data is in effect from the start date. Lead time Collection members The coarse-resolution input field below; This represents the coarse-resolution input field corresponding to the European Centre for Medium-Range Weather Forecasts (ECMWF) sub-seasonal ensemble reforecast data. This represents the terrain auxiliary information corresponding to the coarse-resolution input;
[0024] For branches that take high-resolution input, construct high-resolution input samples:
[0025] ,
[0026] in, This represents the high-resolution CWRF input field after backprojection and resampling. This represents terrain-aided information corresponding to the high-resolution input;
[0027] The corresponding high-resolution target output sample is denoted as: ;
[0028] in, This represents the true high-resolution observation of the target obtained from the grid field construction of CN05.1, CLDAS, or the site.
[0029] Furthermore, the FUNet post-processing model consists of a multi-mode multi-ensemble feature fusion downscaling module, a multi-scale feature extraction encoder, and a sub-pixel convolutional decoder. The model input includes real high-resolution topographic map data and low-resolution atmospheric factor data predicted by multi-mode ensembles.
[0030] The multimodal multi-ensemble feature fusion scaling module first performs sub-pixel convolution × 12 upscaling on the sub-seasonal ensemble forecast data in the low-resolution atmospheric factor data, then fuses it with the dynamic downscaling data to form fused forecast data, and further inputs it into the atmospheric factor fusion module along with real terrain data; after channel attention calibration, it is added to the residual of the fused forecast data, and then outputs the high-resolution atmospheric features after feature fusion through sub-pixel convolution × 2;
[0031] The high-resolution atmospheric features after feature fusion are sequentially fed into a multi-scale feature extraction encoder and a sub-pixel convolutional decoder, and finally output a high-resolution field after bias correction.
[0032] Furthermore, the atmospheric factor fusion module divides the input atmospheric factors into a near-surface thermal branch, a surface layer variable branch, and a pressure layer variable branch. The near-surface thermal branch includes 2-meter temperature, actual topography, and 2-meter dew point temperature. The surface layer variable branch includes surface meteorological variables. The pressure layer variable branch includes pressure layer atmospheric variables. Each branch has its features extracted by the multi-scale feature extraction module, then the features are spliced in the channel dimension, and finally fused by the multi-scale feature extraction module again.
[0033] Furthermore, the multi-scale feature extraction encoder achieves multi-scale feature extraction, channel recalibration, and layer-by-layer downsampling by incorporating batch normalization, ReLU activation, 1×1 convolution, channel attention module, and two-dimensional max pooling operations.
[0034] The subpixel convolutional decoder achieves layer-by-layer high-resolution restoration through two-dimensional convolution, pixel recombination upsampling, and ReLU activation, and completes local detail reconstruction by skip connections with the corresponding layers of the multi-scale feature extraction encoder.
[0035] Furthermore, the first A model in the region Lead time Comprehensive verification error for:
[0036] ,
[0037] in, , and , which are non-negative weighting coefficients used to balance the impact of different evaluation indicators on the fusion decision; RMSE is the root mean square error, Bias is the average bias, and ACC is the outlier correlation coefficient.
[0038] No. A model in the region Lead time Fusion weights Defined as:
[0039] ,
[0040] According to the fusion weight , obtain the area Lead time The final fusion output :
[0041] ,
[0042] in, Indicates the total number of candidate models; This represents the weighted temperature parameter, used to adjust the sensitivity of weight allocation.
[0043] Compared with the prior art, the significant advantages of this invention are as follows:
[0044] 1. This invention addresses the problems of low spatial resolution, large systematic errors, and insufficient local information representation in complex terrain areas in sub-seasonal forecast products for China. It constructs a complete post-processing workflow from multi-source data acquisition, unified preprocessing, model training, and dynamic fusion of multi-model results output, realizing integrated processing of bias correction and spatial refinement, thereby simultaneously improving the accuracy and spatial resolution of forecast results.
