A weather forecasting method and device for fusing multi-scale numerical weather prediction models

Abstract

The present invention provides a weather forecasting method and device for multi-scale numerical weather prediction model fusion. The weather forecasting method includes: obtaining target forecasting factors for multi-scale numerical weather prediction model fusion based on general circulation forecast data and convective weather forecast data; and inputting the target forecasting factors into a pre-trained deep learning forecasting model to obtain a weather forecasting result. In the present invention, by fusing the forecasting factors of numerical weather prediction models at different scales and inputting them into a pre-trained deep learning forecasting model, the weather forecast is obtained, which solves the defect in the prior art that the forecasting effects of weather systems at different scales cannot be taken into account simultaneously, realizes the fusion of multiple-scale numerical weather prediction models, and effectively improves the forecasting effects of weather systems at different scales.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting, and in particular to a weather forecasting method and device integrating multi-scale numerical weather forecasting models. Background Art

[0002] With the development of computer technology and the advancement of data assimilation technology, numerical weather forecasting has become the core and foundation of the current meteorological department's weather forecasting, and has an irreplaceable position and important role that other forecasting methods cannot replace. At the same time, deep learning has also been widely used in the field of weather forecasting due to its powerful feature extraction capabilities.

[0003] At present, numerical weather forecast models can be divided into global numerical weather forecast models and convectively resolved numerical weather forecast models, depending on whether the convective parameterization scheme is used to directly simulate the atmospheric convection process. Among them, the global numerical weather forecast models mainly include the GRAPES global model of the China Meteorological Center, the global forecast model of the European Center for Medium-Range Weather Forecasts, and the GFS model of the U.S. Environmental Forecast Center. Generally speaking, the temporal and spatial resolutions in the global numerical weather forecast models are relatively coarse, with a time resolution of generally 1-3 hours and a spatial resolution of 9-50km. They can have good forecasting effects on short-term large-scale weather changes in the world in the next 0-10 days, but the simulation effects on small and medium-scale convective systems, which are usually only on the order of kilometers, are very limited.

[0004] In order to overcome the inadequacy of the global numerical weather prediction model's ability to predict convective systems, convectively resolvable numerical weather prediction models have developed rapidly around the world. All countries have vigorously developed convectively resolvable weather prediction systems with higher temporal and spatial resolutions that can better characterize convective weather systems, mainly including the GRAPES-meso system developed by China with a temporal and spatial resolution of 3km / h, the European Joint Small-Scale Model COSMO developed by the European Center for Weather Forecasting, and the high-resolution numerical model HRRR forecast system developed by the United States. When the temporal and spatial resolution of the numerical prediction model system is high enough, the convective scale circulation can be explicitly resolved. It can be seen that the convectively resolvable numerical weather forecast has a good forecasting effect on small and medium-scale convective systems, but its overall forecasting effect on large-scale weather systems cannot be taken into account.

[0005] Therefore, how to effectively improve the forecasting effect of weather systems of different scales is an important issue that needs to be urgently solved in the field of weather forecasting technology. Summary of the invention

[0006] The present invention provides a weather forecasting method and device for fusing multi-scale numerical weather forecasting models, aiming to solve the defect in the prior art that the forecasting effects of weather systems of different scales cannot be taken into account simultaneously, and to achieve the fusion of multiple-scale numerical weather forecasting models, thereby effectively improving the forecasting effects of weather systems of different scales.

[0007] On the one hand, the present invention provides a weather forecasting method for fusing multi-scale numerical weather forecasting models, including: obtaining target forecasting factors for fusing multi-scale numerical weather forecasting models according to general circulation forecasting data and convective weather forecasting data; inputting the target forecasting factors into a pre-trained deep learning forecasting model to obtain a weather forecasting result.

[0008] Further, the step of obtaining target forecasting factors for fusing multi-scale numerical models according to general circulation forecasting data and convective weather forecasting data includes: selecting the target forecasting factors from the general circulation forecasting data and the convective weather forecasting data by means of permutation importance analysis.

[0009] Further, the target forecasting factors include forecasting factors of general circulation characteristics and forecasting factors of convective weather forecasting characteristics; among them, the forecasting factors of general circulation characteristics include temperature, geopotential height, and specific humidity; the forecasting factors of convective weather forecasting characteristics include horizontal wind field, vertical velocity, temperature, relative humidity, cloud water mixing ratio, rain water mixing ratio, ice water mixing ratio, snow water mixing ratio, and graupel.

[0010] Further, the weather forecasting method for fusing multi-scale numerical weather forecasting models further includes: determining a training set and a test set according to the target forecasting factors; respectively inputting the training set and the test set into the deep learning forecasting model for training and testing to obtain the pre-trained deep learning forecasting model.

[0011] Further, the step of respectively inputting the training set and the test set into the deep learning forecasting model for training and testing includes: inputting the target forecasting factors in the training set into the corresponding encoders of the deep learning forecasting model; inputting the output results of all the encoders into the same decoder of the deep learning forecasting model to obtain the fusion features of the target forecasting factors; training the deep learning forecasting model according to the fusion features;

[0012] inputting the target forecasting factors in the test set into the corresponding encoders of the deep learning forecasting model; inputting the output results of all the encoders into the same decoder of the deep learning forecasting model to obtain the test fusion features of the target forecasting factors; testing the deep learning forecasting model according to the test fusion features.

[0013] Furthermore, the loss function used for training the deep learning prediction model includes the cross-entropy loss function and the loss function L CSI ; the loss function L CSI has the following formula:

[0014]

[0015] where CSI is the critical success index, y i is the class label of grid point i, p i is the probability of a certain weather occurring at the predicted grid point i, N is the number of grid points, and γ is a constant.

