Numerical model precipitation bias correction method based on deep learning

Abstract

The application discloses a bias correction method for numerical mode precipitation products based on deep learning, and comprises the following specific steps: obtaining ERA5 land surface reanalysis data, ECMWF daily precipitation prediction data and ground station daily precipitation observation data; selecting effective features by using the correlation between meteorological factors and ground observation precipitation values through Pearson correlation coefficients; obtaining relatively accurate daily precipitation labels; establishing a mode precipitation product bias correction model based on deep learning, designing a weight loss function to update model parameters; and training to obtain an optimal model. The application has the advantages that the deep learning model adds historical meteorological sequences, and the model is designed with a double-branch structure, which pays attention to local precipitation and takes into account overall precipitation, the weight loss function is used to solve the problems of large span and extreme imbalance of environmental precipitation, the model can well learn the precipitation error in the numerical mode, and the effect of precipitation bias correction is achieved.

Description

Technical Field

[0001] This invention relates to the field of deep learning-based precipitation forecast bias correction, specifically a deep learning-based numerical model precipitation bias correction method. Background Technology

[0002] Throughout human history, inferring and predicting weather has been a long-standing challenge. In modern weather forecasting, numerical weather prediction (NWP) models, based on physical equations and utilizing high-performance computers, hold a leading position. The European Centre for Medium-Range Weather Forecasts (ECMWF) is a world-leading NWP center. Moreover, ECMWF forecasts offer valuable reference for the accuracy of forecasts in China. However, due to numerous uncertainties in numerical simulations, cumulative errors from integration, and the averaging effect from multiple ensemble members, errors are inevitable in numerical model forecasts. Therefore, reasonable, objective, and quantitative correction is the bridge connecting numerical models and accurate forecasts, and an indispensable step in fully exploring the potential of numerical forecasting.

[0003] Currently, numerous meteorological researchers have conducted extensive research on weather forecast correction. Model error correction mainly falls into two categories: the first focuses on the model forecast itself, improving the quality of initial field data and enhancing the performance of numerical forecasts; the second is based on statistical models to correct post-processing errors in forecast results. Vannitsem and Tian demonstrated good performance on high-resolution models using the Model Output Statistics (MOS) method. Krishnamurti et al. proposed the concept of ensemble forecasts and corrected precipitation forecasts for the global tropics and the United States, while Cui et al. used adaptive Kalman filtering to correct 2-meter temperature ensemble mean forecasts.

[0004] It is worth noting that the process of meteorological evolution is random and discontinuous, and traditional statistical methods have very limited effectiveness in correcting its biases.

[0005] In recent years, deep learning has provided a new approach to addressing uncertainties in numerical weather prediction. Unlike numerical models, deep learning is a data-driven model that discovers new patterns from large amounts of spatiotemporal data. One feasible method to improve bias correction is to continuously learn dynamic weather characteristics from EC model data and enhance correction capabilities through deep learning algorithms. This invention uses topographic and other underlying surface information, ERA5 historical reanalysis data, and EC model 24-hour precipitation forecast products as network inputs, and corresponding ground observations as network targets. It uses data from multiple stations in the study area to correct numerical model precipitation forecast products, ensuring that the shape of the output remains unchanged from the input.

[0006] The new framework consists of three modules: spatial encoding, temporal encoding, and spatial correction. The spatial encoding module encodes contextual information into the study area, taking into account the influence of the surrounding environment on the target area. The temporal encoding module uses historical ERA5 reanalysis data from the past week to predict the spatial variation of meteorological factors in the future. The spatial correction module employs a multi-branch network structure, using one global branch and one local branch to extract spatial correlations of precipitation and focus on details of local precipitation. Ultimately, a relatively accurate regional precipitation forecast is obtained. Summary of the Invention

[0007] Technical issues

[0008] The purpose of this invention is to address the shortcomings and potential improvements of existing technologies. This invention provides a method for correcting precipitation forecast biases in numerical models based on deep learning. It utilizes the learning capabilities of deep learning in big data to uncover potential relationships in precipitation biases and achieve the effect of precipitation correction.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] This invention provides a deep learning-based method for correcting precipitation bias in numerical models, comprising the following steps:

[0012] 1) First, the correlation between meteorological data and ground-observed precipitation was analyzed using ERA5, and meteorological features that are related to precipitation were selected as effective features of the model.

