Strong wind forecasting method and system based on deep learning fusion radar

Abstract

The invention provides a deep learning fusion radar-based strong wind forecasting method and system, and the method comprises the steps: predicting the radar reflectivity factor of a to-be-forecasted region in a future time period through a radar reflectivity factor forecasting model according to the radar reflectivity factor of the to-be-forecasted region in a historical time period; according to the wind speed of the to-be-forecasted area in the historical time period and the radar reflectivity factor in the future time period, using a wind speed forecasting model to predict the wind speed of the to-be-forecasted area in the future time period; the radar reflectivity factor forecasting model and the wind speed forecasting model are obtained by training a deep learning model. According to the invention, the gale forecasting precision is improved and physical consistency is met.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting, and in particular to a gale forecasting method and system based on deep learning fusion radar. Background Art

[0002] Strong winds are highly disastrous and occur in a short time, causing serious damage to life and property safety, and also have a great impact on the safety of power facilities and the operation of public basic services. Therefore, monitoring and early warning of strong winds is an important demand facing the current weather forecast business.

[0003] However, due to the following factors, monitoring and early warning of strong winds are still difficult. First, the observation network has limited coverage. The number and quality of conventional ground wind observation stations are insufficient to support the monitoring of rapid strong wind processes with a small spatial scale. Secondly, early warning and forecasting are difficult. The mechanism of strong wind formation is complex. For example, downbursts or tornadoes may cause disastrous strong winds. At present, there is still a lot of lack of understanding of these complex physical processes and their generation and disappearance mechanisms.

[0004] In solving the technical problems of gale forecasting, previous methods have significant limitations. Existing methods mainly include statistical methods, numerical model methods and the recently popular deep learning extrapolation forecasting method. However, the above three methods all have certain limitations at present. The first is the statistical method. Statistical methods generally select gale-influencing variables for regression calculation, but there are too many variables of factors affecting the atmosphere and land surface on gale, and the selected factors are often highly subjective, making it difficult to describe the quantitative relationship between gale and factors. In addition, there is a large uncertainty in the universality of this prediction method, and the factor weights will change due to different scenarios. Secondly, the numerical model method, because the current model structure and parameterization do not have a good description ability for the atmospheric dynamic formation mechanism of gale and the impact mechanism of complex underlying surface, and the model depends on the accuracy of the initial field, that is, accurate gale model forecasting requires high-resolution initial field information and accurate model parameterization and structural representation. Finally, the deep learning extrapolation forecasting method that has become increasingly popular in recent years, this method uses the highly nonlinear expression ability of deep learning algorithms, and the forecasting effect is better than traditional statistical methods. However, previous technical means either only considered the extrapolation of wind speed data, or simply merged radar data with wind speed data at the input layer. Therefore, it is easy to overfit, and the forecast is not physically consistent, and there are large errors in the initial estimates of storm location and intensity.

[0005] Therefore, it is extremely challenging to provide a new method that can fully integrate radar observation data information and effectively estimate future storm information, thereby improving forecast accuracy and making the forecast physically consistent. Summary of the invention

[0006] The present invention provides a gale forecasting method and system based on deep learning fusion radar, so as to solve the defects in the prior art that the gale forecasting accuracy is not high and the forecasting does not conform to physical consistency, so as to improve the gale forecasting accuracy and make the forecast conform to physical consistency.

[0007] The present invention provides a gale forecasting method based on deep learning fusion radar, comprising:

[0008] According to the radar reflectivity factor of the area to be predicted in the historical time period, using the radar reflectivity factor prediction model to predict the radar reflectivity factor of the area to be predicted in the future time period;

[0009] Predicting the wind speed of the area to be forecasted in the future time period using a wind speed forecast model according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period;

[0010] The radar reflectivity factor prediction model and the wind speed prediction model are obtained by training a deep learning model.