[0016] In a second aspect, the present invention also provides a weather forecasting device for multi-scale numerical weather forecasting model fusion, including: a forecast factor acquisition module for acquiring target forecast factors for multi-scale numerical model fusion according to atmospheric circulation forecast data and convective weather forecast data; a forecast result acquisition module for inputting the target forecast factors into a pre-trained deep learning prediction model to obtain a weather forecasting result.

[0017] In a third aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the weather forecasting method for multi-scale numerical weather forecasting model fusion as described in any one of the above are implemented.

[0018] In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the weather forecasting method for multi-scale numerical weather forecasting model fusion as described in any one of the above are implemented.

[0019] In a fifth aspect, the present invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the weather forecasting method for multi-scale numerical weather forecasting model fusion as described in any one of the above are implemented.

[0020] The weather forecasting method for multi-scale numerical weather forecasting model fusion provided by the present invention obtains target forecast factors for multi-scale numerical weather forecasting model fusion from atmospheric circulation forecast data and convective weather forecast data, and inputs the obtained target forecast factors into a pre-trained deep learning prediction model to effectively fuse the forecast factors under different scale numerical weather forecasting models, and further obtains the predicted weather forecasting result. This method solves the defect in the prior art that the forecasting effects of weather systems of different scales cannot be taken into account simultaneously, realizes the fusion of multiple scale numerical weather forecasting models, and effectively improves the forecasting effects of weather systems of different scales. Description of the Drawings

[0021] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0022] Figure 1 It is a schematic flowchart of the weather forecasting method for multi-scale numerical weather forecasting model fusion provided by the present invention;

[0023] Figure 2 It is a structural diagram of the deep learning forecasting model provided by the present invention;

[0024] Figure 3 It is an ROC curve diagram of the forecasting factors of the global numerical weather forecasting model provided by the present invention;

[0025] Figure 4 It is an ROC curve diagram of the forecasting factors of the convective-resolving numerical weather forecasting model provided by the present invention;

[0026] Figure 5 It is an ROC curve diagram of the forecasting factors of the deep learning forecasting model using different numerical weather forecasting models provided by the present invention;

[0027] Figure 6 It is a preliminary fusion result diagram of the global numerical weather forecasting model and the convective-resolving numerical weather forecasting model provided by the present invention;

[0028] Figure 7 It is a comparison diagram of the forecasting situations of different models provided by the present invention;

[0029] Figure 8 It is a schematic structural diagram of the weather forecasting device for multi-scale numerical weather forecasting model fusion provided by the present invention;

[0030] Figure 9 It is a schematic structural diagram of the electronic device provided by the present invention. Specific embodiments

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0032] At present, according to whether to directly simulate the atmospheric convection process using a convection parameterization scheme, numerical weather prediction models can be divided into global numerical weather prediction models and convective-resolving numerical weather prediction models. Among them, global numerical weather prediction models have better overall prediction effects for large-scale weather systems, while convective-resolving numerical weather prediction models have better characterization capabilities for medium- and small-scale convective systems. In the existing technology, there is no technical solution that can simultaneously take into account the prediction effects of weather forecasting systems of different scales.

[0033] Generally, large-scale weather processes, such as changes in the unstable layer structure, low-level convergence and uplift, and water vapor transport, determine the environmental conditions of moist convection. The convection process, through the release of accumulated latent heat, in turn acts on the evolution of the large-scale circulation. Thus, it can be seen that weather systems of various scales influence each other and are inseparable. Considering this, based on the global prediction model of the European Centre for Medium-Range Weather Forecasts and the GRAPES 3km model developed by the Numerical Prediction Centre of the China Meteorological Administration, the present invention specifically constructs a deep learning prediction model, trains and tests it, and realizes the organic integration of the global numerical weather prediction model and the convective-resolving numerical weather prediction model, that is, the integration of numerical weather prediction models of different scales, so as to effectively improve the prediction effects on weather systems of different scales.

[0034] Figure 1 The flowchart of the weather forecasting method for the multi-scale numerical weather prediction model fusion provided by the present invention is shown, as Figure 1 shown, the forecasting method includes two steps, where:

[0035] S11, according to the atmospheric circulation forecast data and the convective weather forecast data, obtain the target forecast factors for the multi-scale numerical weather prediction model fusion.

[0036] In this step, the atmospheric circulation forecast data can be ERA5 data under the global prediction model of the European Centre for Medium-Range Weather Forecasts. ERA5 provides upper-air and surface analysis fields on a global scale, with a time resolution of 1h, an upper-air analysis field spatial resolution of 0.25°×0.25°, and a surface analysis field spatial resolution of 0.125°×0.125°.

[0037] The convective weather forecast data can be taken from the GRAPES 3km model data developed by the Numerical Prediction Center of the China Meteorological Administration. GRAPES 3km began real-time operation in 2016. Since June 2020, it has started issuing forecasts 8 times a day, with the issuance times being 00, 03, 06, 09, 12, 15, 18, and 21 UTC. The forecast period is 0 - 36 hours, and the time and space resolutions are 1 hour and 0.03° respectively. The covered spatial range is (10° - 60.1°N, 70° - 145°E). GRAPES 3km provides relatively rich physical quantities.

[0038] The multi-scale numerical weather forecast model fusion provided by the present invention refers to the fusion of weather forecast models applicable to different scales. Specifically, it is to fuse the forecast factors under weather forecast models of different scales, so as to effectively improve the forecast effect of weather forecast systems at different scales. For example, in this embodiment, the atmospheric circulation forecast data corresponds to the global numerical weather forecast model, which has a better overall forecast effect for large-scale weather systems; the convective weather forecast data corresponds to the convective-resolving numerical weather forecast model, which has a better characterization ability for medium and small-scale convective systems. The forecast factors obtained based on the atmospheric circulation forecast data and the convective weather forecast data are the forecast factors for the fusion of numerical weather forecast models at different scales.