[0013] 2) The effective meteorological features selected in step 1) are cut into the study area to form a regional feature region, and the feature space is normalized to form a training sample set;

[0014] 3) Using the training sample set described in step 2), the feature range is scaled down to the size of the study area in the spatial feature extraction. Then, the meteorological features are input into the spatiotemporal prediction network to extract the changing trend of meteorological features in the future time period, and the feature spatial resolution is kept consistent with the numerical forecast resolution through deconvolution.

[0015] 4) The prediction results described in step 3) are fused with prior knowledge such as numerical precipitation forecast and DEM. The weight loss function is defined in the input spatial correction module. The optimal model is selected through model training to obtain the final precipitation prediction value.

[0016] 5) Verify and evaluate the precipitation results after deviation correction.

[0017] Furthermore, as a preferred technical solution of the present invention: in step 1), the selection of effective meteorological features and the elimination of interfering features are carried out by using the Pearson correlation coefficient to calculate the correlation between each meteorological feature and ground-observed precipitation. A threshold is set to identify all features greater than the threshold as effective features. The calculation formula is as follows.

[0018]

[0019] In the formula Eigenvalues covariance, Let X and Y represent the variances of features X and Y, respectively.

[0020] Furthermore, as a preferred technical solution of the present invention: the process of constructing the dataset in step 2) involves selecting effective features and then selecting data from the target region as training sample features, and performing normalization on the features. Since the precipitation distribution is discrete and has a large span, the precipitation data is log normalized, and the other meteorological features are Z-SCORE normalized. The formula is as follows.

[0021]

[0022]

[0023] Furthermore, as a preferred technical solution of the present invention: Step 3) involves using the prediction of the future trend of meteorological environment changes by first extracting the historical sequence of the constructed ERA5 and scaling the sample to the size of the study area before inputting it into MotionRNN to predict the future trend of cloud changes in the study area. After deconvolution, the spatial resolution of the predicted trend is kept consistent with the numerical model data.

[0024] Furthermore, as a preferred technical solution of the present invention: the spatial correction module described in step 4) is used to correct biases in order to make the model forecast data more accurate. This module uses the classic U-Net structure and is divided into two branches. One branch is responsible for global correction, and the other focuses on the bias correction of smaller regions. Finally, the results are fused and passed through the output layer to obtain the final model forecast result.

[0025] When updating the network parameters to optimize the model, considering that precipitation data has a long-tailed distribution and needs to be balanced, a weighted loss function is designed. This loss function consists of MSE and MAE, with different weights assigned to different precipitation intensities, where the weight of heavy precipitation is greater than that of light precipitation. The loss function is shown below:

[0026]

[0027]

[0028] Finally, the optimal model is run on the validation set, and its performance is verified using relevant metrics.

[0029] Beneficial effects

[0030] This invention proposes a deep learning-based numerical model precipitation product bias correction method. It fully considers the elemental correlation, temporal dependence, and spatial correlation of precipitation processes. By selecting factors that have potential impact on precipitation, a data structure with temporal characteristics is constructed, and a spatiotemporal fusion precipitation bias correction model is built. Ground observations are used as labels for real precipitation, and the model is continuously trained and optimized to obtain the optimal network parameters.

[0031] The model has a good correction effect on precipitation bias correction of ECMWF model and is easy to operate. After obtaining the optimal model parameters, the precipitation bias correction time for the target area only needs one to five seconds, which has high reliability and good application prospects.

[0032] Based on model precipitation products, this invention constructs a historical meteorological database by analyzing weather changes during rainfall and then uses deep learning methods to correct precipitation intensity. This invention preserves the original spatial resolution of historical data and incorporates sub-pixel convolution into the model to improve resolution, avoiding new errors introduced by manual interpolation. When updating parameters, considering the discrete distribution of precipitation and the extreme imbalance of precipitation data, a weighted loss method is used to alleviate the imbalance. After correction using the model provided by this invention, the precipitation distribution and rainband location are significantly improved compared to numerical models. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the basic process for numerical model precipitation bias correction in the implementation of this invention;

[0034] Figure 2 This is a schematic diagram of the model structure of the present invention;

[0035] Figure 3 Schematic diagram of parameter update for the deviation correction model of this invention;

[0036] Figure 4 This is the TS score in an embodiment of the present invention;

[0037] Figure 5 This is a schematic diagram comparing precipitation errors near certain observation stations within the target area in an embodiment of the present invention. Detailed Implementation

[0038] The following will provide a more detailed description of the deep learning-based numerical model precipitation forecast bias correction method of the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, before performing numerical model precipitation bias correction, the data used needs to be selected. First, the correlation between various meteorological features in the ERA5 data and surface-observed precipitation is calculated using the Pearson correlation coefficient. A threshold of 0.2 is set, and all features greater than this threshold are considered valid features. Second, the data is extracted to the study area, which includes the contextual information of the spatial features of the study area, i.e., the information of the surrounding environment of the study area. This is to improve spatial correlation and better extract the spatial distribution information of precipitation by describing the interaction of local information of adjacent locations when extracting features. Since the time resolution of the ERA5 reanalysis data we collected is 1 hour, which is inconsistent with the time resolution of the EC precipitation forecast data, it is necessary to make the time resolution of the reanalysis data consistent with that of EC.

