Method for correcting precipitation forecast bias based on ensemble prediction

Abstract

The present application belongs to the technical field of deep learning and meteorological prediction, and comprises the following specific steps: taking the reservoir basin in the central part of China as the research object, selecting the control prediction results of 5 prediction modes, dividing the prediction data of the central reservoir according to the latitude and longitude information of the ground observation station, taking the data information near the station as the influencing factor, and performing regularization processing on the data in the small space range; the regularized prediction data of each mode is respectively put into each grouping module to obtain the characteristics of each mode prediction, and finally the accurate precipitation prediction near each station is obtained through the output module after effectively fusing the features such as terrain and station latitude and longitude. The present application can use multiple sub-networks to extract different mode characteristics respectively, and jointly act, give different weights, and can effectively fuse the prediction data of each mode to generate more accurate precipitation value near the station, so as to achieve better correction effect.

Description

Technical Field

[0001] This invention relates to the field of precipitation forecast bias correction methods based on ensemble forecasts, and particularly to precipitation forecast bias correction methods based on ensemble forecasts. Background Technology

[0002] With the development of numerical weather prediction, it has become increasingly apparent that improving individual numerical weather prediction models is difficult, especially for longer-term forecasts. After a long period of development, numerical weather prediction technology has advanced rapidly, evolving from traditional single deterministic forecasts to ensemble numerical forecasts. Numerous studies have shown that ensemble forecasts offer greater economic value than single deterministic forecasts. Ensemble forecasts can effectively utilize the forecast results of various central models to reduce systematic biases. Ensemble forecasts involve statistically post-processing the outputs of individual models, integrating the forecast results from multiple models to reduce forecast errors. This is an effective way to obtain more accurate forecasts.

[0003] Numerous meteorological researchers have conducted extensive research on ensemble forecasting techniques. To reduce the uncertainty of numerical weather predictions, Krishnamurti et al. were among the first to propose the concept of multi-ensemble forecasting and corrected precipitation forecasts for the global tropics and the United States. Ross et al. applied super-ensemble forecasting techniques, using daily data from six global numerical weather prediction models to perform ensemble forecasts of wind fields. Their research also found that super-ensemble forecasts outperformed the best single model and simple ensemble averaging of multiple models. Ji et al. proposed a new method for multi-model ensemble forecasting of precipitation based on MODE assessment. This new method not only improves precipitation forecasting but also provides a more complete picture of precipitation areas, better preserving the spatial correlation of precipitation forecasts. Therefore, previous research has shown that the forecasting skill of multi-model super-ensemble forecasts far surpasses that of single models and simple ensemble averaging, significantly reducing errors in weather and seasonal climate forecasts.

[0004] Traditional multi-model ensemble methods commonly used in meteorology include simple ensemble averaging, bias-free ensemble averaging, and super-ensembles based on linear regression. While these methods can improve forecasting skills, they are all linear statistical post-processing methods. The atmospheric system is a chaotic system with numerous nonlinear relationships between its elements, thus ensemble forecasting methods still have significant room for improvement. The explosive use of deep learning in various fields in recent years has demonstrated that this method can extract hidden nonlinear relationships from large amounts of historical information. Some scholars have achieved satisfactory results using machine learning-based multi-model ensemble forecasting methods such as multiple linear regression and backpropagation neural networks. Yuan et al.'s use of ANNs for statistical post-processing of precipitation ensemble forecasts also demonstrates the feasibility of artificial intelligence in ensemble forecasting. However, these machine learning / deep learning applications are relatively shallow networks, which have limitations in extracting information from complex meteorological features. Summary of the Invention

[0005] The purpose of this invention is to improve model forecasting skills by utilizing the concept of ensemble forecasting, providing data support for flood forecasting and disaster prevention and mitigation in the central reservoir area. The ensemble forecasting method provided by this invention is based on multiple subnetworks that forecast separately from each model, followed by weighted fusion to achieve optimal precipitation forecasting results.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a precipitation forecast bias correction method based on ensemble forecasts, comprising the following steps:

[0007] 1) First, control forecasts and ground station observations of precipitation from five forecast models in the TIGGE system—European Centre for Medium-Range Weather Forecasts (ECMWF), National Center for Environmental Prediction (NCEP), China Meteorological Administration (CMA), Japan Meteorological Agency (JMA), and UK Met Office (UKMO)—were collected. The correlation between the forecast data from each model and the ground observations of precipitation was analyzed, and any unnecessary noise input that might negatively affect performance was removed.