[0039] The target forecast factors are obtained from the atmospheric circulation forecast data and the convective weather forecast data. Among them, the forecast factors in the atmospheric circulation forecast data and the convective weather forecast data include basic elements such as temperature, pressure, humidity, and wind field from the ground to high altitude, as shown in Table 1 below, which will not be elaborated here one by one. Among them, the ECMWF model in Table 1 is the global numerical weather forecast model, and the GRAPES 3km model is the convective-resolving numerical weather forecast model.

[0040]

[0041]

[0042] Specifically, the target forecast factors for the multi-scale numerical weather forecast model fusion include the forecast factors in the global numerical weather forecast model and the convective-resolving numerical weather forecast model in Table 1. That is to say, the finally obtained target forecast factors include at least one forecast factor in the global numerical weather forecast model and one forecast factor in the convective-resolving numerical weather forecast model. The specific number of forecast factors under the two different models can be determined according to the actual situation and will not be specifically limited here.

[0043] It is understandable that this step fuses the relevant forecast factors of the global numerical weather prediction model and the convective-resolving numerical weather prediction model, which can effectively improve the weather forecast effect.

[0044] S12. Input the target forecast factor into the pre-trained deep learning forecast model to obtain the weather forecast result.

[0045] Based on the target forecast factor obtained by fusing the multi-scale numerical weather prediction models in the previous step, this step inputs the obtained target forecast factor into the pre-trained deep learning forecast model to obtain the predicted weather forecast.

[0046] It should be noted that deep learning has a powerful feature extraction ability. Applying it to the field of weather forecasting to extract the spatio-temporal variation characteristics of numerical weather prediction model data and observational data at different scales can effectively realize the fusion of multi-source observational data.

[0047] Figure 2 Shows the structural diagram of the deep learning forecast model provided by the present invention. As Figure 2 shown, the deep learning model provided in this embodiment is a three-dimensional model, including a convolutional layer, a residual connection layer, a pooling layer, an upsampling layer, and a self-attention layer, which can effectively extract the weather system characteristics in three-dimensional space. Specifically, during the deep learning process, the target forecast factors in different modes continuously undergo non-linear transformations, and the deeper the network output shows stronger characteristics. The "depth" of the deep learning network is very important. However, the more layers the network has, the more unstable the gradient will become during the backpropagation process. Therefore, a residual connection layer is set so that the model can automatically learn the identity mapping, thereby accelerating the training of the model. Setting the pooling layer and the upsampling layer can reduce the parameters and computational amount while retaining the main features, prevent overfitting, and can also expand the receptive field. The self-attention layer calculates the coefficients related to the actual situation at each grid point, so that the deep learning forecast model has better feature extraction ability.

[0048] In the deep learning forecast model of this embodiment, the self-attention layer is combined with the residual connection layer to construct a residual and attention mechanism unit, which can improve the model's ability to extract weather forecast characteristics at different scales while ensuring the training efficiency of the model.

[0049] The weather forecast result obtained in this embodiment may include temperature, humidity, wind force, air pressure, etc. according to needs, and is not specifically limited. For example, the weather forecast result is "Beijing, snow tomorrow, the highest daytime temperature is 3 degrees Celsius, the lowest nighttime temperature is minus 10 degrees Celsius, and the northwest wind is level 2-3". According to the weather forecast result, people can plan their travel and clothing in advance.

[0050] It should also be noted that the deep learning prediction model in this embodiment is pre-trained, and the loss function used in the training process of the model can be a logarithmic loss function or a cross entropy loss function, which is not specifically limited here.

[0051] The weather forecast method for multi-scale numerical weather forecast model fusion provided in this embodiment obtains the target forecast factor of the multi-scale numerical weather forecast model fusion from the atmospheric circulation forecast data and the convective weather forecast data, and inputs the obtained target forecast factor into the pre-trained deep learning forecast model to effectively fuse the forecast factors under the numerical weather forecast models of different scales, and further obtain the predicted weather forecast results. This method solves the defect of the prior art that the forecast effect of weather systems of different scales cannot be taken into account at the same time, realizes the fusion of multiple-scale numerical weather forecast models, and effectively improves the forecast effect of weather systems of different scales.

[0052] On the basis of the above embodiments, further, Figure 3 The ROC curve diagram of the prediction factors of the global numerical weather prediction model provided by the present invention is shown; Figure 4 The ROC curve of the prediction factor of the convective resolvable numerical weather forecast model provided by the present invention is shown. Among them, ROC (receiver operating characteristic curve) refers to the receiver operating characteristic curve. A ROC curve represents countless classifiers and describes the process of classifier performance changing with the change of classifier threshold. An important feature of the ROC curve is its area, that is, the AUC value.

[0053] According to the atmospheric circulation forecast data and the convective weather forecast data, target forecast factors for the fusion of the multi-scale numerical weather forecast model are obtained, including: selecting target forecast factors from the atmospheric circulation forecast data and the convective weather forecast data by a permutation importance analysis method.

[0054] It is understandable that fusing all variables in the global numerical weather prediction model and the convective-resolvable numerical weather prediction model can effectively improve the forecasting effect of weather systems of different scales. When there are sufficient training samples and computing power, it is feasible to input all samples. However, in actual situations, the available samples for multi-scale pattern matching in the convective-resolvable numerical weather prediction model are limited due to the short historical data; moreover, if different models provide the same variables, there will be more information redundancy. Therefore, it is very necessary to select the most effective forecast factor, that is, the target forecast factor, from the two models.

[0055] In this embodiment, the permutation importance analysis method is used to analyze the importance of each prognostic factor in the global numerical weather prediction model and the convective-resolving numerical weather prediction model, so as to select the most effective target prognostic factor for weather forecast prediction.