[0040] The specific approach involves summing precipitation characteristics to obtain the 24-hour cumulative daily precipitation. Other variables are downscaled to a 24-hour time resolution based on their own characteristics. To eliminate data loss and errors caused by system failures or data storage issues, missing data and outliers in the collected EC forecast data, rain gauge observation datasets, and ERA5 historical reanalysis data are processed. For missing values, the value is supplemented by the average value of the area before and after that location or a nearby region. Since the EC precipitation forecast data uses UTC time, while the rain gauge observation data uses Beijing time, all datasets require time matching. Considering the different dimensions of various meteorological characteristics in the DEM, EC precipitation forecast data, and historical reanalysis data, which may affect network training, data normalization along the time dimension is necessary. Specifically, except for precipitation characteristics, the remaining characteristics are normalized using the Z-Score algorithm. Due to the large range of discrete precipitation distribution, a log transformation is applied to the precipitation characteristics to alleviate this imbalance. The specific formula is as follows:

[0041]

[0042]

[0043] in It is the overall mean of each element. denoted as the standard deviation of the overall data, z represents the normalized result, and x represents the daily precipitation data. The data used in this invention are from 2017 to 2019. The 24-hour predicted precipitation for 2017-2018 is divided into training and validation sets. To ensure the diversity of validation data, a sample is drawn every 7 days as the validation set. All data from 2019 are used as the test set to evaluate the model performance.

[0044] Precipitation bias correction network such as Figure 2 As shown, the first part is the spatial encoding module, which inputs historical reanalysis data of effective features from the past week to build a historical database. The spatial encoding module consists of a pyramid convolution and two ordinary convolutions. The pyramid convolution has four different scale kernels of 3*3, 5*5, 7*7, and 9*9 to ensure the extraction of multi-scale features. After the multi-scale features are extracted and fused, they enter the ordinary convolution with a stride of 2 to reduce the feature map to the study area. By utilizing contextual information, spatial correlation is improved. The feature map sequence is further input into MotionRNN to predict the future trend of environmental changes of meteorological elements by learning from past meteorological environmental sequences.

[0045] MotionRNN overcomes the shortcomings of existing RNN-based spatiotemporal prediction models, which cannot respond quickly to complex instantaneous changes and cannot model motion trends. It embeds MotionGRU into the model to enhance the modeling of complex motions, and adds Motion Highway connections to balance moving and non-moving components. This is clearly very practical for predicting changes in complex meteorological environments. The working principle of the layer at time t is expressed as follows:

[0046]

[0047]

[0048]

[0049] In the formula This represents the modeling of characteristic motions. It represents the overall movement trend learned. Indicates the first Layer input, This indicates hidden variables in the prediction unit. After obtaining the future trends of various meteorological features through MotionRNN, the feature map is enlarged to maintain the same spatial resolution as the EC precipitation prediction data through transposed convolution. Next, the model prediction data, after being smoothed by a 1*1 convolution operation, is fused with the DEM, slope and aspect static data, and prediction features, awaiting entry into the spatial precipitation correction module.

[0050] The correction module uses the U-Net network structure, which has the advantage of capturing objects of different sizes in the input by using features of different scales. This is especially important for precipitation correction tasks, because some precipitation details are often ignored in model data, and our network must also be able to correct the accurate precipitation values ​​for each grid.

[0051] Traditional deep learning networks mostly adopt a single-branch structure, which can only perceive a single feature and lacks comprehensive attention to global and local features, thus failing to extract data features effectively.

[0052] To improve the overall understanding of features, this invention uses two branches in the precipitation correction network: a global branch that processes the entire region and a local branch that processes the features of a specific region. These two branches, one global and one local, have different functions and coordinate with each other, taking into account both the overall precipitation of the region and the details of local precipitation. Finally, the features are spliced ​​and fused to obtain the spatial features of precipitation within the target region; this information is then fed into the output layer to obtain the final precipitation forecast.