[0008] 2) Secondly, the data of the central reservoir area is divided based on the location (latitude and longitude) of the ground observation station to generate several feature regions containing information about the surrounding stations; these feature regions are then regularized.

[0009] 3) Then, the small grid weather data of each segmented model forecast are input into the corresponding sub-network encoder. Static prior knowledge and convolution operations are performed in the encoding network to capture the influence of different meteorological characteristics on precipitation. Then, the data are trained in a module composed of stacked LSTMs to obtain the initial state of the decoding unit.

[0010] 4) The encoded static prior knowledge and the output features of each encoder are shared as parameters, serving as the initial state of each decoding network. Each decoding network is composed of the same network structure stacked from LSTMs, and the information between the sub-networks corresponding to each mode is not shared. Finally, the decoding network can output the prediction features of each mode.

[0011] 5) The decoding networks work together. One branch assigns different weights to the output features of each network to obtain the final precipitation forecast result. Another branch uses a unit composed of convolutional blocks to make predictions on the precipitation level near the station (no rain, light rain, moderate rain, heavy rain, and rainstorm) to help obtain more accurate precipitation forecast data.

[0012] Preferably, in step 1), the selection of effective meteorological feature information and the removal of any unnecessary noise input that may negatively affect performance are carried out. The stepwise regression method is used to screen variables. Specifically, the grid surface data is interpolated to the vicinity of the station using the inverse distance weighting method (IDW) to obtain the interpolated feature values ​​near the station. Each feature variable is then introduced into the regression model one by one. An F-test is performed after each explanatory variable is introduced, and a t-test is performed on each of the selected explanatory variables. When the previously introduced dependent variable becomes insignificant due to the introduction of subsequent dependent variables, the dependent variable is deleted. This ensures that the regression model contains only significant variables before each new variable is introduced.

[0013] Preferably, the data segmentation and feature normalization steps in step 2) are as follows:

[0014] (a) First, the model forecast time and the ground observation time are uniformly converted to Beijing time. Then, based on the location (latitude and longitude) coordinates of each ground observation station, the station coordinates are aligned in the grid data. Considering the spatial information of precipitation, for each ground observation station, an 11*11 sub-grid centered on the station's latitude and longitude is extracted from the model grid data to represent the characteristic spatial distribution information of the current station. The currently observed precipitation values ​​are then saved to achieve the time and spatial correspondence between the model grid data and the ground observation data as training data.

[0015] (b) Normalization is performed on the segmented features. To avoid introducing large outliers, the segmented data is first standardized using a Min-Max method. The maximum and minimum values ​​of each element are officially defined. Above, this method eliminates differences on the feature scale, where u i σ is the population mean of the i-th feature. i It is the overall standard deviation of the i-th feature.

[0016] Preferably, the steps of the encoding module for each pattern data in step 3) are as follows:

[0017] (a) The encoding information needs to use geographical environment (latitude, longitude and altitude data) and timestamp information as static variables. After static encoding, it is combined with pattern feature data and input into the encoding unit.

[0018] (b) The coding layer contains a 3*3 convolutional kernel to extract features from each mode data and static information, and obtain the dynamic impact of surrounding characteristics on precipitation near the station.

[0019] (c) The features extracted by convolution are flattened and then fed into a sequence encoding module composed of stacked LSTMs. This module can learn the changes in precipitation over time and combine them with static encoding variables as the initial state of the sequence decoding module.

[0020] Preferably, the temporal prediction decoding module described in step 4) uses encoded static prior knowledge and the output features of each encoder as parameters to serve as the initial state of each decoding network. Each decoding network is composed of identical LSTM stacks, and information is not shared between the sub-networks corresponding to each mode. The prediction information at time T is propagated to time T+1, meaning that the prediction data at time T+1 is related to the data at time T. Finally, each sub-network of the temporal decoding network can output the prediction features for each mode.

[0021] Preferably, step 5) involves weight fusion and precipitation level prediction. For the output, the network model is optimized by adding MAE and TS-like scores at different proportions as a loss function, thus obtaining the optimal network model. The expression for the loss function is:

[0022] Loss=λL1+ηLTS

[0023]

[0024] Where Y represents the actual observed precipitation. To enable model fusion for precipitation prediction, y i This refers to the actual precipitation level corresponding to the rainfall. The model obtained through convolutional block branching is used to predict precipitation levels;

[0025] The weighted fusion and precipitation level prediction consists of two branches: one branch outputs the determined precipitation prediction after weighted fusion, and the other branch outputs the precipitation level prediction for each station after a convolutional block.