[0056] The permutation importance analysis method compares the test result one with a missing prognostic factor with the test result two with all prognostic factors participating in the test. The greater the difference between the two, the more important the prognostic factor is for weather forecast prediction; if the change between the two is not significant, it indicates that the prognostic factor is not important or there is information overlap between the prognostic factor and other prognostic factors, such as a linear correlation relationship. Through the permutation importance analysis method, the importance degree of each prognostic factor can be intuitively reflected. In the solution of the present invention, the prognostic factor with a greater impact on the forecast will be selected as the target prognostic factor, and the prognostic factor with an insignificant impact on the forecast will be excluded to avoid information redundancy and improve the calculation efficiency.

[0057] Taking precipitation forecast as an example, by analyzing Figure 3 and Figure 4 from the ROC curve diagrams shown, it can be seen that each prognostic factor in the global numerical weather prediction model and the convective-resolving numerical weather prediction model contains effective information for precipitation forecast, manifested as a different degree of decline in forecast performance in the case of missing.

[0058] Among them, in the prognostic factors of the global numerical weather prediction model, when specific humidity (Q), temperature (T), and geopotential height (H) are missing, the precipitation forecast performance shows an obvious decline, and the impact of temperature is the most obvious. Wind (UVW) has an insignificant impact on the forecast. Among the prognostic factors of the convective-resolving numerical weather prediction model, variables such as temperature (T), graupel (GRLE), and rain water mixing ratio (RWMR) have a greater impact on short-term heavy precipitation forecast. Among them, temperature is the most important for short-term heavy precipitation forecast.

[0059] In this embodiment, through the permutation importance analysis of the prognostic factors, the relatively important prognostic factors in different numerical models are selected as the target prognostic factors and input into the pre-trained deep learning forecast model, so as to obtain more accurate weather forecast results, realizing the fusion of multiple-scale numerical weather prediction models and effectively improving the forecast effect on weather systems of different scales.

[0060] On the basis of the above embodiment, further, Figure 5The ROC curve graph showing the predictors of different numerical weather prediction models used in the deep learning prediction model provided by the present invention is presented. The target predictors include predictors of atmospheric circulation characteristics and predictors of convective weather prediction characteristics; among them, the predictors of atmospheric circulation characteristics include temperature, geopotential height, and specific humidity; the predictors of convective weather prediction characteristics include horizontal wind field, vertical velocity, temperature, relative humidity, cloud water mixing ratio, rain water mixing ratio, ice water mixing ratio, snow water mixing ratio, and graupel.

[0061] After the importance analysis of the predictors, it is necessary to screen the predictors according to Figure 3 and Figure 4 the results of the importance analysis of the predictors. During the precipitation prediction process, the global numerical weather prediction model mainly provides accurate atmospheric circulation information to make the prediction location more accurate. The importance analysis shows that among all the predictors of the global numerical weather prediction model, the importance degrees of the three predictors of temperature (T), geopotential height (H), and specific humidity (Q) are also greater than those of factors such as horizontal wind field (U, V) and vertical velocity (W). Therefore, the three predictors of temperature, geopotential height, and specific humidity that can reflect the basic characteristics of the atmospheric circulation are selected.

[0062] Relatively speaking, the convective-allowing numerical weather prediction model mainly provides the prediction ability of severe convective systems to make the prediction intensity more accurate. Therefore, the predictors that can reflect the occurrence and development process of the convective system are selected. The cloud physical parameters are closely related to the convective process, and at the same time, their importance degrees as predictors are also relatively high. Therefore, all the cloud physical parameters are selected. In addition, factors such as horizontal wind field (U, V) and vertical velocity (W) also reflect the prediction ability of the convective-allowing numerical weather prediction model for the convective system structure. Referring to the results of the importance analysis, they are retained.

[0063] The triggering and development of the convective system are closely related to conditions such as specific humidity and temperature in the upper and lower layers. Therefore, the relative importance degrees of temperature (T) and specific humidity (Q) in the convective-allowing numerical weather prediction model are relatively high. Especially temperature is the most important predictor. Therefore, temperature (T) and specific humidity (Q) in the convective-allowing numerical weather prediction model are selected. The importance degree of geopotential height (H) is the lowest among all the predictors of the convective-allowing numerical weather prediction model. This may be because the global numerical weather prediction model has provided a relatively accurate geopotential height situation field, and the convective environment information can also be better obtained from other predictors. Therefore, its importance degree is relatively low. Therefore, the geopotential height factor in the convective-allowing numerical weather prediction model is not selected.

[0064] Therefore, the finally selected target predictors include predictors of atmospheric circulation characteristics and predictors of convective weather forecast characteristics. Among them, the predictors of atmospheric circulation characteristics include temperature, geopotential height, and specific humidity. The predictors of convective weather forecast characteristics include horizontal wind field, vertical velocity, temperature, relative humidity, cloud water mixing ratio, rain water mixing ratio, ice water mixing ratio, snow water mixing ratio, and graupel.

[0065] According to Figure 5 it can be seen that although the number of predictors has decreased, the forecast performance has been improved to a certain extent compared with using all predictors. The ROC curve using the target predictors selected from the two models completely envelopes the curve using all predictors, and the value of AUC increases from 0.915 to 0.935. The value of AUC is the area under the ROC curve, and the larger the value of AUC, the better the forecast performance. This also indicates that under the condition that other objective conditions are the same, such as using the same deep learning forecast model, the redundancy of data information not only cannot improve the forecast performance, but may even lead to a decline in the forecast performance.

[0066] In addition, according to Figure 5 it can also be concluded that the model for selecting predictors also has obvious performance advantages compared with the model using a single global numerical weather prediction model or a convective-resolving numerical weather prediction model.

[0067] In this embodiment, by using the permutation importance analysis method, the predictors of the global numerical weather prediction model and the convective-resolving numerical weather prediction model are obtained as target predictors and input into a pre-trained deep learning forecast model, so as to obtain more accurate weather forecast results, realize the fusion of multiple-scale numerical weather prediction models, and effectively improve the forecast effect on weather systems of different scales.