[0011] According to a method for forecasting gale winds based on deep learning fusion radar provided by the present invention, according to the radar reflectivity factor of the area to be forecasted in the historical time period, a radar reflectivity factor forecasting model is used to predict the radar reflectivity factor of the area to be forecasted in the future time period, including:

[0012] Using the first radar reflectivity factor training set to train multiple deep learning models, to obtain corresponding multiple radar reflectivity factor prediction models;

[0013] Use the first radar reflectivity factor validation set to score each radar reflectivity factor prediction model, and take the radar reflectivity factor prediction model with the highest score as the optimal radar reflectivity factor prediction model;

[0014] According to the radar reflectivity factor of the area to be predicted in a historical time period, the optimal radar reflectivity factor prediction model is used to predict the radar reflectivity factor of the area to be predicted in a future time period.

[0015] According to a method for forecasting strong winds based on deep learning fusion radar provided by the present invention, according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period, a wind speed forecast model is used to predict the wind speed of the area to be forecasted in the future time period, including:

[0016] Using the optimal radar reflectivity factor prediction model to predict the radar reflectivity factor prediction result corresponding to the second radar reflectivity factor training set;

[0017] Using the wind speed data training set corresponding to the second radar reflectivity factor training set in time and space and the radar reflectivity factor forecast result, multiple deep learning models are trained to obtain corresponding multiple wind speed forecast models.

[0018] Using the second radar reflectivity factor verification set and the wind speed data verification set corresponding to the second radar reflectivity factor verification set in time and space to score each wind speed forecast model, and taking the wind speed forecast model with the highest score as the optimal wind speed forecast model;

[0019] The wind speed of the area to be forecasted in the future time period is predicted using the optimal wind speed forecast model according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period.

[0020] According to a method for forecasting gale winds based on deep learning fusion radar provided by the present invention, according to the radar reflectivity factor of the area to be forecasted in the historical time period, a radar reflectivity factor forecasting model is used to predict the radar reflectivity factor of the area to be forecasted in the future time period, including:

[0021] Using a bilinear interpolation method, the radar reflectivity factor of the area to be forecasted in the historical time period is interpolated to Cartesian coordinates to form radar reflectivity factor grid data;

[0022] According to the radar reflectivity factor gridded data of the area to be predicted in a historical time period, a radar reflectivity factor prediction model is used to predict the radar reflectivity factor gridded data of the area to be predicted in a future time period.

[0023] According to a method for forecasting strong winds based on deep learning fusion radar provided by the present invention, according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period, a wind speed forecast model is used to predict the wind speed of the area to be forecasted in the future time period, including:

[0024] Using a bilinear interpolation method, the wind speed of the area to be forecasted within the historical time period is interpolated to Cartesian coordinates to form wind speed grid data;

[0025] The wind speed gridded data of the area to be forecasted in the future time period is predicted using a wind speed forecast model according to the gridded wind speed data of the area to be forecasted in the historical time period and the gridded radar reflectivity factor data in the future time period.

[0026] According to a gale forecasting method based on deep learning fusion radar provided by the present invention, the loss function for training multiple deep learning models using the first radar reflectivity factor training set is:

[0027]

[0028] in, is the loss function value, n is the number of samples in the first radar reflectivity factor training set, and They are respectively the radar reflectivity factor prediction result and the actual value of the radar reflectivity factor of the sample output by the deep learning model.

[0029] According to a gale forecasting method based on deep learning fusion radar provided by the present invention, the score of the radar reflectivity factor prediction model is a CSI score, the first radar reflectivity factor verification set is grid data, and each radar reflectivity factor prediction model is scored using the first radar reflectivity factor verification set by the following formula:

[0030] CSI = TP / (TP+FA+MS)

[0031] Among them, CSI is the score of the radar reflectivity factor prediction model, TP represents the number of grid points where both the observed and predicted radar reflectivity factors are within the preset range, MS represents the number of grid points where the observed radar reflectivity factor is within the preset range but the predicted radar reflectivity factor is not within the preset range, and FA represents the number of grid points where the observed radar reflectivity factor is not within the preset range but the predicted radar reflectivity factor is within the preset range.

[0032] According to a high wind forecasting method based on deep learning fusion radar provided by the present invention, the score of the wind speed forecasting model is RMSE score.