[0045] 2. This invention effectively utilizes the complementary information between different models and ensemble members by uniformly collecting, filtering, aligning and integrating multi-model and multi-ensemble subseasonal forecast data from the China Meteorological Administration (CMA) and the European Centre for Medium-Range Weather Forecasts (ECMWF), significantly improving the quantity and quality of training samples and enhancing the model's adaptability and generalization ability to different regions, different lead times and different weather and climate backgrounds.
[0046] 3. This invention constructs a multi-model integrated post-processing system that integrates statistical learning, machine learning, and deep learning. The preferred deep learning model is the FUNet model, an improvement on U-Net. Through a multi-mode, multi-set feature fusion downscaling module, an atmospheric factor fusion module, a multi-scale feature extraction module, and a sub-pixel convolutional upsampling mechanism, high-resolution reconstruction and system error correction of the original coarse-resolution model forecast field are achieved. Compared with existing single-model methods, this invention not only enhances the local rationality and detail recovery capabilities of post-processing results in complex terrain areas but also better preserves high-frequency spatial structure and reduces blurring and artifacts.
[0047] 4. This invention adopts a dynamic weighted fusion mechanism based on verification error. According to the historical verification performance of different models in different regions and different lead times, the fusion weights of statistical learning results, machine learning results, deep learning results and ensemble average results are adaptively determined. This significantly improves the overall accuracy, stability and robustness of the system output, reduces the risk of a single model failing in local areas or specific samples, and provides more reliable technical support for extreme weather response, refined meteorological services and disaster risk management. Attached Figure Description
[0048] Figure 1 This is an architecture diagram of the post-processing system of the present invention;
[0049] Figure 2 This is a flowchart of the post-processing method of the present invention;
[0050] Figure 3 The network diagram of the FUNet model;
[0051] Figure 4 This is a structural diagram of a multi-mode, multi-set feature fusion downscaling module;
[0052] Figure 5 This is a structural diagram of the atmospheric factor fusion module;
[0053] Figure 6 Here is a structural diagram of the multi-scale feature extraction module;
[0054] Figure 7 A diagram of the encoder structure for multi-scale feature extraction;
[0055] Figure 8 This is a diagram of the subpixel convolutional decoder structure;
[0056] Figure 9 Comparison chart of the effects of this invention and the benchmark model;
[0057] Figure 10 A comparison chart of the partitioning effects of this invention and a benchmark model. Detailed Implementation
[0058] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0059] like Figure 1 As shown, a multi-model integrated post-processing system for sub-seasonal forecasting in China includes:
[0060] The data layer is responsible for collecting data, processing all collected data in a unified manner, and constructing training and validation sets.
[0061] The model layer uses samples processed in the data layer for model training and prediction. The model library for the model layer can include statistical learning models, machine learning models, and deep learning models. Statistical learning models can employ methods such as QM and QDM; machine learning models can employ methods such as Random Forest (RF) and LightGBM; and for deep learning models, in addition to using ordinary U-Net, the FUNet proposed in this invention is preferred as the system's leading model. FUNet is mainly used to complete the deep fusion of multi-model and multi-ensemble forecast information, high-resolution spatial detail reconstruction, and bias correction of meteorological element fields. It is the core implementation method in the system responsible for "high-precision gridded field post-processing."
[0062] The decision-making layer integrates the outputs of the dominant model FUNet with those of other statistical learning models (such as QM and QDM), machine learning models (such as RF and LightGBM), and deep learning models (such as U-Net) through a dynamically weighted fusion mechanism. The fusion weights are adaptively determined based on the loss performance on the validation set for different regions and lead times, thereby achieving optimal result fusion that addresses regional and time-sensitivity differences.
[0063] The evaluation layer is used to calculate evaluation metrics such as root mean square error (RMSE), bias, and anomaly correlation coefficient (ACC) for each model in the model library. On the one hand, the evaluation results can serve as an important basis for the decision-making layer to set dynamic weighting parameters; on the other hand, they can be used to comprehensively evaluate the overall accuracy, system bias, and anomaly characterization capabilities of the correction results after the system's fusion decision.
