Shallow wind forecasting method, device and equipment and storage medium

Abstract

The invention discloses a shallow wind forecasting method, device and equipment and a storage medium, and relates to the technical field of shallow wind forecasting, and the method comprises the steps: carrying out the training of historical shallow wind observation data of a station, obtaining a first model, and carrying out the forecasting of the shallow wind in a first time period in the future through the first model; and based on the position information of the station, training by using a historical numerical forecasting product to obtain a second model, and forecasting shallow wind in a second time period in the future by using the second model. In this way, according to shallow wind forecast in different time periods, different data are adopted to train the first model and the second model respectively; and in combination with historical observation data of the station and a data forecasting product issued by a forecasting center, the accuracy of shallow wind forecasting can be improved in a targeted manner.

Description

Technical Field

[0001] The present invention relates to the technical field of shallow wind forecasting, and in particular to a shallow wind forecasting method, device, equipment and storage medium. Background Art

[0002] Currently, commonly used shallow wind forecasting methods fall into the following categories, based on their principles: synoptic forecasting methods, numerical forecast products, and methods for interpreting and applying numerical forecast products. Synoptic forecasting methods, including forecast indicator methods, similar situation methods, extrapolation methods, and kinematic methods, utilize historical and current data for analysis and forecasting. These methods are not objective or quantitative, and rely on the forecaster's experience. Classic numerical forecast products have a spatial resolution of 3 kilometers and a temporal resolution of 3 hours, with wind speeds limited to 10 and 100 meters. These spatial and temporal accuracies are far from sufficient for shallow wind forecasts at six levels (10, 30, 50, 70, 80, and 90 meters) at fixed points and times within the launch site. Furthermore, only a subset of meteorological stations participate in data assimilation within numerical forecast products. Consequently, due to corrections to actual weather conditions at the site, numerical forecasts often exhibit errors, making them inadequate for refined shallow wind forecasts. The interpretation and application of numerical forecast products is based on dynamic and statistical methods. Through dynamic interpretation of weather mechanisms and the development of statistical models, objective forecasting methods are developed for different weather phenomena and forecasting requirements. Common dynamic downscaling methods based on the WRF (Weather Research and Forecasting Model) have vertical limitations and cannot provide detailed processing of wind fields below 100 meters. Some methods use microscale CFD (Computational Fluid Dynamics) models to refine wind fields, but these are based on commonly used numerical forecasts and lack assimilation and targeted forecast training tailored to local conditions, resulting in low accuracy.

[0003] It can be seen that how to improve the accuracy of shallow wind forecast is a problem to be solved in this field. Summary of the Invention

[0004] In view of this, the present invention aims to provide a shallow wind forecasting method, apparatus, device, and storage medium. This method uses different data to train a first model and a second model for shallow wind forecasts in different time periods. This method, combined with historical observation data and data forecast products from current weather stations, can improve the accuracy of shallow wind forecasts. The specific solution is as follows:

[0005] In a first aspect, the present application provides a shallow wind forecasting method, comprising:

[0006] A first model is trained using historical shallow wind observation data of the station, and the first model is used to forecast shallow winds in a first future time period;

[0007] Based on the location information of the local station, a second model is trained using historical numerical forecast products, and the second model is used to forecast shallow winds in a second time period in the future.

[0008] Optionally, the first model is obtained by training the historical shallow wind observation data of the station, including:

[0009] Obtaining shallow wind data observed by the shallow wind measurement system of the station in a first preset time period to obtain historical shallow wind observation data;

[0010] Performing an importance assessment on the historical shallow wind observation data based on a random forest algorithm, and screening the historical shallow wind observation data to obtain first training data based on the assessment result;

[0011] The pre-built long short-term memory network is trained using the first training data to obtain a first model.

[0012] Optionally, performing importance evaluation on the historical shallow wind observation data based on a random forest algorithm, and screening the historical shallow wind observation data to obtain first training data based on the evaluation result, includes:

[0013] Eliminating data that meets a preset data error condition from the historical shallow wind observation data to obtain processed observation data after eliminating the erroneous data;

[0014] The importance of the processed observation data is evaluated by using the permutation importance of the random forest out-of-bag data to obtain the corresponding evaluation results;

[0015] Based on the relationship between the evaluation result and a preset importance threshold, first training data is obtained by screening from the processed observation data.