[0033] The present invention also provides a gale forecasting system based on deep learning fusion radar, comprising:

[0034] The first prediction module is used to predict the radar reflectivity factor of the area to be predicted in a future time period using a radar reflectivity factor prediction model according to the radar reflectivity factor of the area to be predicted in a historical time period;

[0035] A second prediction module is used to predict the wind speed of the area to be predicted in the future time period using a wind speed prediction model according to the wind speed of the area to be predicted in the historical time period and the radar reflectivity factor in the future time period;

[0036] The radar reflectivity factor prediction model and the wind speed prediction model are obtained by training a deep learning model.

[0037] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a method for forecasting strong winds based on deep learning fusion radar as described above is implemented.

[0038] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for forecasting strong winds based on deep learning fusion radar as described in any one of the above is implemented.

[0039] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned methods for forecasting strong winds based on deep learning fusion radar.

[0040] The gale forecasting method and system based on deep learning fusion radar provided by the present invention obtain the initial guess value of the radar reflectivity factor in the future time period according to the radar reflectivity factor in the historical time period by using a pre-trained radar reflectivity factor prediction model, and make full use of the physical information of the radar reflectivity factor to provide important falling area and intensity information for gale forecasting; then combine the initial guess value of the radar reflectivity factor in the future time period with the wind speed data in the historical time period, and use the pre-trained wind speed prediction model to perform wind speed forecasting, so as to improve the accuracy of gale forecasting and meet physical consistency, and provide effective information for gale business warning and forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 This is one of the flow charts of the gale forecasting method based on deep learning fusion radar provided by the present invention;

[0043] Figure 2 It is a schematic diagram of the model structure of the gale forecasting method based on deep learning fusion radar provided by the present invention;

[0044] Figure 3 This is the second flow chart of the gale forecasting method based on deep learning fusion radar provided by the present invention;

[0045] Figure 4 This is a forecast example diagram of the gale forecast method based on deep learning fusion radar provided by the present invention;

[0046] Figure 5It is a structural schematic diagram of a strong wind forecasting system based on deep learning fusion radar provided by the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Combine the following Figure 1 A method for forecasting strong winds based on deep learning and radar fusion of the present invention is described, comprising:

[0049] Step 101, predicting the radar reflectivity factor of the area to be predicted in a future time period using a radar reflectivity factor prediction model according to the radar reflectivity factor of the area to be predicted in a historical time period;

[0050] Step 102, predicting the wind speed of the area to be forecasted in the future time period using a wind speed forecast model according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period;

[0051] The radar reflectivity factor prediction model and the wind speed prediction model are obtained by training a deep learning model.

[0052] like Figure 2 As shown, according to the radar reflectivity factor of the historical time period, the pre-trained radar reflectivity factor prediction model is used to obtain the initial guess value of the radar reflectivity factor in the future time period (such as the next 1 hour); the initial guess value of the radar reflectivity factor in the future time period is combined with the wind speed data of the historical time period, and the wind speed forecast model is used to predict the wind speed. Combined with the pre-trained radar reflectivity factor prediction model and the wind speed forecast model, the strong wind forecast result that integrates the radar pre-training results is finally obtained, providing effective information for strong wind business warning and forecasting.

[0053] This embodiment uses a pre-trained radar reflectivity factor prediction model to obtain an initial guess of the radar reflectivity factor for the future time period based on the radar reflectivity factor of the historical time period, and makes full use of the physical information of the radar reflectivity factor to provide important falling area and intensity information for the strong wind forecast; then the initial guess of the radar reflectivity factor for the future time period is combined with the wind speed data of the historical time period, and the pre-trained wind speed forecast model is used to predict the wind speed, so as to improve the accuracy of the strong wind forecast and meet the physical consistency, and provide effective information for the strong wind business warning and forecast.

