Numerical forecasting wind field data error correction method and system based on neural network, medium and equipment

Abstract

The invention relates to the field of numerical weather forecast and deep learning, and discloses a numerical forecast wind field data error correction method and system based on a neural network, a medium and equipment. A data set obtained after data preprocessing is divided into a training set, a verification set and a test set, and space-time matching and normalization are conducted on the data set; establishing a many-to-many variable mapping relationship in a horizontal space by using the training set, expanding a time dimension on the basis of the established multiple mapping relationship, and establishing a 3D U-Net-based deep learning model; performing training and parameter tuning on a 3D U-Net-based deep learning model by using the training set and the verification set, and constructing a deep learning correction model of multi-element multi-time forecast; and correcting the test set by using the trained deep learning correction model. According to the method, the problem of multi-forecast aging and multivariable collaborative correction can be solved.

Description

Technical Field

[0001] The present invention relates to the technical fields of numerical weather prediction and deep learning, and particularly to a method, system, medium and device for correcting errors in numerical prediction wind field data based on a neural network. Background Art

[0002] Although numerical weather prediction has made great progress in the past few decades, due to inaccurate initial fields, imperfect description of relevant physical processes by dynamic models, and insufficiently fine resolution, there are still certain errors in current numerical weather prediction results, which increase with the increase of the prediction time limit. In recent years, deep learning has been applied to many fields such as image processing, target recognition, and diagnostic prediction, and has also provided new ideas for improving numerical weather prediction in the marine meteorology discipline. Current existing research mainly uses simple machine learning models such as random forests, support vector machines, etc. or convolutional neural networks to correct the surface wind field of a single prediction time limit in numerical prediction. However, these methods still have limitations, such as not considering the continuity in the prediction time dimension, and there is less research on the correction of the offshore wind field. Summary of the Invention

[0003] Aiming at the above problems, the purpose of the present invention is to provide a method, system, medium and device for correcting errors in numerical prediction wind field data based on a neural network, which can solve the problem of collaborative correction of multiple prediction time limits and multiple variables.

[0004] To achieve the above purpose, the present invention adopts the following technical solutions: A method for correcting errors in numerical prediction wind field data based on a neural network, which performs collaborative correction of multiple variables for multiple prediction time limits on the wind field data predicted by a numerical weather prediction model, including: collecting internationally publicly released prediction data and reanalysis data, dividing the obtained data set after data preprocessing into a training set, a validation set and a test set, and performing spatio-temporal matching and normalization on the data set; establishing a many-to-many variable mapping relationship in the horizontal space using the training set, expanding the time dimension on the basis of the established multiple mapping relationships, and building a 3D U-Net deep learning model; training and optimizing the parameters of the 3D U-Net deep learning model using the training set and the validation set, and constructing a deep learning correction model for multi-element and multi-time prediction; correcting the test set using the trained deep learning correction model.

[0005] Further, the internationally publicly released prediction data includes historical prediction data of the wind field released by various prediction systems, as the input field of the model;

[0006] The reanalysis data includes various wind field reanalysis data sets, as the target value of the model;

[0007] The variables of the forecast data and reanalysis data are: the longitude component and latitude component of the wind field, and the sea surface pressure field physically related to the wind field.

[0008] Furthermore, the data preprocessing includes: if the collected forecast data is missing at a certain forecast time, then all future forecast samples corresponding to the starting time of this data are removed.

[0009] Furthermore, the spatio-temporal matching and normalization of the dataset include:

[0010] The spatio-temporal matching arranges the dataset in chronological order to make the spatio-temporal fields of the target values and numerical forecast data match one by one;

[0011] The normalization uses the min-max normalization method, and after min-max normalization, the data is distributed between 0 and 1.

[0012] Furthermore, the 3D U-Net deep learning model includes an encoder module, a decoder module, and skip connections;

[0013] The encoder module includes an input layer, three-dimensional convolution, batch normalization, activation function, and max-pooling downsampling operation, which are used to downsample the input spatial information, learn deep features of multi-variables and multi-spatio-temporal, and generate encoder feature maps;

[0014] The decoder module includes three-dimensional convolution, transposed convolution upsampling, batch normalization, and activation function, which are used for upsampling in space, restoring features, and generating decoder feature maps;

[0015] The skip connections are used for feature fusion, which is the stacking of decoder feature maps and encoder feature maps of the same spatial size in the channel dimension.