[0016] Optionally, the using the first training data to train a pre-built long short-term memory network to obtain a first model includes:

[0017] Dividing the first training data based on the first time period to obtain a plurality of first training samples; the first training samples include observation data within the first time period and observation data within a third time period before the first time period;

[0018] Constructing a long short-term memory network based on a preset network architecture and training parameters, and training the long short-term memory network using the first training sample to obtain a first model;

[0019] Accordingly, the forecasting of shallow winds in a first future time period using the first model includes:

[0020] Determining observation data to be processed within the third time period before the current time point;

[0021] The observation data to be processed is input into the first model, and the first model is used to forecast the shallow wind in the first time period after the current time point.

[0022] Optionally, the obtaining of a second model by training historical numerical forecast products based on the location information of the local station includes:

[0023] The historical numerical forecast products are used as the model background field, the forecast data after dynamic downscaling are output, and the forecast data of the local station is extracted;

[0024] Second training data is constructed using historical shallow wind observation data corresponding to the forecast data of the local station;

[0025] The pre-built long short-term memory network is trained using the second training data to obtain a second model.

[0026] Optionally, the using the second training data to train a pre-built long short-term memory network to obtain a second model includes:

[0027] Dividing the second training data into samples based on the second time period to obtain a plurality of second training samples; the second training samples include observation data within the second time period and forecast data within a fourth time period before the second time period;

[0028] The pre-built long short-term memory network is trained using the second training sample to obtain a corresponding second model.

[0029] Accordingly, the forecasting of shallow winds in a second future time period using the second model includes:

[0030] The second model is error-corrected using the forecast data for the target time period output by the second model and the observation data for the target time period to obtain a corresponding corrected model, so that the shallow wind for the future second time period can be forecasted using the corrected model.

[0031] In a second aspect, the present application provides a shallow wind forecasting device, comprising:

[0032] A first prediction module is configured to train a first model using historical shallow wind observation data of the station, and to use the first model to forecast shallow winds in a first future time period;

[0033] The second prediction module is used to obtain a second model based on the location information of the station and trained with historical numerical forecast products, and to use the second model to forecast shallow winds in a second time period in the future.

[0034] In a third aspect, the present application provides an electronic device, comprising:

[0035] Memory, used to store computer programs;

[0036] A processor is used to execute the computer program to implement the shallow wind forecasting method as described above.

[0037] In a fourth aspect, the present application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the shallow wind forecasting method as described above.

[0038] Thus, it can be seen that in the process of training the shallow wind forecast model, this application specifically utilizes the historical shallow wind observation data of this station to train a first model, and uses the first model to forecast the shallow wind for a first time period in the future; and based on the location information of the station, it can use historical numerical forecast products to train a second model, and use the second model to forecast the shallow wind for a second time period in the future. In this way, this application uses different data to train the first model and the second model respectively for shallow wind forecasts in different time periods; combined with the historical observation data and data forecast products of the current meteorological station, the accuracy of the shallow wind forecast can be targetedly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in 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 merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0040] Figure 1 This is a flow chart of a shallow wind forecasting method disclosed in this application;

[0041] Figure 2 A flow chart of a specific shallow wind forecasting method disclosed in this application;

[0042] Figure 3 This is a schematic structural diagram of a shallow wind forecasting device disclosed in this application;

[0043] Figure 4 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0045] See also Figure 1 As shown, an embodiment of the present invention discloses a shallow wind forecasting method, comprising:

[0046] Step S11: Use the historical shallow wind observation data of the station to train a first model, and use the first model to forecast the shallow wind in a first time period in the future.

[0047] It's understandable that leveraging the characteristics and patterns of shallow wind data at this station allows for refined forecasts of shallow winds for a specific timeframe. Furthermore, when the forecast horizon exceeds six hours, the accuracy of the numerical forecast product becomes even more stable. Specifically, when forecasting shallow winds for the next one to six hours, the station's historical shallow wind observations can be combined with machine learning models to maximize the patterns within these data. This historical shallow wind observation data can be used to train a corresponding first model, which can then be used to forecast shallow winds for a specific timeframe.

[0048] In a specific embodiment, training the first model using the station's historical shallow wind observation data may include: obtaining shallow wind data observed by the station's shallow wind measurement system during a first preset time period to obtain the historical shallow wind observation data; performing an importance assessment on the historical shallow wind observation data using a random forest algorithm, and filtering the historical shallow wind observation data based on the assessment results to obtain first training data; and using the first training data to train a pre-constructed long-short-term memory network to obtain the first model. Specifically, in the process of training the first model using the station's historical shallow wind observation data, it is first necessary to obtain shallow wind data observed at the station within a previous period; specifically, shallow wind observation data from more than three months prior may be obtained to obtain the historical shallow wind observation data. The importance of the historical shallow wind observation data may then be assessed using a random forest algorithm. It is understood that the importance of each characteristic variable varies under different forecast timeframes. Here, the importance of each meteorological element in the historical shallow wind observation data may be assessed using a random forest algorithm, and corresponding data for model training, i.e., the first training data, may be filtered based on an importance threshold. The pre-built long short-term memory network is then trained using the first training data to obtain a corresponding first model.