[0068] On the basis of the above embodiment, further, according to the target predictors, a training set and a test set are determined; the training set and the test set are respectively input into the deep learning forecast model for training and testing to obtain the pre-trained deep learning forecast model.

[0069] It can be understood that according to the determined target predictors, a training set and a test set are constructed; the constructed deep learning forecast model is trained and tested by using the training set and the test set to obtain the trained deep learning forecast model; finally, the target predictors are input into the trained deep learning forecast model, and the corresponding weather forecast can be obtained.

[0070] For the determination of the training set and the test set, in a specific embodiment, still taking precipitation forecasting as an example, using the global numerical weather prediction model data, convective-resolving numerical weather prediction model analysis field data, and precipitation observation data from March to September in 2019 and 2020, a training set is constructed, containing 2018 samples. The forecast experiment range is 18 - 42°N, 102 - 126°E.

[0071] Among them, the training samples of the global numerical weather prediction model data are Nh×L×W×Np. Among them, Nh is the number of levels (from the ground to 200hPa, a total of 10 levels), L is the length of the training area, W is the width of the training area, and Np is the number of target forecast factors in the sample. There is no vertical velocity variable in the surface elements, and the altitude with a resolution of 25km is added, and finally a four-dimensional array of 10×97×97×Np is formed.

[0072] Due to the difference in resolution, the training samples of the convective-resolving numerical weather prediction model data are Nh×801×801×Np. Among them, Nh is the number of levels (from the ground to 200hPa, a total of 10 levels), and Np is the number of forecast factors in the sample. There is no vertical velocity variable in the surface elements, and the altitude with a resolution of 3km is added.

[0073] The test set selects July 2019 and June 2020 with stronger precipitation to verify the forecast effect of the model, and the rest are used as the training set.

[0074] Since the number of training samples is limited, a strategy of randomly selecting training areas and using small training areas to achieve large forecast areas can be adopted for training. In a specific embodiment, first, the plane scale of the training area input during the training process is set to 1200×1200km; then, the training area is randomly selected, and from the sample area of 2400×2400km, the training area is randomly selected with a sliding window of 100km; since the input scale of the deep learning forecast model is variable, finally, an area of 2400×2400km or larger can be forecast. In this way, on the one hand, the sample capacity can be effectively expanded, and on the other hand, the training content overhead of the graphics card can be reduced. It is estimated that the number of samples of 1200×1200km divided in this way is about 18162.

[0075] The training of the deep learning forecast model is set with 100 iteration cycles and adopts the Earlystop strategy. When the loss of the model stops decreasing continuously for more than 10 iteration cycles, it will automatically stop, and the weights with the smallest loss on the validation set are saved.

[0076] In this embodiment, a training set and a test set are determined through pre-selected target prediction factors, and the deep learning prediction model is trained and tested with the training set and the test set as inputs, so as to obtain a trained deep learning prediction model, thereby being able to better organically integrate multiple-scale numerical weather prediction models and effectively improving the prediction effect of weather systems at different scales.

[0077] Based on the above embodiment, further, inputting the training set and the test set into the deep learning prediction model for training and testing respectively includes: inputting the target prediction factors in the training set into the corresponding encoders of the deep learning prediction model; inputting the output results of all encoders into the same decoder of the deep learning prediction model to obtain the fusion features of the target prediction factors; training the deep learning prediction model according to the fusion features;

[0078] inputting the target prediction factors in the test set into the corresponding encoders of the deep learning prediction model; inputting the output results of all encoders into the same decoder of the deep learning prediction model to obtain the test fusion features of the target prediction factors; testing the deep learning prediction model according to the test fusion features.

[0079] It can be understood that the deep learning prediction model in this embodiment uses a U-shaped network structure with an encoding-decoding structure. During the encoding process, the deep learning prediction model can continuously extract favorable information for the occurrence of a certain weather. After 4 pooling processes, the feature map is continuously compressed, and finally the three-dimensional size is reduced to 1 / 16 of the original. During the encoding process, the analysis and judgment of the prediction factors of different-scale numerical weather prediction models are realized.

[0080] It should be noted that after 4 pooling processes, the feature map of the target prediction factors is continuously compressed, reduced from the three-dimensional size to 1 / 16 of the original. Some detailed features of the compressed feature map will be lost, and the decoding process can restore it and learn the detailed features, finally realizing the weather prediction for each grid point.

[0081] In this embodiment, there are prediction factors of two models, namely the global numerical weather prediction model and the convective-resolving numerical weather prediction model. Therefore, different encoders are set for the target prediction factors of the two models. The target prediction factors of the two different models in the training set are respectively input into the corresponding encoders. The encoders analyze and judge the target prediction factors, and input the output results after the analysis and judgment of all encoders into the same decoder of the deep learning prediction model to fuse the target prediction factors under the two different models, so as to obtain the fusion features of the target prediction factors under the two models. Finally, the deep learning prediction model is trained according to the obtained fusion features.

[0082] After training a deep learning prediction model, it doesn't mean that the model can be directly put into use. It still needs to be tested to effectively ensure the final prediction effect. Similarly, for the testing of the deep learning prediction model, different encoders are set for the target prediction factors of the two modes respectively. The target prediction factors of the two different modes in the test set are respectively input into the corresponding encoders. The encoders analyze and judge the target prediction factors, and the test output results after all encoders' analysis and judgment are input into the same decoder of the deep learning prediction model to fuse the target prediction factors in the two different modes, so as to obtain the test fusion features of the target prediction factors in the two modes. Finally, based on the obtained test fusion features, the deep learning prediction model is tested.