[0026] The precipitation prediction branch, due to the weight fusion module in the output layer, can be represented by the final precipitation regression as follows:

[0027]

[0028] in Each corresponds to the output value of the sub-network for the corresponding pattern operation. ecmwf w NCEP w CMA w IMA w UKMO These are all dynamic parameters that need to be continuously learned and updated from the network. The final precipitation prediction result is the Hadamard product of the output values ​​of each sub-network and the dynamically learned parameters. For the output results, MAE and TS-like scores are added at different proportions as a loss function to optimize the network model, thereby obtaining the optimal network model. The expression of the loss function is:

[0029] Loss=λL1+ηL TS

[0030]

[0031] Where Y represents the actual observed precipitation. To enable model fusion for precipitation prediction, y i This refers to the actual precipitation level corresponding to the rainfall. The model obtained through convolutional block branching is used to predict precipitation levels.

[0032] The technical effects and advantages of this invention are as follows:

[0033] (1) This invention first uses the stepwise linear regression method to screen out meteorological factors that affect precipitation in each model data, and removes outliers and meteorological factors that may have a negative impact on precipitation, thereby improving the quality of model precipitation data and thus improving precipitation forecasting capabilities.

[0034] (2) The present invention uses multiple identical sub-networks to extract and forecast the forecast data of each model and then perform weighted feature fusion, which can make full use of the advantages of each model. Secondly, the network of the present invention generates another branch for the station precipitation level forecast. The two branches influence each other, which can obtain more accurate precipitation and alleviate the system bias of precipitation forecast of each model. Attached Figure Description

[0035] Figure 1 This is a diagram illustrating the research area of ​​this invention;

[0036] Figure 2 This is a schematic diagram of the model structure of the present invention;

[0037] Figure 3 This is a schematic diagram of the structure of the model encoding and decoding part of the present invention;

[0038] Figure 4 This is a flowchart of the ensemble forecast bias correction method proposed in this invention;

[0039] Figure 5 This is the TS score in an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] This invention provides, for example Figures 1-5 The precipitation forecast bias correction method based on ensemble forecast is shown below: The study area of ​​this invention is as follows: Figure 1 As shown (Central Reservoir Basin, China), the dataset primarily collects forecast data from five model forecasting centers within the TIGGE system: the European Centre for Medium-Range Weather Forecasts (ECMWF), the National Center for Environmental Prediction (NCEP), the China Meteorological Administration (CMA), the Japan Meteorological Agency (JMA), and the UK Met Office (UKMO). Forecasts are issued twice daily, at 00:00 and 12:00 (UTC). The spatial resolution of this model data is 10 km. 24-hour daily cumulative precipitation is used in this study. This invention selects daily precipitation data from national stations in the Central Reservoir Basin as the actual precipitation and as the training data's true label Y, with exceptions as follows: Figure 2 As shown, the present invention also incorporates terrain elevation, observation station latitude and longitude, and time as static coding variables.

[0042] Before performing ensemble forecast precipitation bias correction, it is necessary to select the data to be used. This invention uses a stepwise regression model to screen factors that affect precipitation. The specific steps are as follows:

[0043] (a) Interpolate the grid surface data to the vicinity of the station using the inverse distance weighting (IDW) method to obtain the interpolated feature values ​​near the station. For each model forecast data, there are p regression independent variables X1, X2, ..., X... p Establish univariate regression models for the dependent variable, precipitation value Y:

[0044] Y = β0 + β i X i +∈, i=1,2,...,p, calculate variable X i The value of the F-test statistic for the corresponding regression coefficient is denoted as . final Set a threshold α as the significance level, and the corresponding critical value F.1 ,like Then X i1 The characteristics represented are introduced into the autoregressive model.

[0045] (b) The independent variables of the model forecast data are divided into multiple groups of independent variables (x1, x2), (x1, x3), ..., (x1, x...). p Establish a binary regression equation with the dependent variable precipitation value, and calculate x2, x3, ..., x in the equation. p The regression coefficient test statistic F is denoted as F. The maximum value is denoted as

[0046] If the filtering stops after certain conditions are met, the variable is selected into the variable set. This method is used to select features.

[0047] (c) Consider all dependent variable pairs of variable subsets {X} i1 X i2 ,..,X p The regression repeats step (b).

