A weather prediction method based on deep learning

Abstract

The application discloses a weather prediction method and device based on deep learning, a medium and equipment. The weather prediction method based on deep learning is determined according to n cascaded / parallel weather prediction sub-models based on semantic segmentation. The input end of each weather prediction sub-model except the first-level weather prediction sub-model is communicatively connected to the output end of the previous n-1 weather prediction sub-models. Meteorological data is input into each weather prediction sub-model of each weather prediction model based on deep learning, and corresponding meteorological prediction data within a period of time after the current time is obtained. After weighting the meteorological prediction data, the final meteorological prediction data within a to-be-predicted space range is output. The application greatly improves the calculation efficiency of the weather prediction model by applying the deep learning method, thereby meeting the real-time prediction requirement and providing a fast-response weather prediction result for various application scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather prediction, and in particular to a weather prediction method and device based on deep learning, a medium and equipment. BACKGROUND

[0002] In the field of weather prediction, there are currently two main methods: numerical weather prediction and machine learning (especially deep learning) based weather prediction methods.

[0003] The numerical weather prediction method is based on physical principles for prediction, and has high scientificity and interpretability. By modeling the dynamics and thermodynamics of the atmosphere, the numerical weather prediction method can simulate the evolution of weather systems and obtain prediction results by calculating complex physical equations. Due to the physical model, the numerical weather prediction method has good interpretability and scientificity. The prediction results of this method are more in line with the laws of atmospheric physics, and have high application value for professional research and meteorological analysis. However, the numerical weather prediction method relies on a large amount of computing resources, and the inference speed is relatively slow, usually requiring several hours or even several days of computing time. At the same time, with the accumulation of historical data, although more input data can be provided, the prediction accuracy will not be significantly improved. At present, numerical weather prediction methods have been applied to some national and regional weather prediction systems, such as the Global Forecast System of the United States and the Ensemble Prediction System of the European Medium-Range Weather Forecast Center. The existing weather prediction system performs poorly in terms of prediction effect, inference speed, and data utilization. SUMMARY

[0004] The main purpose of the present application is to provide a weather prediction method, device, medium and equipment based on deep learning, aiming to combine the advantages of data-driven to provide a more efficient and accurate solution for weather prediction.

[0005] To achieve the above-mentioned purpose, the present application provides a weather prediction method based on deep learning, comprising:

[0006] Obtaining meteorological data within a certain period of time before the current time in the space range to be predicted and a weather prediction model based on deep learning, the weather prediction model based on deep learning is determined according to n cascaded / parallel weather prediction sub-models based on semantic segmentation, wherein the input end of each weather prediction sub-model except the first level weather prediction sub-model is communicatively connected to the output end of each weather prediction sub-model of the previous n-1 levels; input the meteorological data into each weather prediction sub-model of each weather prediction model based on deep learning, respectively, to obtain each meteorological prediction data within a certain period of time after the current time; and output the final meteorological prediction data in the space range to be predicted after weighting each meteorological prediction data.

[0007] Optionally, the n cascaded weather prediction sub-models based on semantic segmentation include a plurality of 2D models and at least two of a plurality of 3D models, the 2D models include MedNeXt and PVT-CASCADE, and the 3D models include MISSFormer, ConvNeXt, ConResNet, and FourCastNet.

[0008] Optionally, the meteorological data includes a plurality of pressure layer variables and a plurality of ground variables, each of the pressure layer variables includes a plurality of pressure layers, and the ground variables are various ground meteorological indexes; an input shape of each of the 3D models includes a first channel number, a first depth, a height, and a width, the first channel number is equal to a plurality of sampling time points within a period of time before a current time point, and the first depth is a sum of a number of all pressure layers and a number of all ground meteorological indexes; an input shape of each of the 2D models includes a second channel number, a height, and a width, the second channel number is equal to a dimension after the first channel is combined with the depth; a shape of each of the 3D models and each of the 2D models is a third channel number, a second depth, a height, and a width, the third channel number is a plurality of prediction time points within a period of time after the current time point, and the second depth is equal to a number of ground variables; and the height and the width represent grid points of the to-be-predicted spatial range, respectively.