[0016] Furthermore, using the training set and validation set to train and optimize the parameters of the 3D U-Net deep learning model includes:

[0017] The mean absolute error MAE is used as the loss function to measure the gap between the model output and the true value, and the learning rate and dropout rate are adjusted according to the performance of the validation set to prevent overfitting, and the training results are monitored until the output result of the validation set reaches the optimal effect.

[0018] Furthermore, using the trained deep learning correction model to correct the test set includes:

[0019] The forecast results of the optimized model for the test set are corrected, and the root mean square error RMSE and mean bias MB are used to quantify the performance of the model.

[0020] A numerical prediction wind field data error correction system based on neural network performs multi-variable collaborative correction for wind field data predicted by numerical weather prediction models at multiple prediction timescales, including: a data processing module that collects publicly released international forecast data and reanalysis data, divides the dataset obtained after data preprocessing into a training set, a validation set, and a test set, and performs spatio-temporal matching and normalization on the dataset; a model building module that uses the training set to establish a many-to-many variable mapping relationship in the horizontal space, extends the time dimension based on the established multiple mapping relationships, and builds a 3D U-Net deep learning model; a training and tuning module that uses the training set and the validation set to train and tune the parameters of the 3D U-Net deep learning model, and constructs a deep learning correction model for multi-element and multi-timescale prediction; a correction module that uses the trained deep learning correction model to correct the test set.

[0021] A computer-readable storage medium storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to execute any of the methods described above.

[0022] A computing device, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any of the methods described above.

[0023] Due to the above technical solutions adopted by the present invention, it has the following advantages:

[0024] Through multi-timescale spatio-field joint modeling, the present invention considers the continuity in the prediction time dimension, avoids the fragmentation and inefficient training caused by separate modeling of different prediction timescales, and particularly considers the physical correlation and joint output between different elements. The model learns the effective features between them at the same time, avoiding the inefficient learning of single-variable output. The performance results of the model show that the present invention can significantly improve the wind field results of numerical weather prediction. The present invention can be widely applied to the error correction of numerical prediction results in the field of marine meteorology, guiding production activities such as offshore fisheries and oil and gas drilling.

[0025] Other features and advantages of the present invention will be described in the subsequent specification, and, in part, will become apparent from the specification or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be achieved and obtained by the structures specifically pointed out in the written specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is the overall flowchart of the numerical prediction wind field data error correction method based on neural network in the embodiment of the present invention;

[0027] Figure 2 It is a schematic diagram of the 3D U-Net model structure in an embodiment of the present invention;

[0028] Figure 3 It is a detailed flowchart of a method for correcting errors in numerical prediction wind field data based on a neural network in an embodiment of the present invention. Specific embodiments

[0029] To solve the limitations of existing methods, such as the lack of continuity in the prediction time dimension and the relatively few studies on offshore wind field correction, the present invention proposes a method, system, medium, and device for correcting errors in numerical prediction wind field data based on a neural network, including: collecting the prediction data and reanalysis data of the wind field and surface pressure field publicly released internationally, and constructing a numerical prediction error correction data set; the numerical prediction error correction data set includes a training set, a validation set, and a test set; based on the 3D U-Net deep learning model, combining the physical correlation between the wind field and the surface pressure field, using the training set to establish a horizontal spatial two-dimensional many-to-many variable mapping relationship; expanding the time dimension on the basis of the established multiple mapping relationships, building the model and designing the parameters; using the training set and the validation set to carry out training and optimization, and constructing a multi-variable and multi-time prediction correction model; using the trained deep learning correction model to carry out correction on the test set. The present invention introduces the pressure field to improve the correction effect of wind field prediction errors, taking into account both the physical relationship between the wind field and the pressure field and the continuity in the prediction time dimension, avoiding the fragmentation and inefficient training of separate modeling for different prediction time periods, and supporting the whole process of preprocessing, modeling correction, and evaluation analysis of environmental element data in numerical weather prediction. The correction results can be used to improve the quality of weather forecasts and guide production activities such as offshore fisheries and oil and gas drilling.

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings of the embodiments of 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 based on the described embodiments of the present invention fall within the scope of protection of the present invention.