[0049] In another specific embodiment, the importance evaluation of the historical shallow wind observation data based on the random forest algorithm, and the screening of the first training data from the historical shallow wind observation data based on the evaluation results, may include: removing data that meets a preset data error condition from the historical shallow wind observation data to obtain processed observation data after removing the erroneous data; using the permutation importance of the random forest out-of-bag data to evaluate the importance of the processed observation data to obtain a corresponding evaluation result; and screening the first training data from the processed observation data based on the relationship between the evaluation result and a preset importance threshold. Specifically, the historical shallow wind observation data observed and recorded by this station may contain erroneous records. Here, a data error condition can be set to remove erroneous records from the recorded original observation data to prevent such data from interfering with the subsequent model training process. Furthermore, random forest out-of-bag data can be used to evaluate the importance of each characteristic variable in the historical shallow wind observation data after removing erroneous records. This can assess the importance of different meteorological factors to shallow wind forecasts at different forecast timescales. For example, meteorological factors such as three-hour pressure change, temperature change, wind direction, and wind speed have different importance in influencing future shallow wind changes. Based on the evaluation results and the physical relationship between each meteorological factor and the wind field, characteristic variables with importance below a preset threshold can be eliminated. The remaining characteristic variables are then used as important factors for wind field forecasting, thus obtaining the first training data for training the first model.

[0050] In another specific embodiment, training a pre-constructed long-short-term memory network using the first training data to obtain a first model may include: dividing the first training data based on the first time period to obtain a plurality of first training samples; the first training samples include observation data within the first time period and observation data within a third time period before the first time period; constructing a long-short-term memory network based on a preset network architecture and training parameters, and training the long-short-term memory network using the first training samples to obtain the first model; accordingly, forecasting shallow winds in a first future time period using the first model may include: determining unprocessed observation data within the third time period before the current time point; inputting the unprocessed observation data into the first model, and using the first model to forecast shallow winds in the first time period after the current time point. Specifically, in the process of training the pre-constructed long-short-term memory network using the first training data to obtain the first model, the first training data must first be divided; for example, to predict shallow winds within 6 hours, historical observation data from the past 7 days may be used; that is, corresponding first training samples are divided from the first training data; it is understandable that the first time period here is 6 hours and the third time period is 7 days. Furthermore, when constructing a long short-term memory network, the corresponding network structure and training parameters can be set. For example, an LSTM (Long Short-Term Memory) layer with 200 hidden units can be used with a drop probability of 0.5; a learning rate of 0.01 for the grid can be specified, and the gradient threshold can be set to 1 to prevent gradient explosion. Based on the network architecture and training parameter settings defined above, the network can be trained using the first training sample.

[0051] In a specific embodiment, the validation set can be used to test the model during the training process, the validation data is input into the model, and the prediction results are calculated through forward propagation; the model parameters are updated through back propagation, and the relevant personnel dynamically adjust the parameters of the model during training based on the difference between the prediction results and the actual results corresponding to the validation data; until the accuracy of the model's prediction of the validation set data during the training process meets the expectations, or there is no significant change in the multiple prediction results of different training processes, the training can be terminated to obtain the final model; the final model is subsequently used to forecast the shallow wind at this station.

[0052] Step S12: Based on the location information of the local station, a second model is trained using historical numerical forecast products, and the second model is used to forecast shallow winds in a second future time period.

[0053] In this application, the above steps can be used to train the historical shallow wind observation data of this station to obtain a first model for forecasting the shallow wind in the first time period in the future; since the forecast accuracy of the numerical forecast product will be more stable after the next 6 hours, the first model can be used to forecast the shallow wind within the next 6 hours; and for the shallow wind forecast after 6 hours, the forecast product of this station can be extracted from the original historical numerical forecast product, and the second model can be trained in combination with the historical observation data of the relevant time period of this station; the second model obtained by training forecasts the shallow wind after the next 6 hours, that is, forecasts the shallow wind in the second time period in the future.