[0083] It should be noted that different encoders can be set for the prediction factors of the two modes according to the number of selected prediction factors in different modes. For example, every one or every two target prediction factors in the same mode correspond to the same encoder. In a specific embodiment, the prediction factors temperature, specific humidity, and geopotential height in the global numerical weather prediction mode and the prediction factors vertical velocity, cloud water mixing ratio, rain water mixing ratio, ice water mixing ratio, and snow water mixing ratio in the convective-resolving numerical weather prediction mode are determined as the target prediction factors. Among them, there are three target prediction factors in the global numerical weather prediction mode, and the number of target prediction factors in the convective-resolving numerical weather prediction mode is five. In this case, two encoders corresponding to the prediction factors in the global numerical weather prediction mode are set, and three encoders corresponding to the prediction factors in the convective-resolving numerical weather prediction mode are set, that is, randomly two prediction factors in the same mode correspond to the same encoder, and they are allocated in turn. The remaining one target prediction factor corresponds to a single encoder alone.

[0084] For the training and testing of the deep learning prediction model, corresponding thresholds can be set. When the training and / or testing results of the deep learning prediction model reach a certain threshold, it indicates that the deep learning prediction model has been trained well and / or has passed the test. Otherwise, continue to train and / or test it.

[0085] In this embodiment, by using the determined training set and test set as inputs to train and test the deep learning prediction model, a well-trained and tested deep learning prediction model can be obtained, which can better organically fuse multiple-scale numerical weather prediction modes, thereby effectively improving the prediction effect on weather systems of different scales.

[0086] On the basis of the above embodiment, further, the loss functions used for training the deep learning prediction model include the cross-entropy loss function and loss function L CSI; Loss function L CSI The formula is as follows:

[0087]

[0088] where CSI is the critical success index, y i is the class label of grid point i, p i is the probability of a certain weather occurring at predicted grid point i, N is the number of grid points, and γ is a constant.

[0089] During the training of the deep learning prediction model, the loss function is used to define the performance of the deep learning prediction model. For each training sample, a predicted value is obtained by passing it through the neural network, and then this predicted value is compared with the true value we want to obtain. The smaller the loss function, the smaller the difference between the predicted value and the true value, indicating that the deep learning prediction model is better trained.

[0090] In the training of the neural network model, the commonly used binary classification loss function is the cross-entropy loss function, and its formula is as follows:

[0091]

[0092] where y i is the class label of grid point i. For example, when a certain weather occurs, y i can be set to 1; otherwise y i = 0; p i is the probability of a certain weather occurring at predicted grid point i, and N is the number of grid points.

[0093] The cross-entropy loss function has a good effect on the classification of grid points. At the same time, in order to improve the overall prediction ability of the model for weather systems and considering that CSI is an important index for forecast service verification, in this embodiment, a loss function L CSI based on CSI is also constructed, where the formula of the loss function L CSI is:

[0094]

[0095] where CSI is the critical success index, y i is the class label of grid point i, p i is the probability of a certain weather occurring at predicted grid point i, N is the number of grid points, and γ is a constant used to smooth the loss function, usually set to 1.

[0096] Specifically, the cross-entropy loss function L CE focuses on making predictions for each grid point, while the loss function L TSIt can enhance the model's forecasting ability for large-scale weather systems. On this basis, this embodiment also combines the two, so as to increase the overall forecasting ability of the model from mesoscale to large-scale weather systems, that is, the loss function L can also be:

[0097] L = L CE + L CSI

[0098] It can be understood that when training the deep learning forecasting model in this embodiment, the loss functions used include the cross-entropy loss function and the CSI-based loss function. Among them, either the cross-entropy loss function or the CSI-based loss function can be used alone, or the two can be combined, and no specific limitation is made here.

[0099] In this embodiment, by using the cross-entropy loss function and the CSI-based loss function to train the deep learning forecasting model to obtain a trained deep learning forecasting model, it can better organically integrate multiple-scale numerical weather prediction models, thereby effectively improving the forecasting effect on weather systems of different scales.

[0100] Furthermore, Figure 6 shows the preliminary fusion result diagram of the global numerical weather prediction model and the convective-resolving numerical weather prediction model provided by the present invention.

[0101] In this specific embodiment, all the forecasting factors in the global numerical weather prediction model and the convective-resolving numerical weather prediction model in Table 1 are input into the deep learning forecasting model to obtain a trained deep learning forecasting model. At the same time, in order to compare the performance differences before and after the fusion of different models, the forecasting factors under the global numerical weather prediction model and the convective-resolving numerical weather prediction model are trained separately, and the corresponding forecasting models are obtained.

[0102] The specific results are as Figure 6 shown. From Figure 6 it can be obtained that by using the deep learning method and using the forecasting factors in the global numerical weather prediction model and the convective-resolving numerical weather prediction model respectively, forecasting models with good forecasting performance can be obtained, and their AUCs reach 0.876 and 0.887 respectively, indicating that they have high forecasting skills.

[0103] The deep learning model fuses the predictors of the global numerical weather prediction model and the convective - resolution numerical weather prediction model, which is the deep learning prediction model in the present invention. Its prediction performance has been significantly improved. The value of AUC has increased to 0.915, and the ROC curve completely encompasses the ROC curves of the global numerical weather prediction model and the convective - resolution numerical weather prediction model. Thus, it can be seen that by using the deep learning method to fuse numerical weather prediction models at different scales, the prediction effect of a single model on weather systems can be effectively enhanced.

[0104] Based on the above - mentioned embodiments, Figure 7 Figure 5 shows a comparison chart of the prediction situations of different models provided by the present invention. Among them, Figure (a) is the precipitation prediction under the convective - resolution numerical weather prediction model; Figure (b) is the heavy precipitation probability prediction under the convective - resolution numerical weather prediction model; Figure (c) is the heavy precipitation probability prediction under the global numerical weather prediction model; Figure (d) is the heavy precipitation probability prediction after fusing numerical weather prediction models at different scales.