[0031] It should be noted that the terms used here are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used here, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0032] In one embodiment of the present invention, a method for correcting the error of numerical prediction wind field data based on a neural network is provided. In this embodiment, the present invention is used to perform multi-variable collaborative correction on the wind field data predicted by the numerical weather prediction model for multiple prediction time scales, such as Figure 1 as shown, the method includes the following steps:

[0033] 1) Collect the forecast data and reanalysis data publicly released internationally, obtain the numerical prediction error correction data set after data preprocessing, divide the data set into a training set, a validation set, and a test set, and perform spatio-temporal matching and normalization on the data set;

[0034] 2) Use the training set to establish a many-to-many variable mapping relationship in the horizontal space, expand the time dimension on the basis of the established multiple mapping relationships, and build a 3D U-Net deep learning model;

[0035] 3) Use the training set and the validation set to train and optimize the parameters of the 3D U-Net deep learning model, and construct a deep learning correction model for multi-element and multi-time-scale prediction;

[0036] 4) Use the trained deep learning correction model to correct the test set.

[0037] When the present invention is used, it can effectively integrate multi-variable spatio-temporal features, solve the limitations that the meridional and zonal components of the vector wind field cannot be jointly corrected and the existing models can only correct the prediction results of a single time scale at a time, and improve the accuracy of the correction of the offshore wind field prediction. The correction results of the present invention are of great significance for improving the quality of weather forecasting and guiding production activities such as offshore fishery and oil and gas drilling.

[0038] In the above step 1), the forecast data publicly released internationally includes the historical forecast data of the wind field released by various international forecast systems, which is used as the input field of the model;

[0039] The reanalysis data includes various wind field reanalysis data sets, which are used as the target values of the model;

[0040] The variables of the forecast data and the reanalysis data are: the longitude component and the latitude component of the wind field, and the sea surface pressure field physically related to the wind field.

[0041] In the above step 1), the data preprocessing includes: if the forecast data collected is missing at a certain forecast time scale, then all the future forecast samples corresponding to the starting time of the data are removed.

[0042] In the above step 1), the spatio-temporal matching and normalization of the data set are respectively:

[0043] The spatio-temporal matching is to arrange the data set in time order so that the spatio-temporal fields of the target values and the numerical prediction data are matched one by one;

[0044] Normalization is performed using the min-max normalization method. After min-max normalization, the data is distributed between 0 and 1.

[0045] In step 2) above, as Figure 2 shown, the 3D U-Net deep learning model includes an encoder module, a decoder module, and skip connections;

[0046] The encoder module includes an input layer, 3D convolution, batch normalization, activation function, and max pooling downsampling operations, which are used to downsample the input spatial information, learn deep features of multi-variables and multi-spatio-temporal, and generate encoder feature maps;

[0047] The decoder module includes 3D convolution, transposed convolution upsampling, batch normalization, and activation function, which are used for upsampling in space, restoring features, and generating decoder feature maps;

[0048] The skip connections are used for feature fusion, which is to stack the decoder feature maps and encoder feature maps of the same spatial size in the channel dimension.

[0049] In step 3) above, the 3D U-Net deep learning model is trained and its parameters are tuned using the training set and the validation set, including the following steps:

[0050] 3.1) The mean absolute error MAE is used as the loss function to measure the gap between the model output and the true value;

[0051] 3.2) Adjust the learning rate and dropout rate according to the performance of the validation set to prevent overfitting;

[0052] 3.3) Monitor the training results until the output results of the validation set reach the optimal effect.

[0053] In step 4) above, the trained deep learning correction model is used to correct the test set, including the following steps:

[0054] 4.1) Correct the prediction results of the tuned model for the test set;

[0055] 4.2) The root mean square error RMSE and the mean bias MB are used to quantify the performance of the model.

[0056] Example, an error correction method for 10m wind field data in offshore numerical prediction based on deep learning, jointly models and corrects the wind field prediction results of multiple time steps. As Figure 3 shown, the numerical prediction wind field data error correction method based on neural network provided in this example includes the following steps:

[0057] S1. Collect the historical forecast and reanalysis datasets of the 10m wind field and sea surface pressure in the Chinese offshore area, divide them into training set, validation set and test set, and perform spatio-temporal matching and normalization on the datasets;

[0058] S2. Combine the physical correlation between the wind field and the pressure field, and use the training set to establish a many-to-many variable mapping relationship in the horizontal space;

[0059] S3. Expand the time dimension based on the established multiple mapping relationships, build and design the parameters of a 3D U-Net deep learning model, use the training set and the validation set for training and optimization, and construct a multi-time-scale forecast correction model applicable to the 10m wind field of offshore numerical weather prediction;

[0060] S4. Use the trained deep learning correction model to correct the test set. Taking the 10m wind field in the offshore area as an example, since the 10m wind field at sea is a basic and important meteorological element, which not only plays an important synoptic role in forecasting basic elements such as air temperature, but also directly affects the production and life of people at sea. Therefore, improving the forecasting level of the sea wind field has important practical significance.