[0054] Based on the above embodiments, Figure 3 As shown, in this embodiment, according to the radar reflectivity factor of the area to be predicted in the historical time period, the radar reflectivity factor prediction model is used to predict the radar reflectivity factor of the area to be predicted in the future time period, including:

[0055] Using the first radar reflectivity factor training set to train multiple deep learning models, to obtain corresponding multiple radar reflectivity factor prediction models;

[0056] Use the first radar reflectivity factor validation set to score each radar reflectivity factor prediction model, and take the radar reflectivity factor prediction model with the highest score as the optimal radar reflectivity factor prediction model;

[0057] According to the radar reflectivity factor of the area to be predicted in a historical time period, the optimal radar reflectivity factor prediction model is used to predict the radar reflectivity factor of the area to be predicted in a future time period.

[0058] First, a radar reflectivity factor dataset is established. The time span of each sample can be 3 hours, including 1 hour of input and 2 hours of forecast. The dataset can be divided into the first radar reflectivity factor training set and the first radar reflectivity factor verification set in a ratio of 8:2.

[0059] Optionally, the radar reflectivity factors in the radar reflectivity factor dataset are interpolated to Cartesian coordinates using a bilinear interpolation method to form a radar reflectivity factor gridded dataset, and the dataset is divided into a non-overlapping training set and a validation set.

[0060] The radar reflectivity factor prediction model was constructed using commonly used deep learning libraries (such as PyTorch, etc.), and the model was trained using the first radar reflectivity factor training set data samples. After the training was completed, the first radar reflectivity factor validation set was used for scoring to obtain the optimal radar reflectivity factor prediction model.

[0061] The radar reflectivity factor prediction model can be scored using the RMSE (Root Mean Square Error) indicator and / or the CSI (Characteristic Stability Index) forecast indicator.

[0062] Based on the above embodiments, Figure 3 As shown, in this embodiment, according to the wind speed of the to-be-forecasted area in the historical time period and the radar reflectivity factor in the future time period, the wind speed forecast model is used to predict the wind speed of the to-be-forecasted area in the future time period, including:

[0063] Using the optimal radar reflectivity factor prediction model to predict the radar reflectivity factor prediction result corresponding to the second radar reflectivity factor training set;

[0064] Using the wind speed data training set corresponding to the second radar reflectivity factor training set in time and space and the radar reflectivity factor forecast result, multiple deep learning models are trained to obtain corresponding multiple wind speed forecast models.

[0065] Using the second radar reflectivity factor verification set and the wind speed data verification set corresponding to the second radar reflectivity factor verification set in time and space to score each wind speed forecast model, and taking the wind speed forecast model with the highest score as the optimal wind speed forecast model;

[0066] The wind speed of the area to be forecasted in the future time period is predicted using the optimal wind speed forecast model according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period.

[0067] The second radar reflectivity factor training set may be the same as the first radar reflectivity factor training set, or may be extracted from a radar reflectivity factor data set, which is not limited in this embodiment.

[0068] The radar reflectivity factor prediction results corresponding to each sample in the second radar reflectivity factor training set are predicted and stored using the optimal radar reflectivity factor prediction model obtained through training. The saved radar reflectivity factor prediction results are used as new feature inputs to participate in the training of the wind speed prediction model.

[0069] The wind speed data of ground stations corresponding to the time and space of the radar reflectivity factor dataset are collected to establish a wind speed dataset. The wind speed dataset can be divided into a training set and a validation set in a ratio of 8:2.

[0070] Optionally, the site wind speed data in the wind speed data set are bilinearly interpolated to Cartesian coordinates to form a wind speed gridded data set, which corresponds to the radar reflectivity factor gridded data set in time and space, and then the wind speed forecast model is trained.

[0071] The deep learning model is trained by combining the saved radar reflectivity factor forecast results and the constructed wind speed data training set to establish a wind speed forecast model, and evaluate and select the optimal wind speed forecast model. The wind speed forecast model can be scored using the RMSE index and / or the CSI forecast index.

[0072] On the basis of the above embodiment, in this embodiment, according to the radar reflectivity factor of the area to be predicted in the historical time period, the radar reflectivity factor prediction model is used to predict the radar reflectivity factor of the area to be predicted in the future time period, including:

[0073] Using a bilinear interpolation method, the radar reflectivity factor of the area to be forecasted in the historical time period is interpolated to Cartesian coordinates to form radar reflectivity factor grid data;

[0074] According to the radar reflectivity factor gridded data of the area to be predicted in a historical time period, a radar reflectivity factor prediction model is used to predict the radar reflectivity factor gridded data of the area to be predicted in a future time period.