[0054] In a specific embodiment, based on the location information of the local station, the second model is trained using historical numerical forecast products, which can include: using the historical numerical forecast products as the model background field, outputting the forecast data after dynamic downscaling, and extracting the forecast data of the local station; using the historical shallow wind observation data corresponding to the time of the local station forecast data to construct the second training data; using the second training data to train a pre-constructed long short-term memory network to obtain the second model. Specifically, the numerical forecast product needs to be dynamically downscaled first to obtain data with higher temporal and spatial resolution; the original numerical forecast product is used as the model background field, and the forecast data of the station is extracted according to the latitude and longitude of the station; the observation data of the historical period corresponding to the forecast data of the station are selected to form the second training data; in a specific embodiment, the forecast data of the station may include atmospheric factors such as air pressure, temperature, total cloud cover, wind field components, and visibility; because the wind field of general model forecast products is expressed in wind volume, in order to be consistent with the observed full wind direction and full wind speed, the 10-meter and 100-meter wind field components can be further synthesized into 10-meter wind direction, 10-meter wind speed and 100-meter wind direction, 100-meter wind speed. The formula used in wind field synthesis is as follows:

[0055] ;

[0056] ;

[0057] Where S is wind speed, D is wind direction, U and V are the east-west and north-south components of the wind field, respectively. The synthesized S and D are then merged with the aforementioned meteorological elements.

[0058] In a specific embodiment, the use of the second training data to train a pre-constructed long short-term memory network to obtain a second model may include: dividing the second training data into samples based on the second time period to obtain a plurality of second training samples; the second training samples include observation data within the second time period and forecast data within a fourth time period before the second time period; and using the second training samples to train the pre-constructed long short-term memory network to obtain a corresponding second model. Specifically, in the process of training the second model using the second training data, the eigenvectors in the second training data can be divided according to timeliness to obtain corresponding second training samples. It can be understood that the second training samples include observation data for the second time period and forecast data for the fourth time period before the second time period. For example, when the forecast time period is 7 hours and the fourth time period is t1, t2, ...tm, the second time period can be t1+7, t2+7, ...tm+7; that is, the shallow wind forecast value with a forecast time of k hours from the start time t1 (that is, the shallow wind at time t1+k predicted by the model forecast product after dynamic downscaling) should correspond to the shallow wind observation value at time t1+k, and so on. It can be understood that the second model is trained based on the input historical forecast data and the corresponding observation data, that is, the model is trained based on the error between the past model forecast product and the observation data.

[0059] In another specific embodiment, after the second model is used to forecast the shallow wind in the second time period in the future, it can also include: generating an error correction value for the second model based on the difference between the forecast data of the target time period output by the second model and the observed data of the target time period, and using the error correction value to correct the forecast data output by the second model. Specifically, after the second model is used to forecast the shallow wind in the second time period in the future, the forecast data output by the second model can also be corrected based on the difference between the forecast and the actual observed data. For example, the second model predicts the wind speed for the next day. On the 1st day, the wind speed for the 2nd day is predicted to be 4 meters / s, on the 2nd day, the wind speed for the 3rd day is predicted to be 2 meters / s, and on the 3rd day, the wind speed for the 4th day is predicted to be 3 meters / s. The actual observed wind speed every day is 2 meters / s. Then the errors for these three days are 2, 0, and 1 respectively, that is, the pattern of the second model forecasting shallow wind is generally 0-2 meters / s larger. If the model predicts today that tomorrow's wind speed is 5 meters / s, after correction by the machine learning model, the final forecast output today for tomorrow's wind speed may be 4 meters / s.

[0060] As can be seen from this, this application uses different data to train the first and second models for shallow wind forecasts in different time periods. Machine learning can maximize the historical characteristics and patterns of shallow wind at this station, making a refined shallow wind forecast for this station. For forecasts beyond a certain time limit, relevant numerical forecast products can be used, combined with the station's historical data, to construct a station-specific correction model for the numerical forecast product. This can combine the historical observation data of the current shallow wind measurement system and historical data forecast products to improve the accuracy of shallow wind forecasts in local areas based on timeliness and targetedness.