[0105] The present invention also provides a specific precipitation prediction case on June 2, 2020.

[0106] On June 2, 2020, there was a relatively large - scale precipitation weather in South China, and obvious precipitation processes occurred in Guizhou, Guangxi, Hunan and other places. Among them, the northern part of Guangxi was in the center of heavy precipitation, and the short - time heavy precipitation with an intensity exceeding 20 mm / h had a large range. After analysis, it was found that the heavy precipitation process in Guangxi was a typical warm - sector precipitation, which was a short - time heavy precipitation weather triggered by local heat convection without the influence of obvious weather systems.

[0107] Regarding the comparison of the precipitation prediction situation of this time as Figure 7 , where the filled color is the actual precipitation distribution, and the contour lines are the short - time heavy precipitation prediction probability distribution. From the comparison of the four predictions: The convective - resolution numerical weather prediction model has better strong precipitation analysis ability. By using deep learning, the characteristic signals of this heavy precipitation can be extracted, thus achieving a better prediction of the heavy precipitation process in the northern part of Guangxi (Figure (a)). Although the precipitation product of the convective - resolution numerical weather prediction model itself has a smaller predicted area, it also successfully predicted this strong convection process to a certain extent (Figure (b)).

[0108] For this warm - sector heavy precipitation process, the global numerical weather prediction model basically has no prediction ability (Figure (c)). The global numerical weather prediction model missed the heavy precipitation process in the northern part of Guangxi, indicating that the global numerical weather prediction model has relatively weak prediction ability for heavy precipitation generated by local convective systems.

[0109] By using deep learning to fuse the global numerical weather prediction model and the convective-resolving numerical weather prediction model, the prediction of heavy precipitation processes in the northern region of Guangxi is better achieved (Figure (d)). After fusion, the heavy precipitation falling areas predicted by the algorithm are more accurate. Compared with the convective-resolving numerical weather prediction model, the false alarms are reduced; compared with the global numerical weather prediction model, the hit rate is increased. Therefore, overall, the prediction performance is significantly improved compared with that of a single numerical weather prediction model.

[0110] It can be seen from this that this embodiment well demonstrates the advantages of multi-scale numerical weather prediction model fusion: Although the global numerical weather prediction model can predict large-scale weather systems with relatively accurate impacts, its prediction ability for when and how convection is triggered is weak; the convective-resolving numerical weather prediction model has the ability to predict convection, but there will be a certain deviation in the predicted location. After the fusion of different numerical weather prediction models, better predictions of the location and intensity of heavy precipitation can be achieved.

[0111] In addition, the embodiment of the present invention also systematically verified the precipitation prediction results in June 2020 by using the traditional point-to-point verification scoring method.

[0112] Classic meteorological prediction verification factors, such as POD (probability of detect), false alarm rate FAR (false alarm rate), critical success index CSI (critical success index), accuracy (Accuracy), bias (Bias), and ETS (equitable threat score). Specifically, the systematic verification results carried out by the present invention are shown in Table 2 below:

[0113]

[0114]

[0115] Table 2 compares the scoring situations of deep learning using different models, multi-model fusion, and the original precipitation prediction of GRAPES 3km. It can be seen from Table 2 that by using the deep learning method, based on the basic variables of the global numerical weather prediction model and the convective-resolving numerical weather prediction model, effective fitting of short-term heavy precipitation can be achieved, thus obtaining more accurate prediction results than GRAPES 3km. That is to say, for short-term heavy precipitation weather, by using the deep learning method, relevant features of heavy precipitation weather systems can be effectively extracted, thus achieving a prediction ability superior to that of the global model based on the convective parameterization scheme and the convective-resolving numerical weather prediction model.

[0116] In the case of fusing all variables of the global numerical weather prediction model and the convective-resolving numerical weather prediction model, the forecast score has been significantly improved. After the selection of forecast factors, the forecast performance has been further enhanced. Relatively speaking, the results of the multi-scale numerical weather prediction model fusion forecast have increased by 80.0% and 66.7% respectively compared with the results of only using the global numerical weather prediction model or the convective-resolving numerical weather prediction model.

[0117] The batch inspection results show that the POD of the deep learning-based forecast using the global numerical weather prediction model is low, and its ability to capture heavy precipitation events is limited; the POD of the deep learning-based forecast using the convective-resolving numerical weather prediction model has been significantly improved, but the FAR and BIAS are high, and there are too many false alarms; after the fusion of the two numerical weather prediction models, while the POD has increased, the FAR has been reduced to a certain extent, effectively improving the overall forecast performance. Therefore, the weather forecasting method using deep learning for multi-scale numerical weather prediction model fusion can effectively improve the forecast effect.

[0118] Figure 8 The structural schematic diagram of the weather forecasting device for multi-scale numerical weather prediction model fusion provided by the present invention is shown. As Figure 8 shown, the device includes: a forecast factor acquisition module 81, configured to acquire target forecast factors for multi-scale numerical model fusion according to the atmospheric circulation forecast data and the convective weather forecast data; a forecast result acquisition module 82, configured to input the target forecast factors into a pre-trained deep learning forecast model to obtain the weather forecast result.

[0119] A weather forecasting device for multi-scale numerical weather prediction model fusion provided in this embodiment corresponds to and refers to the weather forecasting method for multi-scale numerical weather prediction model fusion described above, and will not be elaborated here one by one.

[0120] Figure 9 The entity structural schematic diagram of an electronic device provided by the present invention is shown. As Figure 9As shown in the figure, the electronic device may include: a processor 910, a communications interface 920, a memory 930, and a communication bus 940. Among them, the processor 910, the communications interface 920, and the memory 930 complete communication with each other through the communication bus 940. The processor 910 may call the logical instructions in the memory 930 to execute a weather forecasting method for multi-scale numerical weather forecasting model fusion. The method includes: obtaining target forecasting factors for multi-scale numerical weather forecasting model fusion according to the general circulation forecast data and convective weather forecast data; inputting the target forecasting factors into a pre-trained deep learning forecasting model to obtain a weather forecasting result.