[0048] Repeat the above steps, selecting one value from the unadded features for each regression model, until no variables are introduced after testing. After stepwise regression screening to identify relevant meteorological factors affecting precipitation in each model, align the ground observation station data with the model grid data, and divide the data features into smaller ranges centered on the station's latitude and longitude coordinates, ultimately forming an 11*11 grid size. To eliminate data anomalies caused by system failures or data storage issues, Min-Max standardization was performed on the collected forecast data from the five model centers and the rain gauge observation dataset, shrinking all data to the range provided by the official data provider. The calculation formula is as follows: To avoid the impact of outlier data on the model, after data scaling, due to differences in the feature dimensions of different meteorological factors, a Z-score normalization is required. The formula is as follows: Where u is the population mean of each element, σ is the standard deviation of the population data, z is the normalized result, and x is the daily precipitation data. For example... Figure 2 As shown, since the precipitation forecast data collected from the five model centers uses UTC time, while the ground observation data uses Beijing time, it is necessary to unify the time between the grid data and the station data.

[0049] Ensemble precipitation bias correction network model, such as Figure 3 As shown. The model consists of multiple parts. Values ​​such as terrain elevation, latitude and longitude, and time information are incorporated into an embedding and fed into the network. Each encoded factor is then mixed with the five model center prediction features before being fed into the network's encoding layer (e.g., ...).Figure 4 a) After extracting mixed features using 3x3 convolutional layers, the data enters an encoding structure composed of stacked LSTMs. The encoding structure has no neuron outputs; its context vector is c = q(h1, h2, ..., h...). N Using the hidden states h1, h2, ..., h of all neurons N The context vector c is calculated and then passed to the next layer's decoding structure (e.g., ...). Figure 4 (b) The decoding structure transmits the output of the encoding layer and the static encoded features as a mixed feature. Furthermore, the input of the decoding unit includes the neuron output y′ from the previous time step. That is, the input structure of each neuron includes: the hidden vector h′ of the previous neuron, the output y′ of the previous neuron, and the current input is the context vector of the encoding layer and the encoded features of the static variables. The calculation formula is shown below:

[0050] h′ t =σ(Uc+Wh′) t-1 +Vy′ t-1 +b)

[0051] y′ t =σ(Vh′) t +c)

[0052] After passing through the sequence encoding / decoding subnetwork, the model obtains encoded prediction data for five pattern centers. Next, it employs hybrid feature fusion. The features output from the five subnetworks are cross-fused and passed to two branches. One branch undergoes a 1x1 convolution and activation function to output a definite precipitation result. The other branch undergoes a 3x3 and 1x1 convolution followed by a sigmoid operation to output a predicted precipitation level for each station. The two branches influence each other, with the precipitation level branch assisting in determining the quantitative precipitation branch, resulting in a more accurate precipitation prediction.

[0053] In this invention, L1 loss is used to optimize deterministic precipitation. However, a major problem with L1 loss is that it is often considered to smooth the entire data. For precipitation forecasting, it's clear that L1 cannot perfectly update parameters because precipitation can vary significantly across different locations. Therefore, another branch of this invention focuses on precipitation level prediction and considers a loss that can distinguish local differences while maintaining overall similarity. In meteorology, binary classification is frequently used to evaluate model performance; for example, the TS score is widely used to evaluate the accuracy of precipitation prediction across different rainfall amounts. In this invention, precipitation levels are divided into five categories: no rain, light rain, moderate rain, heavy rain, and torrential rain. Directly using the TS score to optimize network parameters is very challenging; therefore, this invention uses the TS score concept to maximize the prediction accuracy of station precipitation levels. In short, all station precipitation levels are treated as a single category and used as the cross-entropy loss function to train the network. This loss is defined as follows:

[0054]

[0055] Where y i The precipitation level corresponds to the ground-observed precipitation. This represents the precipitation level predicted for this station by the precipitation level prediction branch. The overall loss of the proposed network is expressed as follows:

[0056]

[0057] Y represents the specific precipitation observed on the ground. To predict precipitation using the model, we used the data preprocessing and model building methods outlined in the implementation plan. We divided the data using surface observation data of central reservoirs from 2017 to 2019, with 80% used for model training and 20% for model validation. We then used 2020 data to validate the final results of our proposed model. The Threat Score (TS) was used to validate model performance. The TS score is defined as follows: σ represents the precipitation thresholds: 10mm, 25mm, 50mm, 100mm, and 150mm correspond to light rain, moderate rain, heavy rain, torrential rain, and extremely heavy rain, respectively. The evaluation results are as follows: Figure 5 As shown in the figure, it is clear from the figure that the proposed method of using the network to integrate and correct the ensemble members has achieved better results, and its precipitation forecast performance at all levels is better than the best case of the single model.