[0061] Specifically, the method for constructing and processing the dataset is as follows: In this embodiment, it is necessary to construct a sufficiently rich sample for the deep learning model. The sample construction is sourced from the historical forecast data of the Global Forecast System (GFS) developed by the National Oceanic and Atmospheric Administration of the United States and the data such as the meridional component U10m, zonal component V10m and mean sea level pressure MSLP of the 10m wind field in the fifth-generation atmospheric reanalysis dataset (ERA5) of the European Centre for Medium-Range Weather Forecasts. The time span of the dataset is from 2015 to 2023, with an interval of every 6 hours, once at 0:00, 6:00, 12:00 and 18:00 every day; the spatial range covers the Chinese offshore area (3°N - 42.75°N, 100°E - 139.75°E), and the spatial resolution is 0.25°×0.25°.

[0062] To avoid the influence of missing values in the data on the training results, this embodiment preprocesses the data. Specifically, when the forecast data collected is missing at a certain forecast time scale, all future forecast samples corresponding to the starting time of the data are removed.

[0063] The dataset of this embodiment is divided into a training set, a validation set, and a test set. The time span of the training set is from January 15, 2015 to December 31, 2021, and each variable has 10,132 data samples in the training set; the time span of the validation set is from January 1, 2022 to December 31, 2022, and each variable has 1,459 data samples in the training set; the time span of the test set is from January 1, 2023 to December 31, 2023, and each variable has 1,454 data samples in the training set. The dimension of the input array is (n, 3, 6, 160, 160), where n represents the number of samples; 3 represents U10m, V10m, and MSLP in the GFS forecast; 6 represents 6 time steps from the initial forecast time to the next 5 days, with a 24-hour time interval; the two 160s respectively represent the number of grid points in the meridional and zonal directions of the study area. The dimension of the true value array is (n, 2, 6, 160, 160), and the dimension of the corrected array, which is the model output, is also (n, 2, 6, 160, 160).

[0064] For each variable of GFS and ERA5, calculate the maximum and minimum values of their respective offshore elements and normalize them, and set the land part to 0 uniformly. The minimum-maximum normalization method is used for normalization, and the specific formula is:

[0065] x norm =(x - x min ) / (x max - x min ),

[0066] where x min and x max represent the minimum and maximum values of the numerical weather prediction error correction dataset respectively, x is the original data of the numerical weather prediction error correction dataset, and after minimum-maximum normalization, the data is distributed between 0 and 1.

[0067] In this embodiment, deep learning modeling and training are specifically as follows: The model uses 3D U-Net, adding a time dimension to the two-dimensional U-Net model, which is convenient for taking different times as a whole input for learning, avoiding the fragmentation of considering different time steps separately, and improving the calculation efficiency at the same time. The 3D U-Net model based on it includes an encoder module, a decoder module, and skip connections. The construction of the 3D U-Net deep learning model, the design of specific model parameters, and training specifically include the following steps:

[0068] S21. Encoder module: Downsample the input spatial information to learn the deep features of U10m, V10m, and MSLP in the offshore area in the next 5 days, and generate encoder feature maps;

[0069] S22. Decoder module: Implement upsampling in space, restore the features of U10m, V10m, and MSLP in the offshore area for the next 5 days, and generate decoder feature maps.

[0070] S23. Skip connection: Fuse the decoder feature maps and encoder feature maps of the same spatial size in the channel dimension.

[0071] S24. During the training process, use the mean absolute error (MAE) of the loss function to monitor the output results of the model.

[0072] S25. Select the Adam optimization algorithm, set the initial learning rate to 0.001, and decay it in the ExponentialLR manner for each additional round of training for gradient update of the loss function with respect to the model weights. Set the number of training samples in each mini-batch to 4 and the number of training rounds to 30.

[0073] In this embodiment, the verification of the correction result is as follows: During testing, the model output, that is, the dimension of the corrected array, is the same as (1459, 3, 6, 160, 160). The maximum and minimum values of each variable's offshore elements in GFS and ERA5 in the training set are used for inverse normalization to obtain the final correction result, and the error evaluation is based on this result.

[0074] For the numerical weather prediction wind field data error correction method based on neural network provided in this embodiment, its correction performance is evaluated on the validation set, and the root mean square error (RMSE) and mean bias (MB) are used to quantify the performance of the model. In this embodiment, the improvement amplitude of the root mean square error of the predicted wind field within the next 5 days is more than 15%.