[0064] The application output layer, after determining the optimal weighting parameters for each region and lead time, outputs corresponding post-processing forecast results according to user needs. Output content may include high-resolution grid correction products, regional statistical results, and visualization products geared towards business applications.
[0065] This invention provides a multi-model ensemble post-processing method for sub-seasonal forecasting in China, such as... Figure 2 As shown, this method takes raw sub-seasonal model ensemble forecast data, dynamic downscaling results, and terrain-aided data as input. Through unified data preprocessing, collaborative training of statistical learning models, machine learning models, and deep learning models, and dynamic weighted fusion based on validation errors, it generates high-resolution gridded forecast products. This method is suitable for bias correction, spatial refinement, and operational post-processing of meteorological elements such as temperature. The steps include the following:
[0066] Step 1: Acquisition of multi-mode, multi-set data;
[0067] The method of this invention is designed for post-processing tasks of sub-seasonal forecasts in China. The required datasets can be obtained from the official websites of the China Meteorological Administration and the European Centre for Weather and Geosciences. The data used mainly include the following categories:
[0068] (A1) Original sub-seasonal ensemble reforecast data;
[0069] The data primarily includes China Meteorological Administration's subseasonal ensemble recast data (CMA S2S reforecast) and the European Centre for Medium-Range Weather Forecasts (ECMWF S2S reforecast). Both types of data serve as the system's raw large-scale forecast inputs, providing subseasonal atmospheric background and ensemble forecast information.
[0070] (A2) Dynamic downscaling results data;
[0071] This mainly includes CWRF (Climate-Weather Research and Forecasting model) dynamic downscaling data obtained by further regional dynamic downscaling based on CMA S2S re-prediction. This type of data contains more detailed spatial structure information compared to the original coarse-resolution ensemble re-prediction.
[0072] (A3) High-resolution observational true data;
[0073] This primarily includes high-resolution land surface assimilation gridded data (CLDAS) from the China Meteorological Administration's Land Surface Assimilation System or a 0.25°×0.25° high-resolution gridded observation dataset for the Chinese region (CN05.1 version). This type of data serves as the target output for the supervised learning phase of the system, used to construct input-output sample pairs.
[0074] (A4) Terrain-aided data;
[0075] This includes high-resolution topographic map data of China. This type of data serves as a static auxiliary factor to enhance the FUNET model's ability to express the meteorological distribution characteristics of complex topographic regions.
[0076] (A5) Supplementary data from site observations (optional);
[0077] When gridded observation products have insufficient accuracy in some areas, or when there are missing measurements for the target date, discrete station observation data can be further introduced and converted into a regular gridded field using the SetConv method to supplement the true value of the high-resolution target.
[0078] Step 2: Multi-mode, multi-set data processing;
[0079] First, time alignment and sample alignment are performed. Firstly, the seasonal ensemble recast data from the China Meteorological Administration (CMAS2S reforecast) and the European Centre for Medium-Range Weather Forecasts (ECMWF S2S reforecast) are time-filtered, retaining only samples corresponding to the same start date to ensure consistency of inputs from different models in the time dimension. Then, according to the requirements of the target mission, weekly average samples or pentad average samples are constructed from the retained samples, respectively. Let the... The time window corresponding to each lead time is Then, the average sample of a certain input variable over that time window can be expressed as:
[0080] (1)
[0081] in, Indicates the reporting start date as The forecast sample in the first The variable field corresponding to each forecast day The first one obtained by construction Weekly average or time-average sample.
[0082] For CWRF dynamic downscaling data, the data is filtered and time-aligned according to the same reporting start date as CMA S2S data, and then the corresponding weekly or pentad average samples are constructed. For high-resolution ground truth observation data such as CLDAS, the same lead time window as the input data is used for averaging, thus forming an observational learning target field corresponding to the input. When high-resolution gridded observations have insufficient accuracy in some areas, or when there are missing measurements on the target date, discrete station observation data can be further introduced to supplement them. Station observations are converted into a regular gridded field using the SetConv method, which is used to construct or supplement the target output samples for learning. Its form can be expressed as:
[0083] (2)
[0084] in, Indicates the location of the target grid point. Indicates the first Location of each observation station This represents the total number of observation stations participating in the interpolation. Indicates the corresponding observation value. This represents a kernel function with learnable parameters. To prevent extremely small positive numbers with a denominator of zero.