[0009] Optionally, the meteorological data is input into each of the weather prediction sub-models, including: performing sea-land type marking and / or soil type marking and / or elevation marking on the meteorological data to obtain first input data; obtaining a difference value of the meteorological data at a sampling time point within a period of time before a current time point, and splicing the difference value of the meteorological data with the meteorological data at the current time point to obtain second input data; adding Gaussian noise to the second input data to obtain third input data; and inputting the first input data, the second input data, or the third input data into each of the weather prediction sub-models.

[0010] Optionally, before the sea-land type marking and / or soil type marking and / or elevation marking on the meteorological data to obtain the first input data, the method further includes: obtaining a mean value and a standard deviation of the elevation data, obtaining standardized elevation data according to a ratio of the mean value and the standard deviation, and obtaining corresponding elevation marking data according to the meteorological data.

[0011] Optionally, the inputting the meteorological data into each weather prediction sub-model of the deep learning-based weather prediction model respectively, and outputting each meteorological prediction data corresponding to a period of time after the current time, comprises: determining a weather prediction model group according to the parallel weather prediction sub-models, and determining the deep learning-based weather prediction model based on a plurality of weather prediction model groups; and inputting the meteorological data into each weather prediction model group in the deep learning-based weather prediction model respectively, so as to obtain the meteorological prediction data of each sub-period of time in the period of time after the current time, wherein each sub-period of time is continuous in time.

[0012] Optionally, the method further comprises: obtaining final meteorological prediction data and real label data, and determining a loss function based on the final meteorological prediction data and the real label data; and training the deep learning-based weather prediction model by using a gradient descent method according to the loss function, to obtain a deep learning-based weather prediction model with updated parameters.

[0013] In addition, to achieve the above object, the present application also provides a deep learning-based weather prediction device, comprising: an acquisition module configured to acquire meteorological data in a period of time before the current time in a to-be-predicted spatial range and a deep learning-based weather prediction model, wherein the deep learning-based weather prediction model is determined according to n cascaded / parallel weather prediction sub-models based on semantic segmentation, and the input end of each weather prediction sub-model except the first weather prediction sub-model is communicatively connected to the output end of the corresponding n-1 weather prediction sub-model; a processing module configured to input the meteorological data into each weather prediction sub-model of each deep learning-based weather prediction model respectively, so as to obtain each meteorological prediction data in a period of time after the current time; and an output module configured to output final meteorological prediction data in the to-be-predicted spatial range after weighting each meteorological prediction data.

[0014] To achieve the above object, the present application also provides a computer readable storage medium comprising instructions, which, when executed on a computer, cause the computer to perform the deep learning-based weather prediction method provided in the above embodiments.

[0015] An electronic device, comprising: at least one processor, a memory and an input-output unit; wherein the memory is configured to store a computer program, and the processor is configured to invoke the computer program stored in the memory to execute the deep learning-based weather prediction method provided in the above embodiments.

[0016] The embodiment of the present application provides a weather prediction method, device, medium and equipment based on deep learning, which acquires meteorological data in a period of time before the current time in a to-be-predicted space range and a weather prediction model based on deep learning, and determines the weather prediction model based on deep learning according to n cascaded / parallel weather prediction sub-models based on semantic segmentation, wherein the input end of each weather prediction sub-model except the first weather prediction sub-model is communicatively connected with the output end of each weather prediction sub-model of the previous n-1 levels; the meteorological data is respectively input into each weather prediction sub-model of each weather prediction model based on deep learning, and each meteorological prediction data in a period of time after the current time is correspondingly obtained; and the final meteorological prediction data in the to-be-predicted space range is output after weighting each meteorological prediction data, so that the calculation efficiency of the weather prediction model is greatly improved by the application of the deep learning method, the prediction operation can be completed within milliseconds to seconds, thereby meeting the real-time prediction requirement and providing a fast-response weather prediction result for various application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A flowchart is provided for the embodiment of the weather prediction method based on deep learning of the present application.

[0018] Figure 2 A principle diagram is provided for the embodiment of the weather prediction method based on deep learning of the present application.

[0019] Figure 3 A function module schematic diagram is provided for the embodiment of the weather prediction device based on deep learning of the present application.