[0075] In an embodiment of the present invention, a numerical weather prediction wind field data error correction system based on neural network is provided to perform multi-variable collaborative correction of wind field data predicted by the numerical weather prediction model for multiple prediction time scales, including:

[0076] Data processing module: Collect publicly released forecast data and reanalysis data internationally, divide the dataset obtained after data preprocessing into training set, validation set, and test set, and perform spatio-temporal matching and normalization on the dataset.

[0077] Model building module: Use the training set to establish a many-to-many variable mapping relationship in the horizontal space, extend the time dimension based on the established multiple mapping relationships, and build a 3D U-Net deep learning model.

[0078] Training and tuning module: Use the training set and validation set to train and tune the parameters of the 3D U-Net deep learning model, and construct a deep learning correction model for multi-element and multi-time-scale prediction.

[0079] Correction module, which corrects the test set using the trained deep learning correction model.

[0080] In the above embodiment, the internationally publicly released forecast data includes the historical wind field forecast data released by various forecast systems as the model input field.

[0081] The reanalysis data includes various wind field reanalysis data sets as the target value of the model.

[0082] The variables of the forecast data and the reanalysis data are: the longitude component and the latitude component of the wind field, and the sea surface pressure field physically associated with the wind field.

[0083] In the above embodiment, the data preprocessing includes: if the collected forecast data is missing at a certain forecast time, then all future forecast samples corresponding to the starting time of the data are removed.

[0084] In the above embodiment, the spatio-temporal matching and normalization of the data set include:

[0085] The spatio-temporal matching arranges the data set in chronological order to make the spatio-temporal fields of the target value and the numerical forecast data match one by one.

[0086] The normalization uses the min-max normalization method, and after min-max normalization, the data is distributed between 0 and 1.

[0087] In the above embodiment, the 3D U-Net deep learning model includes an encoder module, a decoder module, and skip connections.

[0088] The encoder module includes an input layer, three-dimensional convolution, batch normalization, activation function, and max pooling downsampling operations, which are used to downsample the input spatial information, learn deep features of multiple variables and multiple spatio-temporal dimensions, and generate encoder feature maps.

[0089] The decoder module includes three-dimensional convolution, transposed convolution upsampling, batch normalization, and activation function, which are used for upsampling in space, restoring features, and generating decoder feature maps.

[0090] The skip connections are used for feature fusion, which is the stacking of decoder feature maps and encoder feature maps of the same spatial size in the channel dimension.

[0091] In the above embodiment, training and parameter tuning of the 3D U-Net deep learning model using the training set and the validation set include:

[0092] The mean absolute error MAE is used as the loss function to measure the gap between the model output and the true value, and the learning rate and dropout rate are adjusted according to the performance of the validation set to prevent overfitting, and the training results are monitored until the output result of the validation set reaches the optimal effect.

[0093] In the above embodiments, correcting the test set by using the trained deep learning correction model includes:

[0094] Correcting the prediction results of the tuned model for the test set, and using the root mean square error RMSE and the mean bias MB to quantify the performance of the model.

[0095] The system provided in this embodiment is used to execute the above method embodiments. For the specific process and detailed content, please refer to the above embodiments and will not be elaborated here.

[0096] In an embodiment of the present invention, a computing device is provided. The computing device may be a terminal, and it may include: a processor, a communications interface, a memory, a display screen, and an input device. Among them, the processor, the communications interface, and the memory complete mutual communication through a communication bus. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. When the computer program is executed by the processor, the methods in the above embodiments are implemented; the internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communications interface is used to communicate with an external terminal in a wired or wireless manner. The wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen may be a liquid crystal display screen or an electronic ink display screen. The input device may be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computing device, or an external keyboard, touchpad, or mouse, etc. The processor can call the logical instructions in the memory.

[0097] In addition, when the logical instructions in the above memory 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 this 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: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical disks, and other various media that can store program codes.

[0098] In one embodiment of the present invention, there is provided a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions that, when executed by a computer, enable the computer to execute the methods provided in the above-described method embodiments.

[0099] In one embodiment of the present invention, there is provided a non-transitory computer-readable storage medium that stores server instructions, and these computer instructions cause the computer to execute the methods provided in the above embodiments.

[0100] For a computer-readable storage medium provided in the above embodiments, its implementation principle and technical effects are similar to those of the above method embodiments, and will not be elaborated herein.