[0075] On the basis of the above embodiment, in this embodiment, according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period, the wind speed forecast model is used to predict the wind speed of the area to be forecasted in the future time period, including:

[0076] Using a bilinear interpolation method, the wind speed of the area to be forecasted within the historical time period is interpolated to Cartesian coordinates to form wind speed grid data;

[0077] The wind speed gridded data of the area to be forecasted in the future time period is predicted using a wind speed forecast model according to the gridded wind speed data of the area to be forecasted in the historical time period and the gridded radar reflectivity factor data in the future time period.

[0078] Based on the above embodiment, the loss function for training multiple deep learning models using the first radar reflectivity factor training set in this embodiment is:

[0079]

[0080] in, is the loss function value, n is the number of samples in the first radar reflectivity factor training set, and They are respectively the radar reflectivity factor prediction result and the actual value of the radar reflectivity factor of the sample output by the deep learning model.

[0081] In this embodiment, the loss function of the radar reflectivity factor prediction model training adopts a joint loss function that averages the L1 loss and the L2 loss, the optimizer can adopt the AdamW optimizer, the learning rate can be 5e-4, and the other parameters take the default values. After the training is completed, the validation set data is used for scoring to obtain the pre-trained optimal radar reflectivity factor prediction model.

[0082] The loss function of the wind speed forecast model training may be the same as the loss function of the radar reflectivity factor forecast model training, and the training hyperparameters may be the same.

[0083] On the basis of the above embodiment, the score of the radar reflectivity factor prediction model in this embodiment is a CSI score, the first radar reflectivity factor verification set is grid data, and each radar reflectivity factor prediction model is scored using the first radar reflectivity factor verification set by the following formula:

[0084] CSI = TP / (TP+FA+MS)

[0085] Among them, CSI is the score of the radar reflectivity factor prediction model, TP represents the number of grid points where both the observed and predicted radar reflectivity factors are within the preset range, MS represents the number of grid points where the observed radar reflectivity factor is within the preset range but the predicted radar reflectivity factor is not within the preset range, and FA represents the number of grid points where the observed radar reflectivity factor is not within the preset range but the predicted radar reflectivity factor is within the preset range.

[0086] On the basis of the above-mentioned embodiment, the score of the wind speed forecast model in this embodiment is the RMSE score.

[0087] This embodiment first uses a bilinear interpolation method to interpolate the radar reflectivity factor to Cartesian coordinates, and then constructs a deep learning model to train the data to obtain a pre-trained radar reflectivity factor prediction model; constructs a ground station data set corresponding to the time of strong wind, and interpolates it to the corresponding Cartesian coordinate grid points using a bilinear interpolation method; uses the pre-trained radar reflectivity prediction model to obtain an initial guess of the radar reflectivity for a future time period; combines the initial guess of the radar reflectivity for the future time period with the ground grid wind speed data, and performs deep learning model training for strong wind forecasting, and finally obtains a strong wind forecast result that integrates the radar pre-training results.

[0088] This embodiment makes fuller use of radar physical information through the secondary pre-training of the deep learning model, and enhances the adaptability and generalization ability of the model to complex meteorological conditions. In addition, this embodiment shows obvious advantages in both forecast accuracy and practicality, providing more reliable technical support for weather forecasting services and related fields, and has important practical application value and broad development prospects.

[0089] Table 1 compares the characteristics of this method with other wind speed forecasting methods:

[0090] Table 1

[0091] feature Traditional statistical methods Relying on limited variable selection, poor generalization, and difficulty in fully capturing the complex relationship between strong winds and multiple influencing factors, resulting in high uncertainty in the prediction results Numerical modeling methods Limited by the ability to describe the formation mechanism of atmospheric dynamics and the impact mechanism of the underlying surface, and highly dependent on the accuracy of initial field information Existing deep learning methods Radar data information is not fully utilized, and there are large errors in the initial estimates of storm location and intensity. There is much room for improvement in forecast accuracy and physical consistency. This method By using a deep learning pre-trained model, combined with radar reflectivity factors and ground wind speed data, the storm's location and intensity can be predicted more accurately, significantly improving the accuracy of the forecast and enhancing the model's adaptability and generalization ability to complex meteorological conditions.