[0020] Figure 4 A function module schematic diagram is provided for the embodiment of the weather prediction device based on deep learning of the present application.

[0021] The implementation, functional characteristics and advantages of the present application will be further described with reference to the accompanying drawings. DETAILED DESCRIPTION

[0022] It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0023] As the existing technology based on deep learning weather forecasting method gradually becomes an important means. This method learns and pattern recognizes historical weather data through deep neural network and other models to quickly generate prediction results. The deep learning based weather prediction method has excellent prediction performance and great potential. The computing resources required by this method are relatively small, and the inference speed is fast, which can reach millisecond level or even second level computing speed. In addition, the deep learning method can effectively utilize the historical input data to improve the accuracy of the prediction. But the current deep learning method has certain deficiency in explainability, and the model output result is often difficult to fully understand, which limits the wide application of this method in the field of meteorology. At present, some deep learning based weather forecasting systems have been put into application, such as Pangu model, Fuxi model, and GraphCast system. These systems make full use of the advantages of deep learning in data processing and pattern recognition, and provide an efficient solution for short-term and medium-term weather forecasting. Although the numerical weather prediction method has high scientificity and explainability, its computing resource demand is huge, the prediction speed is slow, and the effect of historical data utilization is not significant.

[0024] The present application provides a solution, which greatly improves the computing efficiency of the weather prediction model through the application of deep learning method, so that it can complete the prediction operation within milliseconds to seconds, thereby meeting the real-time prediction demand and providing fast response weather prediction results for various application scenarios. The present application uses deep learning model to learn and pattern extract a large amount of historical meteorological data, thereby improving the prediction accuracy, so that the model can capture the complex change rule of the weather system and provide technical support for higher accuracy weather forecasting. The present application makes full use of the learning and mining ability of deep learning model on historical data except the current time, which can find the complex pattern and potential relationship of weather system in large-scale data, so that the model shows higher prediction accuracy under different meteorological conditions. Compared with the traditional numerical weather prediction method, the deep learning method proposed in the present application can realize efficient training and prediction under the condition of relatively less computing resources, which is suitable for application in resource limited environment and can reduce the operation cost.

[0025] The technical solutions proposed in the present application will be described in detail below with reference to the accompanying drawings.

[0026] Referring to Figure 1 The deep learning based weather prediction method provided by the first embodiment of the present application can include:

[0027] S10, acquire meteorological data in a period of time before the current time in a space range to be predicted and a weather prediction model based on deep learning, the weather prediction model based on deep learning is determined according to n cascaded / parallel weather prediction sub-models based on semantic segmentation, wherein the input end of each weather prediction sub-model except the first level weather prediction sub-model is communicatively connected to the output end of the respective n-1 level weather prediction sub-model;

[0028] Reference Figure 2 In particular, the present application adopts a phased prediction method, which refers to using the input and prediction result of the current stage as the input of the next stage, so that the prediction accuracy can be significantly improved. Model integration improves the accuracy and robustness of overall prediction by combining the prediction results of multiple weather prediction sub-models based on semantic segmentation, thereby reducing the error and bias of a single model. Reference Figure 3 The present application adopts a weighted summation integration method, that is, the weight of each weather prediction sub-model based on semantic segmentation is set according to the prediction effect, and finally the prediction results of all models are multiplied by the weight to obtain the fusion result. The sum of the weights of different weather prediction sub-models based on semantic segmentation is 1.

[0029] In the selection process of the sub-model, 6 3D and 2D models in the field of semantic segmentation are mainly selected: MISSFormer, ConvNeXt, ConResNet, MedNeXt, FourCastNet and PVT-CASCADE. Among them, the 2D models include MedNeXt and PVT-CASCADE, and the 3D models include MISSFormer, ConvNeXt, ConResNet and FourCastNet.