[0085] After completing time alignment and spatial processing, training samples are further organized according to unified rules. This process does not directly merge all source data into a single file; instead, it uses the reporting start date, lead time, and set members as indices to create separate input samples for different source data, and calls them according to the corresponding branches during the model input stage. For the coarse resolution input branch, the reporting start date is used as the index. Lead time and collection members As an index, coarse-resolution input samples can be constructed:
[0086] (3)
[0087] in, This indicates that the China Meteorological Administration's sub-seasonal ensemble reforecast data is in effect from the start date. Lead time Collection members The coarse resolution input field below, This represents the corresponding coarse-resolution input field of the European Centre for Medium-Range Weather Forecasts (ECMWF) sub-seasonal ensemble reforecast data. This represents the terrain auxiliary information corresponding to the coarse-resolution input. For the high-resolution input branch, a high-resolution input sample can be constructed:
[0088] (4)
[0089] in, This represents the high-resolution CWRF input field after backprojection and resampling. This represents terrain-aided information corresponding to the high-resolution input.
[0090] The corresponding high-resolution target output sample is denoted as , can represent the true high-resolution observation of the target obtained by constructing a grid field supplemented by CN05.1, CLDAS, or the site.
[0091] After organizing the samples, data from different sources and with different variables are normalized to reduce the impact of dimensional differences on model training and accelerate model convergence. The normalization parameters are obtained statistically from the training set and are consistently used in the validation and test sets. The final samples available for use by the system model layer include coarse-resolution input samples, high-resolution input samples, and corresponding target output samples.
[0092] Step 3: Train the deep learning post-processing model based on FUNet;
[0093] The FUNet post-processing model consists of three modules: a multi-model, multi-ensemble feature fusion and downscaling module, a multi-scale feature extraction encoder, and a sub-pixel convolutional decoder. The model input includes real high-resolution topographic map data and low-resolution atmospheric factor data predicted by multi-model ensembles. The overall network structure is as follows: Figure 3 As shown.
[0094] First, the low-resolution atmospheric factors obtained from the real high-resolution topographic map and multimodal ensemble forecast, i.e., those processed in step two, are combined. and The input is then fed into the multimodal, multi-set feature fusion scaling module. This module, as shown... Figure 4 As shown, the sub-seasonal ensemble forecast data in the low-resolution atmospheric factor data from the multi-model ensemble forecast is first upscaled by subpixel convolution × 12, then fused with dynamical downscaling data to form fused forecast data, which is further input into the atmospheric factor fusion module along with real topographic data. After channel attention calibration, it is added to the residual of the fused forecast data, and then the high-resolution atmospheric features after feature fusion are output through subpixel convolution × 2.
[0095] The atmospheric factor fusion module structure included in the multi-mode multi-ensemble feature fusion downscaling module is as follows: Figure 5 As shown, the atmospheric factor fusion module divides the input atmospheric factors into near-surface thermal branches, surface layer variable branches, and pressure layer variable branches. The near-surface thermal branch includes 2-meter temperature, actual topography, and 2-meter dew point temperature; the surface layer variable branch includes surface meteorological variables; and the pressure layer variable branch includes pressure layer atmospheric variables. Each branch's features are extracted by a multi-scale feature extraction module, then concatenated along the channel dimension, and finally fused again by the multi-scale feature extraction module. The multi-scale feature extraction module is shown below. Figure 6 As shown, the multi-scale feature extraction module uses three parallel branches: 1×1 convolution, 3×3 convolution, and 5×5 convolution to extract atmospheric features from different receptive fields and splice them in the channel dimension to form a multi-scale fusion representation.