[0061] like Figure 2 As shown, the embodiment of the present application discloses a shallow wind forecasting method, wherein 1 to 6 hours is used as a first time period and 6 to 24 hours is used as a second time period, specifically including:

[0062] This embodiment is to establish a shallow wind short-term forecast model; it includes an RL (Reinforcement Learning) model for the first time period and an RL correction model for the second time period; wherein the RL model uses historical shallow wind actual observation data, adopts a random forest algorithm to calculate the importance of meteorological elements in the historical shallow wind observation data, selects characteristic variables with greater importance as input values of the forecast model, and inputs them into the long short-term memory network model, thereby constructing a sliding long short-term memory network RL wind field forecast model based on random forest feature selection; correspondingly, the RL correction model refers to obtaining background field data of the forecast product as a model, using WRF to perform dynamic downscaling and output forecast data with higher spatiotemporal resolution, and then using the random forest algorithm to screen the importance of meteorological elements in the forecast data, inputting the screened characteristic variables and the corresponding actual data into the long short-term memory network model, and constructing a sliding RL wind field forecast correction model.

[0063] Specifically, during the training of the first model, we first select the historical hourly observation data from the shallow wind measurement system at this station and sort it chronologically. For example, if we are predicting the hourly maximum wind speed at time t, where k is the prediction time (k=1, 2, 3, …, 6), the feature variables selected at time tk are air pressure (six layers), sea level pressure, 3-hour pressure variation (six layers), 24-hour pressure variation (six layers), temperature (six layers), 3-hour temperature variation (six layers), dew point temperature (six layers), relative humidity (six layers), maximum wind speed (six layers), wind direction at maximum wind speed (six layers), maximum wind speed (six layers), maximum wind direction (six layers), visibility, and rainfall. The six layers of data refer to meteorological element data at 10 meters, 30 meters, 50 meters, 70 meters, 80 meters, and 90 meters (this represents the installation height of the shallow wind measurement equipment for an example; actual altitudes can range from 0 to 100 meters). At the same time, data quality control is required to remove erroneous records in the original observation data to prevent such data "noise" from interfering with the learning process, thus obtaining a historical shallow wind observation dataset. Then, based on the calculation and selection of the importance of the feature variables of random forest, the feature vectors are screened according to their importance. It is understandable that the importance of each feature variable is different at different forecast time periods in wind speed and wind direction forecasts. The importance of feature variables can be calculated using the permutation importance of random forest out-of-bag data, for example, by setting the number of decision trees to 50. The importance value of each feature variable at each forecast time period is calculated. Based on the calculation results and the physical meaning between each meteorological element and the wind field, the feature variables with an importance value less than 1.0 are removed to reduce the impact of "noise" on the accuracy of the forecast model. The remaining feature variables are used as important factors in wind field forecasts.

[0064] Furthermore, to train the LSTM network, training samples are prepared using filtered historical shallow wind observation data. For each time period (k=1, 2, 3, 4, 5, 6), the corresponding raw data is a two-dimensional array of shape (total time steps, total number of feature variables corresponding to time period k). A sliding window approach is used to extract multiple overlapping samples from the raw data, forecasting the six shallow winds at time t+k (where t is the start time and k is the forecast time period). Hourly shallow wind data from the seven days preceding and including time t is used to construct a training sequence (shape 7*24, total number of feature variables corresponding to time period k). The response sequence is the maximum wind speed and direction of the six shallow winds at time t+k. This response sequence is used during the model training phase as labels or target values for the training data, helping the model learn the mapping from input features to wind speed output. For each forecast time t+k, a fixed-length 7-day window of historical data is selected as input. This data window slides along the time axis as the forecast time t+k progresses. The input to the LSTM model requires converting the raw data into a three-dimensional tensor of shape (batch_size, time_step, input_size). Batch_size is the number of data samples fed into the model, time_step is the number of time steps in each sample, and input_size is the number of features in each time step. For example, the raw data is arranged in chronological order: t1, t2, ..., t168, t169, t170, t171, t172, t173, t174, .... The first training sample (sliding window) is constructed as follows: the data corresponding to times t1-t168 is selected as training data at time t168, and the data at time t174 is the response data with a validity period of six hours from time t168. The next sample is: the data corresponding to times t2-t169 is selected as training data at time t169, and the data at time t175 is the response data with a validity period of six hours from time t169. Furthermore, it is necessary to normalize all variables used for training, calculate the mean and standard deviation of the eigenvalues, and normalize the training variables to zero mean and unit variance. In addition, it is necessary to predefine the network architecture and build a dynamic long short-term memory network, which may specifically include an LSTM layer with 200 hidden units and a dropout layer with a dropout probability of 0.5; specify training option parameters, such as specifying a grid learning rate of 0.01 and setting the gradient threshold to 1 to prevent gradient explosion. Based on the network architecture and training parameter settings defined above, the network can be trained using a training sequence. In a specific embodiment, the prediction results can be calculated by forward propagation, and the model parameters can be updated by backpropagation; this process will be iterated multiple times until the performance of the model on the validation set is no longer significantly improved.