[0101] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one or more flows and / or one or more blocks. Figure 1 one or more flows and / or Figure 1 one or more blocks

[0102] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in one or more flows and / or one or more blocks. Figure 1 one or more flows and / or Figure 1 one or more blocks

[0103] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more flows and / or one or more blocks. Figure 1 one or more flows and / or Figure 1 one or more blocks

[0104] 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 cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for correcting errors in numerical wind forecast data based on a neural network, characterized in that: The wind data predicted by the numerical weather prediction model are corrected with multi-variable coordination for multiple forecast periods, including: Collect internationally published forecast data and reanalysis data, divide the data sets obtained after data preprocessing into training sets, validation sets and test sets, and perform spatiotemporal matching and normalization on the data sets; Use the training set to establish a many-to-many variable mapping relationship in the horizontal space, expand the time dimension based on the established multiple mapping relationship, and build a deep learning model based on 3D U-Net; The training set and validation set are used to train and optimize the parameters of the 3D U-Net deep learning model, and a deep learning correction model for multi-factor and multi-time prediction is constructed; The trained deep learning correction model is used to correct the test set.

2. The method for correcting the error of numerical wind forecast data based on a neural network as claimed in claim 1, characterized in that: The forecast data released internationally include historical wind forecast data released by various forecast systems, which are used as model input fields; The reanalysis data include various wind field reanalysis data sets, which serve as the target values ​​of the model; The variables of forecast data and reanalysis data are: the longitude and latitude components of the wind field, and the sea surface pressure field that is physically related to the wind field.

3. The method for correcting the error of numerical wind forecast data based on a neural network as claimed in claim 1, characterized in that: Data preprocessing includes: if the collected forecast data is missing in a certain forecast time period, all future forecast samples corresponding to the starting time of the data will be removed.

4. The method for correcting the error of numerical wind forecast data based on a neural network as claimed in claim 1, characterized in that: Temporal and spatial matching and normalization of datasets include: Time-space matching is to arrange the data set in time order so that the time-space fields of the target value and the numerical forecast data match one by one; Normalization uses the minimum and maximum normalization method. After minimum and maximum normalization, the data is distributed between 0 and 1.

5. The method for correcting the error of numerical wind forecast data based on neural network according to claim 1, characterized in that: Based on the 3D U-Net deep learning model, the model includes an encoder module, a decoder module, and a skip-layer connection; The encoder module includes an input layer, 3D convolution, batch normalization, activation function, and max pooling downsampling operations, which are used to downsample the input spatial information, learn multivariate and multi-temporal deep features, and generate encoder feature maps; The decoder module includes three-dimensional convolution, transposed convolution upsampling, batch normalization and activation function, which are used for spatial upsampling, feature restoration and generation of decoder feature maps; The skip connection is used for feature fusion, which is to stack the decoder feature map and encoder feature map of the same spatial size in the channel dimension.

6. The method for correcting the error of numerical wind forecast data based on a neural network as claimed in claim 1, characterized in that: Use the training set and validation set to train and optimize the parameters of the 3D U-Net deep learning model, including: The mean absolute error (MAE) is used as the loss function to measure the gap between the model output and the true value. The learning rate and the dropout rate are adjusted according to the performance of the validation set to prevent overfitting. The training results are monitored until the output of the validation set reaches the optimal effect.

7. The method for correcting the error of numerical wind forecast data based on a neural network as claimed in claim 1, characterized in that: Use the trained deep learning correction model to correct the test set, including: The prediction results of the tuned model for the test set are corrected, and the root mean square error RMSE and mean deviation MB are used to quantify the performance of the model.

8. A numerical forecast wind field data error correction system based on neural network, characterized in that: The wind data predicted by the numerical weather prediction model are corrected with multi-variable coordination for multiple forecast periods, including: The data processing module collects internationally published forecast data and reanalysis data, divides the data sets obtained after data preprocessing into training sets, validation sets, and test sets, and performs spatiotemporal matching and normalization on the data sets; The model building module uses the training set to establish a many-to-many variable mapping relationship in the horizontal space, expands the time dimension based on the established multiple mapping relationships, and builds a deep learning model based on 3D U-Net; The training and tuning module uses the training set and validation set to train and tune the parameters of the 3D U-Net deep learning model, and build a deep learning correction model for multi-factor and multi-time forecasting; The correction module uses the trained deep learning correction model to correct the test set.

9. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any one of the methods of claims 1 to 7.

10. A computing device, characterized in that include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for executing any one of the methods described in claims 1 to 7.

Images