[0030] In the data definition process, the meteorological data includes a plurality of pressure layer variables and a plurality of ground variables, each pressure layer variable includes a plurality of pressure layers, and the ground variable is a plurality of ground meteorological indexes; the input shape of each 3D model includes: a first channel number, a first depth, a height and a width, the first channel number is equal to a plurality of sampling times in a period of time before the current time, and the first depth is the sum of the number of all pressure layers and the number of all ground meteorological indexes; the input shape of each 2D model includes: a second channel number, a height and a width, and the second channel number is equal to the dimension after the first channel and the depth are combined; the shape of each 3D model and each 2D model is: a third channel number, a second depth, a height and a width, the third channel number is a plurality of prediction times in a period of time after the current time, and the second depth is equal to the number of ground variables; wherein the height and the width respectively represent the grid points of the space range to be predicted.

[0031] Exemplarily, the data set used in this application is mainly derived from the ERA5 reanalysis data set, covering meteorological historical data and meteorological variables required for future predictions, specifically including 5 pressure layer variables (each variable has 13 pressure layers), 5 ground variables, a total of 70 variables. The 5 pressure layer variables are geopotential height, temperature, zonal wind, meridional wind, and relative humidity, and the 5 ground variables are common surface meteorological indicators, namely 2-meter temperature, 10-meter zonal wind, 10-meter meridional wind, mean sea level pressure, and 6-hour cumulative precipitation. The submodel predicts the values ​​of 5 ground variables for the next 20 moments by inputting historical data (70 variables) at 2 moments, with the time interval between adjacent 2 moments being 6 hours. The spatial range covered by the data is 10 to 50 degrees north latitude and 100 to 140 degrees east longitude, with a spatial resolution of 0.25 degrees, thus obtaining a grid point of 161×161. The shape of the input data is [2, 70, 161, 161] (2 historical moments, 70 input variables, 161 × 161 grid points), and the shape of the prediction results and true labels is [20, 5, 161, 161] (20 future moments, 5 ground variables, 161 × 161 grid points).

[0032] For 3D models, the input data is of shape [2, 70, 161, 161], representing the number of channels, depth, height, and width of the 3D model input, respectively. At output, the number of channels is set to 20 (representing the next 20 moments). Finally, a fully connected layer maps the depth to 5 (representing the five ground variables), resulting in an output shape of [20, 5, 161, 161]. For 2D models, the first two dimensions of the input data are first merged to obtain data of shape [140, 161, 161]. At this time, 140, 161, and 161 are used as the number of channels, height, and width of the 2D model input, respectively. At output, the number of channels is set to 100, and then expanded to 20×5, resulting in an output shape of [20, 5, 161, 161].

[0033] The application also includes an optimization process for input data. In the optimization process for input data, the weather prediction method based on deep learning provided by the application can further include: marking meteorological data by sea-land type and / or soil type and / or altitude to obtain first input data; obtaining a difference value of meteorological data at a sampling time within a period of time before the current time, and splicing the difference value of meteorological data with meteorological data at the current time to obtain second input data; adding Gaussian noise to the second input data to obtain third input data; and inputting the first input data, the second input data or the third input data into each weather prediction sub-model. Specifically, before the meteorological data is marked by sea-land type and / or soil type and / or altitude to obtain the first input data, the data optimization process further includes: obtaining a mean value reduction and a standard deviation of the altitude data, and obtaining standardized altitude data according to a ratio of the mean value reduction and the standard deviation, and obtaining corresponding altitude marking data according to the meteorological data. Because the meteorological changes in different regions are affected by factors such as the sea-land type, soil type and altitude of the region. Therefore, the application will splice additional supplementary data in the original input, specifically including sea-land type marking, soil type marking and altitude. The altitude data range varies greatly and needs to be standardized by dividing the mean value reduction by the standard deviation before use. In addition, in order to make the model more explicitly model the change trend of historical data, the difference value of the historical 2 time data is also spliced with the original input. Data enhancement can increase data diversity and improve model generalization performance. Here, a learnable random Gaussian noise is added to the input data. The specific implementation is to add the Gaussian noise as a learnable weight of the model, which is optimized together with the training process.

[0034] In the embodiments of the application, the weather prediction method based on deep learning provided by the application can further include:

[0035] obtaining final meteorological prediction data and real label data, and determining a loss function based on the final meteorological prediction data and the real label data;

[0036] training the weather prediction model based on deep learning by using a gradient descent method according to the loss function, to obtain the weather prediction model based on deep learning with updated parameters.