[0065] Accordingly, the training process for the second model first requires WRF dynamical downscaling and data processing of the numerical forecast products. Specifically, the model background field can be selected from a global numerical forecast product published daily at 00:00 UTC, with a temporal resolution of 3 hours and a spatial resolution of 0.25°×0.25°. Timescales are selected for 3, 6, 9, 12, 15, 18, 21, and 24 hours. The model domain is designed using a triple nesting approach, with horizontal resolutions of 27 km×27 km, 9 km×9 km, and 3 km×3 km, respectively, from the outermost layer to the innermost layer. The grid dimensions are 300×200, 103×103, and 115×109, respectively. The longitude and latitude of the center point of the innermost layer serve as the longitude and latitude coordinates of the forecast point. The middle layer is located over South China, and the outermost layer encompasses the northwest Pacific. Model static data, including terrain elevation and land use types, are derived from the Moderate Resolution Imaging Spectroradiometer (MODIS). The model integration step is set to six times the model horizontal resolution, with an integration step of 300 seconds for the outermost layer and 18 seconds for the innermost layer. The model outputs forecasts every hour. The three-layer model uses the WSM6 scheme for regional microphysics, the RRTMG (Rapid Radiative Transfer Model of an Atmosphere) scheme for longwave and shortwave radiation, the Noah land surface model for land surface processes, the New Tiedtke (New Tiedtke Mass Flux Convection Scheme) scheme for cumulus convection parameterization, and the Shin-Hong (Shin and Hong Improved Boundary Layer Scheme) or YSU (Yonsei University Boundary Layer Parameterization Scheme) scheme for boundary layer parameterization. Cumulus convection parameterization is disabled in the innermost nested layer of the model. Then, select and output hourly forecast data for 15 meteorological elements: air pressure, sea level pressure, 3-hour pressure change, 24-hour pressure change, temperature, dew point temperature, relative humidity, low cloud cover, total cloud cover, 10-meter gusts over the past 3 hours, 10-meter wind field U component, 10-meter wind field V component, 100-meter wind field U component, 100-meter wind field V component, and visibility. The local station data is extracted based on longitude and latitude. It can be understood that the WRF model consists of two main parts: pre-processing and integral forecasting. The pre-processing part mainly performs: setting the simulation area range, preparing terrain, surface and soil information; decoding and format conversion of global background field data; horizontal interpolation and spatial matching of meteorological fields, etc.The integral prediction component primarily involves constructing the model's initial fields and boundary conditions, and performing iterative calculations for numerical integral prediction. A second-order Runge-Kutta explicit time-difference scheme is employed for time integration. The physical process parameterization scheme encompasses key modules such as radiative transfer, boundary layer turbulent mixing, convective triggering and organization, grid-scale turbulent diffusion, and cloud microphysics.

[0066] In a specific embodiment, in order to correspond to the shallow wind observation data, the 10-meter and 100-meter wind field components in the extracted forecast product can be synthesized into 10-meter wind direction, 10-meter wind speed and 100-meter wind direction, 100-meter wind speed; the synthesized elements are merged into the 15 meteorological elements of the forecast data, and a total of 19 meteorological elements can be obtained as characteristic variables.

[0067] Accordingly, based on the calculation and selection of feature variable importance using the random forest algorithm, the feature variables for each time period (k=7, 8, ... 24) were screened to obtain the corresponding second training data. For each time period (k=7, 8, ... 24), the feature variables at all onset times were selected to construct a training sequence with the shape of (total onset time step, total number of feature variables corresponding to the k time period). The response sequence was constructed with the shape of (total forecast time step, maximum shallow wind speed and direction at the sixth layer). For forecast time period k=7, the onset times were t1, t2, ... tm, and the forecast times were t1+7, t2+7, ... tm+7. After inputting the feature variables corresponding to onset time t1, the response sequence was the observed shallow wind direction and speed at forecast time t1+7 (the time point required for the forecast), and so on. It can be understood that for each forecast time t+7, a forecast data window of fixed length 1h is selected as input. As the forecast time progresses, this data window will also slide on the time axis; the number of data samples input to the model is m.

[0068] In a specific embodiment, using a trained model to forecast shallow winds requires constructing a forecast sequence. This is done using the same method used to construct training data, with the normalized forecast sequence selected according to the aforementioned criteria. The constructed forecast sequence values are then fed into the trained forecast network model, which then outputs a sequence of predicted shallow wind directions and speeds.