[0096] After the downscaling process of the multi-mode, multi-set feature fusion downscaling module is completed, the system proceeds to the multi-scale feature extraction encoder and sub-pixel convolutional decoder structure. The encoder, for example... Figure 7 The multi-scale feature extraction encoder shown achieves multi-scale feature extraction, channel recalibration, and layer-by-layer downsampling by further incorporating batch normalization, ReLU activation, 1×1 convolution, channel attention modules, and two-dimensional max pooling operations. The sub-pixel convolutional decoder is as follows: Figure 8The subpixel convolutional decoder shown achieves layer-by-layer high-resolution restoration through 2D convolution, pixel recombination upsampling, and ReLU activation, and combines this with skip connections to the corresponding levels of the encoder to complete local detail reconstruction. Finally, FUNet outputs a high-resolution field after bias correction.
[0097] During training, the corresponding high-resolution ground truth observations at that time. As a supervisory signal (i.e., label data), Charbonnier Loss is used as the primary objective function. :
[0098] (5)
[0099] in, Represents the true value of high-resolution observations. This indicates the model output result. is the smoothing constant; N represents the number of samples.
[0100] Step 4: Based on the constructed multi-model library, a dynamic weighted fusion module is used to calculate the comprehensive error of each model at the partition location, and the fusion parameters are updated according to the comprehensive error;
[0101] After completing multi-source data preprocessing, unified sample construction, and training of the main model FUNet, the system enters the multi-model collaborative inference and result fusion stage. To improve the stability and adaptability of the post-processing results of the next season's forecast in different regions and with different lead times, this invention does not rely on the output of a single model, but instead simultaneously constructs a candidate result set composed of the results of statistical learning models, machine learning models, and deep learning models.
[0102] Candidate models for each model include: quantile mapping models (QM) and quantile gain mapping models (QDM) in statistical learning models; random forest models (RF) and LightGBM models in machine learning models; U-Net models in deep learning models; and the FUNet post-processing model as the dominant model. If necessary, the ensemble average of the original forecasts can also be used as a baseline candidate. For the same input sample, after inference using the above different models, multiple sets of high-resolution correction results can be obtained, denoted as... ,in, Indicates the candidate model number. Indicates the area code. Indicates the lead time number. Indicates the first A model in the region Lead time The correction results are output above.
[0103] This invention does not simply average the candidate results; instead, it constructs a dynamic fusion mechanism that adapts to regional and temporal factors based on the historical performance of each model on the validation set. Specifically, on the validation set, the error performance of each candidate model is statistically analyzed in different regions and with different lead times. Evaluation metrics include Root Mean Square Error (RMSE), Mean Bias, and Anomaly Correlation Coefficient (ACC). RMSE characterizes the overall error magnitude, Bias characterizes whether the system is excessively high or low, and ACC characterizes the consistency of anomaly patterns. Further, the following is defined: A model in the region Lead time The overall verification error is:
[0104] (6)
[0105] in, , and These are non-negative weighting coefficients used to balance the impact of different evaluation metrics on the fusion decision. This approach simultaneously constrains the model's absolute error, system bias, and ability to characterize anomalous structures, thus avoiding the bias caused by relying solely on a single error metric.
[0106] Based on the comprehensive verification error, this invention further employs an adaptive inverse error weighting method to determine the fusion weights. A model in the region Lead time Fusion weights Defined as:
[0107] (7)
[0108] Based on the above weights, the region can be obtained. Lead time The final fusion output :
[0109] (8)
[0110] in, This represents the total number of candidate models. This represents the weight temperature parameter, used to adjust the sensitivity of weight allocation. The smaller the model validation error, the larger the corresponding fusion weight; when multiple models perform similarly, the weights of each model tend to be balanced.
[0111] Step 5: Post-processing effect evaluation;
[0112] After completing the dynamic fusion of multiple models, the system enters the evaluation layer to uniformly evaluate the fusion results and the results of each candidate model. The evaluation layer is mainly used for two purposes: firstly, to provide a verification basis for the dynamic weighting in step four; secondly, to comprehensively verify the overall accuracy, system bias, and ability to characterize abnormal field patterns of the final output product.