[0121] In addition, when the logical instructions in the above-mentioned memory 930 are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0122] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the weather forecasting method for multi-scale numerical weather forecasting model fusion provided by the above-mentioned various methods. The method includes: obtaining target forecasting factors for multi-scale numerical weather forecasting model fusion according to the general circulation forecast data and convective weather forecast data; inputting the target forecasting factors into a pre-trained deep learning forecasting model to obtain a weather forecasting result.

[0123] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements a weather forecasting method for multi-scale numerical weather forecasting model fusion provided by the above-mentioned various methods. The method includes: obtaining target forecasting factors for multi-scale numerical weather forecasting model fusion based on general circulation forecasting data and convective weather forecasting data; inputting the target forecasting factors into a pre-trained deep learning forecasting model to obtain a weather forecasting result.

[0124] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0125] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A weather forecasting method integrating multi-scale numerical weather prediction models, characterized in that, Including: According to the general circulation forecast data and convective weather forecast data, obtain the target forecast factors for the fusion of multi-scale numerical weather prediction models through the permutation importance analysis method; Among them, the general circulation forecast data corresponds to the global numerical weather prediction model, the convective weather forecast data corresponds to the convective-resolving numerical weather prediction model, and the multi-scale numerical weather prediction model includes the global numerical weather prediction model and the convective-resolving numerical weather prediction model; The target forecast factors include forecast factors of general circulation characteristics and forecast factors of convective weather forecast characteristics; the forecast factors of general circulation characteristics include temperature, geopotential height, and specific humidity; the forecast factors of convective weather forecast characteristics include horizontal wind field, vertical velocity, temperature, relative humidity, cloud water mixing ratio, rain water mixing ratio, ice water mixing ratio, snow water mixing ratio, and graupel; Input the target forecast factors into a pre-trained deep learning forecast model to obtain the weather forecast result; among them, the deep learning forecast model is a "U"-shaped network with a dual encoder-single decoder structure. Different encoders are set for the target forecast factors under the global numerical weather prediction model and the convective-resolving numerical weather prediction model. Input the target forecast factors of the two different models into the corresponding encoders respectively. The encoder analyzes and judges the target forecast factors, and inputs the output results after all encoders' analysis and judgment into the same decoder of the deep learning forecast model to fuse the target forecast factors under the two different models.

2. The weather forecasting method for multi-scale numerical weather forecasting model fusion according to claim 1, characterized in that Also including: Determine the training set and the test set according to the target forecast factors; Input the training set and the test set into the deep learning forecast model for training and testing respectively to obtain the pre-trained deep learning forecast model.

3. The weather forecasting method for multi-scale numerical weather forecasting model fusion according to claim 2, characterized in that, The inputting the training set and the test set into the deep learning forecast model for training and testing respectively includes: Input the target forecast factors in the training set into the corresponding encoders of the deep learning forecast model; input the output results of all the encoders into the same decoder of the deep learning forecast model to obtain the fusion features of the target forecast factors; train the deep learning forecast model according to the fusion features; Input the target forecast factors in the test set into the corresponding encoders of the deep learning forecast model; input the output results of all the encoders into the same decoder of the deep learning forecast model to obtain the test fusion features of the target forecast factors; test the deep learning forecast model according to the test fusion features.

4. The weather forecasting method for multi-scale numerical weather forecasting model fusion according to claim 1, characterized in that, The loss function used for training the deep learning prediction model includes a cross-entropy loss function and a loss function L CSI ; The loss function L CSI has the following formula: where CSI is the critical success index, y i is the class label of grid point i, p i is the probability of a certain weather occurring at grid point i, N is the number of grid points, and γ is a constant.

5. A weather forecasting device for fusing multi-scale numerical weather prediction models, characterized in that, Including: A forecast factor acquisition module, configured to obtain the target forecast factors for the fusion of multi-scale numerical weather prediction models according to the general circulation forecast data and the convective weather forecast data through the permutation importance analysis method; Among them, the general circulation forecast data corresponds to a global numerical weather prediction model, the convective weather forecast data corresponds to a convective-resolving numerical weather prediction model, and the multi-scale numerical weather prediction model includes a global numerical weather prediction model and a convective-resolving numerical weather prediction model; The target forecast factors include forecast factors of general circulation characteristics and forecast factors of convective weather forecast characteristics; the forecast factors of general circulation characteristics include temperature, geopotential height, and specific humidity; the forecast factors of convective weather forecast characteristics include horizontal wind field, vertical velocity, temperature, relative humidity, cloud water mixing ratio, rain water mixing ratio, ice water mixing ratio, snow water mixing ratio, and graupel; A forecast result acquisition module, configured to input the target forecast factors into a pre-trained deep learning forecast model to obtain a weather forecast result; wherein, the deep learning forecast model is a "U"-shaped network with a dual encoder-single decoder structure, and different encoders are respectively set for the target forecast factors under the global numerical weather prediction model and the convective-resolving numerical weather prediction model. The target forecast factors of the two different models are respectively input into the corresponding encoders, and the encoders analyze and judge the target forecast factors, and the output results after the analysis and judgment of all encoders are input into the same decoder of the deep learning forecast model to fuse the target forecast factors under the two different models.

6. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, the steps of the weather forecasting method for multi-scale numerical weather prediction model fusion according to any one of claims 1 to 4 are implemented.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the weather forecasting method for multi-scale numerical weather prediction model fusion according to any one of claims 1 to 4 are implemented.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, the steps of the weather forecasting method for multi-scale numerical weather prediction model fusion according to any one of claims 1 to 4 are implemented.

Images