[0069] It can be seen that this application trains the first model and the second model respectively for shallow wind forecasts for different time periods, combining the historical observation data of the current shallow wind measurement system and the historical data forecast products; uses the random forest algorithm to calculate the importance of meteorological elements in the above data, selects characteristic variables with greater importance as input values of the forecast model and inputs them into the long-short-term memory network model to construct shallow wind forecast models for 1 to 6 hours and 6 to 24 hours, which can specifically improve the accuracy of shallow wind forecasts in local areas.

[0070] like Figure 3 As shown, the embodiment of the present application discloses a shallow wind forecasting device, comprising:

[0071] A first prediction module 11 is configured to train a first model using historical shallow wind observation data of the station, and use the first model to forecast shallow winds in a first time period in the future;

[0072] The second prediction module 12 is used to obtain a second model based on the location information of the station and trained with historical numerical forecast products, and to use the second model to forecast shallow winds in a second time period in the future.

[0073] It can be seen that this application uses different data to train the first model and the second model respectively for shallow wind forecasts in different time periods; combined with the historical observation data of the current shallow wind measurement system and the historical data forecast products, the accuracy of shallow wind forecasts can be targetedly improved.

[0074] In a specific embodiment, the first prediction module 11 may include:

[0075] A first data acquisition unit is configured to acquire shallow wind data observed by a shallow wind measurement system of the station during a first preset time period to obtain historical shallow wind observation data;

[0076] a data evaluation submodule, configured to perform importance evaluation on the historical shallow wind observation data based on a random forest algorithm, and to filter and obtain first training data from the historical shallow wind observation data based on the evaluation result;

[0077] The first training submodule is used to train the pre-built long short-term memory network using the first training data to obtain a first model.

[0078] In another specific embodiment, the data evaluation submodule may include:

[0079] a data elimination unit, configured to eliminate data that meets a preset data error condition from the historical shallow wind observation data, thereby obtaining processed observation data after eliminating the erroneous data;

[0080] a data evaluation unit, configured to perform importance evaluation on the processed observation data using the permutation importance of random forest out-of-bag data to obtain corresponding evaluation results;

[0081] A data screening unit is used to screen the processed observation data to obtain first training data based on the relationship between the evaluation result and a preset importance threshold.

[0082] In another specific embodiment, the first training submodule may include:

[0083] a first data partitioning unit, configured to partition the first training data based on the first time period to obtain a plurality of first training samples; the first training samples comprising observation data within the first time period and observation data within a third time period before the first time period;

[0084] A first training unit is configured to construct a long short-term memory network based on a preset network architecture and training parameters, and train the long short-term memory network using the first training sample to obtain a first model;

[0085] Correspondingly, the first training unit is specifically used to: determine the observation data to be processed in the third time period before the current time point; input the observation data to be processed into the first model, and use the first model to forecast the shallow wind in the first time period after the current time point.

[0086] In a specific embodiment, the second prediction module 12 may include:

[0087] The data extraction unit is used to use the historical numerical forecast products as the model background field, output the forecast data after dynamic downscaling, and extract the forecast data of the local station;

[0088] A training data construction unit, configured to construct second training data using historical shallow wind observation data corresponding to the forecast data of the local station;

[0089] The second training submodule is used to train the pre-built long short-term memory network using the second training data to obtain a second model.

[0090] In another specific embodiment, the second training submodule may include:

[0091] a second data partitioning unit configured to partition the second training data into samples based on the second time period to obtain a plurality of second training samples; the second training samples comprising observation data within the second time period and forecast data within a fourth time period prior to the second time period;

[0092] The second training unit is used to train the pre-built long short-term memory network using the second training sample to obtain a corresponding second model.

[0093] In a specific embodiment, the device may further include:

[0094] The error correction module is used to use the forecast data of the target time period output by the second model and the observation data of the target time period to correct the error of the second model and obtain a corresponding corrected model so as to use the corrected model to forecast the shallow wind in the second time period in the future.

[0095] Furthermore, the embodiment of the present application also discloses an electronic device, Figure 4 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram should not be considered as any limitation to the scope of application of the present application.