[0092] Collect radar data and ground station wind speed data of a strong wind event in a certain area. The schematic diagram of the thunderstorm and strong wind process is as follows: Figure 4 As shown. Among them, Figure 4(a1) to (a5) are the input strong wind point data, (b1) to (b5) are the wind speed grid feature importance of the wind speed forecast model obtained by interpretability analysis, (c1) to (c5) are the initial guess values ​​of future radar forecasts, (d1) to (d5) are the initial guess value feature importance of the strong wind forecast model obtained by interpretability analysis, and (e1) to (e5) are the final future wind speed forecast grid results.

[0093] The trained wind speed forecast model is used to test the data to obtain the forecast result. The gale forecast result obtained by the present invention is compared and analyzed with the actual observation data to evaluate the forecast accuracy and perform interpretability analysis.

[0094] Through comparative analysis and explainability analysis, it can be seen that the gale forecasting method of the present invention can make full use of the physical information of the radar reflectivity factor, and provide an effective estimate of the future storm location and intensity through a secondary pre-trained deep learning model, thereby providing key information for the forecast of gale, and effectively improving the deep learning model's forecast accuracy for gale.

[0095] The strong wind forecasting system based on deep learning fusion radar provided by the present invention is described below. The strong wind forecasting system based on deep learning fusion radar described below and the strong wind forecasting method based on deep learning fusion radar described above can be referenced to each other.

[0096] like Figure 5 As shown, the system includes a first prediction module 501 and a second prediction module 502, wherein:

[0097] The first prediction module 501 is used to predict the radar reflectivity factor of the area to be predicted in the future time period using a radar reflectivity factor prediction model according to the radar reflectivity factor of the area to be predicted in the historical time period;

[0098] The second prediction module 502 is used to predict the wind speed of the area to be predicted in the future time period using a wind speed prediction model according to the wind speed of the area to be predicted in the historical time period and the radar reflectivity factor in the future time period;

[0099] The radar reflectivity factor prediction model and the wind speed prediction model are obtained by training a deep learning model.

[0100] This embodiment uses a pre-trained radar reflectivity factor prediction model to obtain an initial guess of the radar reflectivity factor for the future time period based on the radar reflectivity factor of the historical time period, and makes full use of the physical information of the radar reflectivity factor to provide important falling area and intensity information for the strong wind forecast; then the initial guess of the radar reflectivity factor for the future time period is combined with the wind speed data of the historical time period, and the pre-trained wind speed forecast model is used to predict the wind speed, so as to improve the accuracy of the strong wind forecast and meet the physical consistency, and provide effective information for the strong wind business warning and forecast.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A gale forecasting method based on deep learning fusion radar, characterized in that: include: According to the radar reflectivity factor of the area to be predicted in the historical time period, using the radar reflectivity factor prediction model to predict the radar reflectivity factor of the area to be predicted in the future time period; Predicting the wind speed of the area to be forecasted in the future time period using a wind speed forecast model according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period; The radar reflectivity factor prediction model and the wind speed prediction model are obtained by training a deep learning model.

2. The gale forecasting method based on deep learning fusion radar according to claim 1 is characterized in that: According to the radar reflectivity factor of the area to be predicted in the historical time period, the radar reflectivity factor prediction model is used to predict the radar reflectivity factor of the area to be predicted in the future time period, including: Using the first radar reflectivity factor training set to train multiple deep learning models, to obtain corresponding multiple radar reflectivity factor prediction models; Use the first radar reflectivity factor validation set to score each radar reflectivity factor prediction model, and take the radar reflectivity factor prediction model with the highest score as the optimal radar reflectivity factor prediction model; According to the radar reflectivity factor of the area to be predicted in a historical time period, the optimal radar reflectivity factor prediction model is used to predict the radar reflectivity factor of the area to be predicted in a future time period.