[0053] The 24-hour precipitation is highly unbalanced across different magnitudes (from 10mm of light rain to 100mm of torrential rain, and even larger), exhibiting a long-tailed distribution. Using a conventional loss function would degrade the model's correction performance because most samples are negative, providing no effective signal. Therefore, this invention proposes using a weighted loss function to help address this issue. The specific distribution proportions of precipitation magnitudes are shown in the table below:

[0054]

[0055] Specifically, we assign a weight w(x) to each precipitation intensity x.

[0056]

[0057] The loss function consists of the weighted mean squared error (MSE) and the weighted average absolute error (MAE), and its formula is as follows:

[0058]

[0059] Where N represents the number of rain gauges in the area. and These represent the actual and corrected precipitation amounts at the corresponding latitude and longitude, respectively. The weight corresponding to the rainfall intensity at the i-th station is given, and the specific reverse parameter update process is as follows: Figure 3 As shown.

[0060] The trained optimal model is applied to the validation set, and the model training results are statistically analyzed to calculate the TS score to evaluate the model performance.

[0061] Specifically, such as Figure 4 5. Figure 4 The TS score, or risk score, is shown in the following formula:

[0062]

[0063] in The boundaries between positive and negative samples are defined; specifically, the label is greater than... Samples with a marker size smaller than 0.5 are considered positive samples. Samples that do not meet the precipitation threshold are considered negative samples. Under these conditions, NA represents the number of correctly predicted grid points after correction, NB represents the number of missed grid points after correction, NC represents the number of false alarms after correction, and ND represents the number of precipitation points where both the predicted and observed precipitation are less than the precipitation threshold. The number of grid points.

Claims

1. A method for correcting precipitation bias in numerical models based on deep learning, characterized in that, Includes the following steps: 1) First, the correlation between meteorological data and ground-observed precipitation was analyzed using ERA5, and influencing factors that have a potential relationship with precipitation were selected as effective features of the model; 2) The influencing factors that have a potential relationship with precipitation selected in step 1) are segmented into the study area to form a characteristic region, and the characteristic region is normalized to form a training sample set; 3) Using the training sample set described in step 2), the feature range is scaled down to the size of the study area in the spatial feature extraction. Then, the meteorological features are input into the spatiotemporal prediction network to extract the changing trend of meteorological features in the future time period, and the feature spatial resolution is kept consistent with the numerical forecast resolution through deconvolution. 4) Overlay the prediction results described in step 3) with numerical precipitation products and DEM prior knowledge, input them into the spatial correction module, update the model parameters using a custom weight loss function, and select the optimal model to verify the effect of model bias correction on the test set. 5) Verify and evaluate the precipitation results after deviation correction; Step 2) involves the following steps for creating the dataset and normalizing the features: (a) The selected effective features are segmented into the study area and diffused outwards, taking into account the impact of the surrounding environment on precipitation in the target area; (b) Normalization is performed on the segmented features, with Z-SCORE normalization used for all features except precipitation. The calculation formula is as follows: Data features are then normalized to the same dimension, where It is the overall mean of each element. It is the standard deviation of the overall data; Because precipitation is discretely distributed and varies significantly in magnitude, log normalization is used to mitigate the large differences in precipitation. The calculation formula is as follows: ; The spatial correction module in step 4) is divided into two branches: The global and local branches take into account both global and local precipitation. The entire model is trained using a weighted loss function to optimize parameters and alleviate the problem of extremely uneven precipitation distribution. The expression for the weighted loss function is as follows: Where N represents the number of rain gauges in the area, and These represent the actual and corrected precipitation amounts at the corresponding latitude and longitude, respectively. Here are the weights corresponding to the rainfall intensity at the i-th station, with heavy rain having a greater weight than light rain, as shown below: ; The weight corresponding to each precipitation intensity x; Finally, the optimal model is used to evaluate its performance using a validation set.

2. The method for correcting precipitation bias in numerical models based on deep learning according to claim 1, characterized in that, In step 1), the Pearson correlation coefficient is used to calculate the correlation between various meteorological features and precipitation observed at ground stations. A threshold is set, and all meteorological features exceeding this threshold are considered valid features. The Pearson correlation coefficient is calculated as follows: In the formula Eigenvalues covariance, Let X and Y represent the variances of features X and Y, respectively.

3. The method for correcting precipitation bias in numerical models based on deep learning according to claim 1, characterized in that, Step 3) The spatiotemporal prediction network first considers the impact of surrounding meteorological environment information on the target area, then uses MotionRNN to extract changes in meteorological characteristics over a period of time from the constructed historical meteorological sequence information to infer the future trend of the target area, and uses deconvolution to obtain a spatial resolution consistent with the precipitation in the numerical model.

Images