[0058] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for correcting precipitation forecast bias based on ensemble forecasts, characterized in that, Includes the following steps: 1) First, collect control forecasts and ground station observations of precipitation from at least 5 forecast models in the TIGGE system, and remove the negative impact on performance by inputting noise between the forecast data of each model and the ground observations of precipitation. 2) The data of the central reservoir area is divided into several feature regions containing information about the surrounding areas of the stations, based on the location of the ground observation stations. The feature regions are then regularized. 3) The small grid weather data of each model forecast after segmentation are input into the corresponding sub-network encoder. Static prior knowledge and convolution operations are performed in the encoding network to capture the influence of different meteorological characteristics on precipitation. Then, the data is trained in a module composed of stacked LSTMs to obtain the initial state of the decoding unit. 4) The encoded static prior knowledge is shared as parameters with the output features of each encoder; 5) The decoding networks work together. One branch assigns different weights to the output features of each network to obtain the final precipitation forecast result, while the other branch uses a unit composed of convolutional blocks to predict the precipitation level near the station. Step 5) Use MAE and class TS scores added together at different proportions as the loss function to optimize the network model, thereby obtaining the optimal network model. The expression for the loss function is: ； 。 2. The precipitation forecast bias correction method based on ensemble forecasting according to claim 1, characterized in that: Step 1) Use stepwise regression to screen variables. The specific steps are as follows: interpolate the grid surface data to the vicinity of the station using the inverse distance weighting method to obtain the feature values ​​after interpolation near the station. Then, introduce each feature variable into the regression model one by one. After each explanatory variable is introduced, perform an F test and perform a t test on each of the selected explanatory variables. When the original dependent variable becomes no longer significant due to the introduction of subsequent dependent variables, delete the dependent variable to ensure that the regression model only contains significant variables before each new variable is introduced.

3. The precipitation forecast bias correction method based on ensemble forecasting according to claim 1, characterized in that: The specific steps for step 2) are as follows: (a) First, the model forecast time and the ground observation time are uniformly converted to Beijing time. Then, based on the location coordinates of each ground observation station, the station coordinates are aligned in the grid data. Considering the spatial information of precipitation, for each ground observation station, a sub-grid containing 11*11 grids is extracted from the model grid data with the station's latitude and longitude as the center to represent the characteristic spatial distribution information of the current station. The current observed precipitation value is saved to achieve the training data for the temporal and spatial correspondence between the model grid data and the ground observation data. (b) Normalization is performed on the segmented features. To avoid introducing large outliers, the segmented data is first standardized using a Min-Max method. The maximum and minimum values ​​for each element are provided on the official website, and the calculation formula is as follows: Then, the standardized features are normalized using the Z-SCORE method.

4. The precipitation forecast bias correction method based on ensemble forecasting according to claim 1, characterized in that, The specific steps for step 3) are as follows: (a) Geographic environment and timestamp information need to be used as static variables in the encoding information. After static encoding, they are combined with pattern feature data and input into the encoding unit. (b) The coding layer contains a 3*3 convolutional kernel to extract features from each mode data and static information, and obtain the dynamic impact of surrounding characteristic information on precipitation near the station. (c) The features extracted by convolution are flattened and then fed into a sequence encoding module composed of stacked LSTMs.

5. The precipitation forecast bias correction method based on ensemble forecasting according to claim 1, characterized in that: Step 4) is the initial state of a decoding network, where each decoding network is a stack of LSTMs with the same network structure, and the information between the sub-networks corresponding to each mode is not shared. The prediction information at time T will be passed on to time T+1. After passing through the temporal decoding network, each sub-network can output the prediction features of each mode. In step 2), the calculation formula is: Data features are normalized to the same dimension to eliminate differences in feature scale. It is the first The overall mean of each feature. It is the first The overall standard deviation of each feature characteristic; In step 5), where To accurately observe precipitation, To enable model fusion for precipitation prediction, This refers to the actual precipitation level corresponding to the rainfall. The model obtained through convolutional block branching is used to predict precipitation levels.

6. The precipitation forecast bias correction method based on ensemble forecasting according to claim 1, characterized in that: In step 4), as the initial state of a decoding network, each decoding network is a stack of LSTMs with the same network structure, and the information between the sub-networks corresponding to each mode is not shared. Finally, the decoding network can output the prediction features of each mode.

7. The precipitation forecast bias correction method based on ensemble forecasting according to claim 4, characterized in that: In step (c) of step 3), the module learns that the time-varying precipitation and the combination of static coding variables together serve as the initial state of the sequence decoding module.

Images