[0113] The evaluation layer uses RMSE, Bias, and ACC as core evaluation indicators, and can be expanded to auxiliary indicators such as MAE and PCC as needed. The evaluation can be conducted nationwide or in separate regions across the country to analyze the differences in the system's applicability in different regional contexts.
[0114] In one embodiment of the present invention, based on the main model FUNet, a multi-model library is constructed by selecting the statistical learning method QM, the machine learning method LightGBM, and the deep learning method UNet. The test set samples from July to August are divided into regions according to Northeast, North China, East China, Central China, South China, Southwest, Northwest, and the Qinghai-Tibet Plateau, and the average results for the first six pentads are evaluated. Figure 9 As shown, the national average RMSE, MAE, and ACC of the dynamic weighted fusion system are 2.4875 K, 1.8287 K, and 0.6535, respectively; the corresponding indices for the dominant model FUNet are 2.5515 K, 1.8984 K, and 0.6347; and the corresponding indices for the U-Net / CWRF benchmark model are 2.7463 K, 2.0680 K, and 0.5686. This demonstrates that the dynamic weighted fusion mechanism proposed in this invention can further reduce the overall error and improve the consistency of anomalous field patterns based on a single model.
[0115] At the same time, in the regional assessment, such as Figure 10 As shown, the RMSE of the system proposed in this invention is generally lower than that of the single model in most regions, including Northeast China, North China, East China, Central China, South China, Southwest China, Northwest China, and the Qinghai-Tibet Plateau, while the ACC also remains at a relatively high level. This indicates that the multi-model dynamic fusion mechanism proposed in this invention can provide more robust correction results in complex terrain areas and situations with significant regional differences.
[0116] Step Six: Results Output and Business Applications;
[0117] After obtaining the fusion weights for different regions and lead times, the system enters the application output layer, generating corresponding post-processing forecast products according to user needs. The output includes: high-resolution gridded temperature correction products for the Chinese region; statistical evaluation results for different regions and lead times; and comparative analysis products of the fusion system and single-model results.
Claims
1. A multi-model integrated post-processing system for sub-seasonal forecasting in China, characterized in that, include: The data layer is responsible for collecting data, processing the collected data in a unified manner, and building training and validation sets. The model layer uses training and validation sets to train and predict various models in the model library, which includes statistical learning models, machine learning models, and deep learning models. The decision-making layer integrates the output of the FUNet post-processing model with the outputs of the statistical learning model, the machine learning model, and the deep learning model through a dynamic weighted fusion module. The evaluation layer is used to calculate the evaluation index of each model in the model library; the evaluation results are used as the basis for the decision layer to set dynamic weighting parameters, and are used to comprehensively evaluate the overall accuracy, system bias and anomaly characterization ability of the correction results of the dynamic weighted fusion module. The application output layer, after determining the optimal weighting parameters for each region and lead time, outputs corresponding post-processing forecast results according to user needs; the output content includes high-resolution grid correction products, regional statistical results, and visualization products for business applications.
2. A multi-model ensemble post-processing method for sub-seasonal forecasts in China, based on the multi-model ensemble post-processing system as described in claim 1, to generate high-resolution gridded forecast products, characterized in that... Including the following steps: S1, acquire multi-model multi-ensemble data, including raw sub-seasonal ensemble re-forecast data, dynamic downscaling results data, high-resolution observation true data, topographic auxiliary data and station observation supplementary data; S2 processes multi-mode, multi-set data to obtain coarse-resolution input samples, high-resolution input samples, and corresponding target output samples that can be used by the model layer. S3, after training the constructed FUNet post-processing model, constructs a dynamic weighted fusion module consisting of a statistical learning model, a machine learning model, and a deep learning model; S4, the dynamic weighted fusion module uses the comprehensive verification error to calculate the comprehensive error of each model at the partition position; then it updates the fusion parameters based on the comprehensive error, and determines the fusion weights based on the adaptive inverse error weighting method based on the comprehensive verification error, and outputs the final fusion prediction result.