[0096] Figure 4 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps of the shallow wind forecast method disclosed in any of the aforementioned embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0097] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device. The communication protocol it follows is any communication protocol that can be applied to the technical solution of this application and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world. Its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0098] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or CD, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0099] The operating system 221 is used to manage and control the hardware devices and computer program 222 on the electronic device 20, and can be Windows Server, Netware, Unix, Linux, etc. In addition to including computer programs capable of implementing the shallow wind forecasting method performed by the electronic device 20 disclosed in any of the aforementioned embodiments, the computer program 222 can further include computer programs capable of implementing other specific tasks.

[0100] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when executed by a processor, the computer program implements the aforementioned shallow wind forecasting method. The specific steps of this method can be referred to the corresponding contents disclosed in the aforementioned embodiments and will not be repeated here.

[0101] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from the other embodiments. Reference can be made to the descriptions of the identical or similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and the relevant parts can be referred to the descriptions of the methods.

[0102] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0103] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0104] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0105] The above is a detailed introduction to the technical solution provided by the present application. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for those skilled in the art, according to the ideas of the present application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A shallow wind forecasting method, characterized in that: include: A first model is trained using historical shallow wind observation data of the station, and the first model is used to forecast shallow winds in a first future time period; Based on the location information of the local station, a second model is trained using historical numerical forecast products, and the second model is used to forecast shallow winds in a second time period in the future.

2. The shallow wind forecasting method according to claim 1, characterized in that: The first model is obtained by training the historical shallow wind observation data of the station, including: Obtaining shallow wind data observed by the shallow wind measurement system of the station in a first preset time period to obtain historical shallow wind observation data; Performing an importance assessment on the historical shallow wind observation data based on a random forest algorithm, and screening the historical shallow wind observation data to obtain first training data based on the assessment result; The pre-built long short-term memory network is trained using the first training data to obtain a first model.

3. The shallow wind forecasting method according to claim 2, characterized in that: The performing importance evaluation on the historical shallow wind observation data based on the random forest algorithm, and screening the historical shallow wind observation data to obtain first training data based on the evaluation result, includes: Eliminating data that meets a preset data error condition from the historical shallow wind observation data to obtain processed observation data after eliminating the erroneous data; The importance of the processed observation data is evaluated by using the permutation importance of the random forest out-of-bag data to obtain the corresponding evaluation results; Based on the relationship between the evaluation result and a preset importance threshold, first training data is obtained by screening from the processed observation data.

4. The shallow wind forecasting method according to claim 2, characterized in that: The method of training a pre-built long short-term memory network using the first training data to obtain a first model includes: Dividing the first training data based on the first time period to obtain a plurality of first training samples; the first training samples include observation data within the first time period and observation data within a third time period before the first time period; Constructing a long short-term memory network based on a preset network architecture and training parameters, and training the long short-term memory network using the first training sample to obtain a first model; Accordingly, the forecasting of shallow winds in a first future time period using the first model includes: Determining observation data to be processed within the third time period before the current time point; The observation data to be processed is input into the first model, and the first model is used to forecast the shallow wind in the first time period after the current time point.

5. The shallow wind forecasting method according to claim 1, characterized in that: The second model is obtained by training the historical numerical forecast products based on the location information of the local station, including: The historical numerical forecast products are used as the model background field, the forecast data after dynamic downscaling are output, and the forecast data of the local station is extracted; Second training data is constructed using historical shallow wind observation data corresponding to the forecast data of the local station; The pre-built long short-term memory network is trained using the second training data to obtain a second model.

6. The shallow wind forecasting method according to claim 5, characterized in that: The method of training the pre-built long short-term memory network using the second training data to obtain a second model includes: Dividing the second training data into samples based on the second time period to obtain a plurality of second training samples; the second training samples include observation data within the second time period and forecast data within a fourth time period before the second time period; The pre-built long short-term memory network is trained using the second training sample to obtain a corresponding second model.

7. The shallow wind forecasting method according to any one of claims 1 to 6, characterized in that: After forecasting the shallow wind in the second time period in the future by using the second model, the method further includes: The second model is error-corrected using the forecast data for the target time period output by the second model and the observation data for the target time period to obtain a corresponding corrected model, so that the shallow wind for the future second time period can be forecasted using the corrected model.

8. A shallow wind forecasting device, characterized in that: include: A first prediction module is configured to train a first model using historical shallow wind observation data of the station, and to use the first model to forecast shallow winds in a first future time period; The second prediction module is used to obtain a second model based on the location information of the station and trained with historical numerical forecast products, and to use the second model to forecast shallow winds in a second time period in the future.

9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the shallow wind forecasting method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that Used to store a computer program, which, when executed by a processor, implements the shallow wind forecasting method according to any one of claims 1 to 7.

Images