3. The gale forecasting method based on deep learning fusion radar according to claim 2 is characterized in that: Predicting the wind speed of the area to be forecasted in the future time period using a wind speed forecast model according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period, including: Using the optimal radar reflectivity factor prediction model to predict the radar reflectivity factor prediction result corresponding to the second radar reflectivity factor training set; Using the wind speed data training set corresponding to the second radar reflectivity factor training set in time and space and the radar reflectivity factor forecast result, multiple deep learning models are trained to obtain corresponding multiple wind speed forecast models; Using the second radar reflectivity factor verification set and the wind speed data verification set corresponding to the second radar reflectivity factor verification set in time and space to score each wind speed forecast model, and taking the wind speed forecast model with the highest score as the optimal wind speed forecast model; The wind speed of the area to be forecasted in the future time period is predicted using the optimal wind speed forecast model according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period.

4. The gale forecasting method based on deep learning fusion radar according to claim 2 is characterized in that: According to the radar reflectivity factor of the area to be predicted in the historical time period, the radar reflectivity factor prediction model is used to predict the radar reflectivity factor of the area to be predicted in the future time period, including: Using a bilinear interpolation method, the radar reflectivity factor of the area to be forecasted in the historical time period is interpolated to Cartesian coordinates to form radar reflectivity factor grid data; According to the radar reflectivity factor gridded data of the area to be predicted in a historical time period, a radar reflectivity factor prediction model is used to predict the radar reflectivity factor gridded data of the area to be predicted in a future time period.

5. The gale forecasting method based on deep learning fusion radar according to claim 4 is characterized in that: Predicting the wind speed of the area to be forecasted in the future time period using a wind speed forecast model according to the wind speed of the area to be forecasted in the historical time period and the radar reflectivity factor in the future time period, including: Using a bilinear interpolation method, the wind speed of the area to be forecasted within the historical time period is interpolated to Cartesian coordinates to form wind speed grid data; The wind speed gridded data of the area to be forecasted in the future time period is predicted using a wind speed forecast model according to the gridded wind speed data of the area to be forecasted in the historical time period and the gridded radar reflectivity factor data in the future time period.

6. The method for forecasting strong winds based on deep learning fusion radar according to claim 2, characterized in that: The loss function for training multiple deep learning models using the first radar reflectivity factor training set is: ； in, is the loss function value, n is the number of samples in the first radar reflectivity factor training set, and They are respectively the radar reflectivity factor prediction result and the actual value of the radar reflectivity factor of the sample output by the deep learning model.

7. The gale forecasting method based on deep learning fusion radar according to claim 2 is characterized in that: The score of the radar reflectivity factor prediction model is the CSI score, the first radar reflectivity factor validation set is grid data, and each radar reflectivity factor prediction model is scored using the first radar reflectivity factor validation set using the following formula: CSI = TP / (TP+FA+MS) Among them, CSI is the score of the radar reflectivity factor prediction model, TP represents the number of grid points where both the observed and predicted radar reflectivity factors are within the preset range, MS represents the number of grid points where the observed radar reflectivity factor is within the preset range but the predicted radar reflectivity factor is not within the preset range, and FA represents the number of grid points where the observed radar reflectivity factor is not within the preset range but the predicted radar reflectivity factor is within the preset range.

8. The method for forecasting strong winds based on deep learning fusion radar according to claim 3, characterized in that: The score of the wind speed forecast model is the RMSE score.

9. A strong wind forecasting system based on deep learning fusion radar, characterized in that: include: The first prediction module is used to predict the radar reflectivity factor of the area to be predicted in a future time period using a radar reflectivity factor prediction model according to the radar reflectivity factor of the area to be predicted in a historical time period; A second prediction module is used to predict the wind speed of the area to be predicted in the future time period using a wind speed prediction model according to the wind speed of the area to be predicted in the historical time period and the radar reflectivity factor in the future time period; The radar reflectivity factor prediction model and the wind speed prediction model are obtained by training a deep learning model.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the strong wind forecasting method based on deep learning fusion radar as described in any one of claims 1 to 8 is implemented.

Images