3. The multi-model ensemble post-processing method for sub-seasonal forecasting in China according to claim 2, characterized in that, In step S2, after completing time alignment and spatial processing for the multi-mode, multi-set data, coarse-resolution input samples are constructed for branches that take coarse resolution as input, using the reporting start date, lead time, and set members as indices: ， in, This indicates that the China Meteorological Administration's sub-seasonal ensemble reforecast data is in effect from the start date. Lead time Collection members The coarse resolution input field below; This represents the coarse-resolution input field corresponding to the European Centre for Medium-Range Weather Forecasts (ECMWF) sub-seasonal ensemble reforecast data. This represents the terrain auxiliary information corresponding to the coarse-resolution input; For branches that take high-resolution input, construct high-resolution input samples: ， in, This represents the high-resolution CWRF input field after backprojection and resampling. This represents terrain-aided information corresponding to the high-resolution input; The corresponding high-resolution target output sample is denoted as: ; in, This represents the true high-resolution observation of the target obtained from the grid field construction of CN05.1, CLDAS, or the site.
4. The multi-model ensemble post-processing method for sub-seasonal forecasting in China according to claim 2, characterized in that, The FUNet post-processing model consists of a multi-mode multi-ensemble feature fusion downscaling module, a multi-scale feature extraction encoder, and a sub-pixel convolutional decoder. The model input includes real high-resolution topographic map data and low-resolution atmospheric factor data predicted by multi-mode ensembles. The multimodal multi-ensemble feature fusion scaling module first performs sub-pixel convolution × 12 upscaling on the sub-seasonal ensemble forecast data in the low-resolution atmospheric factor data, then fuses it with the dynamic downscaling data to form fused forecast data, and further inputs it into the atmospheric factor fusion module along with real terrain data; after channel attention calibration, it is added to the residual of the fused forecast data, and then outputs the high-resolution atmospheric features after feature fusion through sub-pixel convolution × 2; The high-resolution atmospheric features after feature fusion are sequentially fed into a multi-scale feature extraction encoder and a sub-pixel convolutional decoder, and finally output a high-resolution field after bias correction.
5. The multi-model ensemble post-processing method for sub-seasonal forecasting in China according to claim 4, characterized in that, The atmospheric factor fusion module divides the input atmospheric factors into a near-surface thermal branch, a surface layer variable branch, and a pressure layer variable branch. The near-surface thermal branch includes 2-meter temperature, actual topography, and 2-meter dew point temperature. The surface layer variable branch includes surface meteorological variables. The pressure layer variable branch includes pressure layer atmospheric variables. Each branch extracts features through a multi-scale feature extraction module, which then splices the features in the channel dimension and performs a unified fusion through another multi-scale feature extraction module.
6. The multi-model ensemble post-processing method for sub-seasonal forecasting in China according to claim 4, characterized in that, The multi-scale feature extraction encoder achieves multi-scale feature extraction, channel recalibration, and layer-by-layer downsampling by incorporating batch normalization, ReLU activation, 1×1 convolution, channel attention module, and two-dimensional max pooling operations. The subpixel convolutional decoder achieves layer-by-layer high-resolution restoration through two-dimensional convolution, pixel recombination upsampling, and ReLU activation, and completes local detail reconstruction by skip connections with the corresponding layers of the multi-scale feature extraction encoder.
7. The multi-model ensemble post-processing method for sub-seasonal forecasting in China according to claim 2, characterized in that, No. A model in the region Lead time The overall verification error for: ， in, , and , which are non-negative weighting coefficients used to balance the impact of different evaluation indicators on the fusion decision; RMSE is the root mean square error, Bias is the average bias, and ACC is the outlier correlation coefficient. No. A model in the region Lead time Fusion weights Defined as: ， According to the fusion weight , obtain the area Lead time The final fusion output : ， in, Indicates the total number of candidate models; This represents the weighted temperature parameter, used to adjust the sensitivity of weight allocation.