[0037] Specifically, in the training stage, the model receives input data, obtains a prediction result by a certain prediction method, calculates a loss between the prediction result and a real label, calculates a gradient of the loss with respect to the model parameters, and updates the model parameters by a gradient descent method, so that the prediction result of the model is closer to the real label. In order to capture the complex rules contained in the meteorological data, 6 different architecture models are trained in the training stage, each model modeling different pattern relationships. In the inference stage, the prediction results of the 6 models are fused to obtain a final prediction result, that is, the weather prediction accuracy is further increased by model integration.

[0038] S20, input the meteorological data in each weather prediction sub-model of each deep learning-based weather prediction model respectively, to correspond to obtain each meteorological prediction data in a period of time after the current time;

[0039] In an optional embodiment, inputting the meteorological data in each weather prediction sub-model of each deep learning-based weather prediction model respectively, to correspond to output each meteorological prediction data in a period of time after the current time, comprises: determining a weather prediction model group according to the parallel weather prediction sub-models, and determining the deep learning-based weather prediction model based on a plurality of weather prediction model groups; inputting the meteorological data in each weather prediction model group of the deep learning-based weather prediction model respectively, to correspond to obtain meteorological prediction data in each sub-time period in a period of time after the current time, wherein each sub-time period is continuous in time.

[0040] In another optional embodiment, the weather prediction method can also include: a classical autoregressive method, which is simple and can flexibly cope with different prediction lengths, but has relatively long reasoning time and error accumulation problems; directly predicting all future time points, which has no error accumulation problems, but it is difficult for a single model to accurately predict 20 future time points only relying on 2 historical time points.

[0041] In another optional embodiment, the weather prediction method can also include: a variable grouping + time grouping + phased prediction method, when the prediction length is longer, the correlation between the future and the input data is greatly weakened, and more adjacent time data will be relied on, and the direct prediction method cannot solve this problem.

[0042] Referring to Figure 4 The second embodiment of the present application provides a deep learning-based weather prediction device. The deep learning-based weather prediction device 100 can include an acquisition module 1001, a processing module 1002, and an output module 1003. The acquisition module 1001 is used to acquire meteorological data in a period of time before the current time in a to-be-predicted spatial range and a deep learning-based weather prediction model. The deep learning-based weather prediction model is determined according to n cascaded / parallel semantic segmentation-based weather prediction sub-models. Except for the input end of the first weather prediction sub-model, the input end of each weather prediction sub-model is communicatively connected to the output end of the previous n-1 weather prediction sub-models. The processing module 1002 is used to input the meteorological data in each weather prediction sub-model of each deep learning-based weather prediction model respectively, to correspond to obtain each meteorological prediction data in a period of time after the current time. The output module 1003 is used to output the final meteorological prediction data in the to-be-predicted spatial range after weighting each meteorological prediction data.

[0043] On the basis of the above-mentioned embodiments, the application further provides a computer readable storage medium, which comprises instructions, when the instructions are executed on a computer, causing the computer to execute the deep learning-based weather prediction method provided by the above-mentioned embodiments.

[0044] On the basis of the above-mentioned embodiments, an electronic device comprises at least one processor, a memory and an input-output unit; wherein the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the deep learning-based weather prediction method provided by the above-mentioned embodiments.

[0045] The above is only the preferred embodiments of the application, and does not limit the patent scope of the application, and any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the application.

Claims

1. A weather forecasting method based on deep learning, characterized in that: include: Obtaining meteorological data for a period of time before a current moment within a spatial range to be predicted and a deep learning-based weather prediction model, wherein the deep learning-based weather prediction model is determined based on n cascaded / parallel semantic segmentation-based weather prediction sub-models, wherein the input ends of each level of weather prediction sub-model except the first level of weather prediction sub-model are communicatively connected to the output ends of the respective weather prediction sub-models of the previous n-1 levels; Inputting the meteorological data into each weather prediction sub-model of each weather prediction model based on deep learning, respectively, to obtain meteorological forecast data for a period of time after the current moment; After weighting each of the weather forecast data, output the final weather forecast data within the spatial range to be predicted; In the sub-model selection process, the six 3D and 2D models based on the semantic segmentation field are: MISSFormer, ConvNeXt, ConResNet, MedNeXt, FourCastNet and PVT-CASCADE. Among them, the 2D models include MedNeXt and PVT-CASCADE, and the 3D models include MISSFormer, ConvNeXt, ConResNet and FourCastNet; The meteorological data includes a plurality of pressure layer variables and a plurality of ground variables, each of the pressure layer variables includes a plurality of pressure layers, and the ground variables include various surface meteorological indicators; The input shape of each of the 3D models includes: a first channel number, a first depth, a height, and a width, wherein the first channel number is equal to a plurality of sampling moments in a period of time before the current moment, and the first depth is the sum of the number of all the pressure layers and the number of all the surface meteorological indicators; The input shape of each of the 2D models includes: a second number of channels, height, and width, where the second number of channels is equal to the dimensions of the first channels combined with the first depth; The shapes of the 3D models and the 2D models are: a third number of channels, a second depth, a height, and a width, wherein the third number of channels is a plurality of predicted moments within a period of time after the current moment, and the second depth is equal to the number of ground variables; The height and the width respectively represent the grid points of the spatial range to be predicted. Before inputting the meteorological data into each of the weather forecast sub-models, the method further includes: Marking the meteorological data by sea and land type and / or soil type and / or altitude to obtain first input data; Obtaining a difference value of meteorological data at sampling moments within a period of time before the current moment, and concatenating the difference value of meteorological data with the meteorological data at the current moment to obtain second input data; adding Gaussian noise to the second input data to obtain third input data; The first input data, the second input data or the third input data is input into each of the weather prediction sub-models.

2. The weather forecasting method based on deep learning according to claim 1, characterized in that: Before performing sea and land type marking and / or soil type marking and / or altitude marking on the meteorological data to obtain the first input data, the method further includes: The mean value and standard deviation of the altitude data are obtained, and the standardized altitude data is obtained according to the ratio of the mean value to the standard deviation, and corresponding altitude mark data is obtained according to the meteorological data.

3. The weather forecasting method based on deep learning according to claim 1, wherein: The meteorological data is input into each weather prediction sub-model of the deep learning-based weather prediction model, and meteorological prediction data for a period of time after the current moment is outputted accordingly, including: Determining a weather prediction model group according to the parallel weather prediction sub-models, and determining the deep learning-based weather prediction model based on multiple weather prediction model groups; The meteorological data is respectively input into each weather prediction model group in the deep learning-based weather prediction model, and the meteorological forecast data for each sub-time period in a period of time after the current moment is correspondingly obtained, wherein each sub-time period is continuous in time.

4. The weather forecasting method based on deep learning according to claim 1, wherein: The method further comprises: Obtaining final weather forecast data and true label data, and determining a loss function based on the final weather forecast data and the true label data; According to the loss function, the gradient descent method is used to train the weather prediction model based on deep learning to obtain the weather prediction model based on deep learning with updated parameters.

5. A weather forecasting device based on deep learning, applying the weather forecasting method based on deep learning according to any one of claims 1 to 4, characterized in that: include: An acquisition module, configured to acquire meteorological data for a period of time prior to a current moment within a spatial range to be predicted and a deep learning-based weather prediction model, wherein the deep learning-based weather prediction model is determined based on n cascaded / parallel semantic segmentation-based weather prediction sub-models, wherein the input end of each level of weather prediction sub-model other than the first level of weather prediction sub-model is communicatively connected to the output end of each of the previous n-1 levels of weather prediction sub-models; A processing module, configured to input the meteorological data into each weather prediction sub-model of each weather prediction model based on deep learning, and obtain meteorological prediction data for a period of time after the current moment; The output module is used to weight each of the weather forecast data and output the final weather forecast data within the spatial range to be predicted.

6. A computer-readable storage medium, characterized in that It includes instructions that, when executed on a computer, enable the computer to execute the weather forecasting method based on deep learning according to any one of claims 1 to 4.

7. An electronic device, characterized in that: The electronic device comprises: at least one processor, memory, and input-output unit; The memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the weather forecasting method based on deep learning according to any one of claims 1 to 4.

Images