Wind field prediction data processing method and system based on deep learning

Abstract

The application discloses a kind of based on deep learning wind field forecast data processing method and system, belong to meteorological forecast big data analysis technical field.The application can make full use of a variety of meteorological data resources and is effectively processed according to integration instruction.Through observation element mining, past observation element embedding vector is obtained, and feature pooling and vector mining can deeply extract meteorological observation information features.Based on past cross-modal space mapping vector, meteorological element derivation and trend deduction are carried out, which expands the connotation of meteorological information.The generated meteorological forecast reference vector subsequently improves the accuracy of the vector.Discriminant analysis obtains wind field forecast training results and obtains prediction confidence features, which increases the credibility of the prediction results.Data assimilation thermal feature set locates target interactive variable to determine target wind field forecast processing network, and finally can accurately generate wind field forecast results according to current information, which improves the accuracy, reliability and practicality of wind field forecast as a whole.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meteorological forecast big data analysis, and particularly relates to a wind field forecast data processing method and system based on deep learning. BACKGROUND

[0002] In the field of meteorological forecasting, the accuracy and reliability of wind field forecasting have always been key issues. Traditional wind field forecasting methods often have many limitations. On the one hand, the use of meteorological observation information is not sufficient, relying only on data from a single source or a small number of sources, which cannot fully and accurately reflect the wind field conditions. For example, relying only on ground meteorological station data, without the fusion of multi-source data such as satellites and upper air sounding, will miss a lot of information of key meteorological elements. On the other hand, traditional methods lack the ability to deeply mine and integrate meteorological data. There are complex relationships between meteorological elements, and the ability to extract features, derive elements and deduce trends is insufficient, making it difficult to accurately grasp the internal mechanism of wind field changes. In addition, there is a lack of effective evaluation and adjustment mechanism in the forecasting process, making it difficult to determine the credibility of the forecasting results and to optimize and adjust for different forecasting situations. SUMMARY

[0003] The application provides a wind field forecast data processing method and system based on deep learning, which can solve or partially solve the technical problems involved in the background technology.

[0004] The embodiment of the present application provides a wind field prediction data processing method based on deep learning, which is applied to a wind field prediction data processing system. The method comprises the following steps: obtaining past multi-source meteorological observation information and past data integration indication information; in an initial wind field prediction processing network, the past data integration indication information is used to mine observation elements from the past multi-source meteorological observation information, so as to obtain a past observation element embedding vector; the observation element mining is used for feature pooling and vector mining; the past data integration indication information is used to perform cross-modal space mapping on the past observation element embedding vector, so as to obtain a past cross-modal space mapping vector; the past observation element embedding vector and the past cross-modal space mapping vector are subjected to vector interaction and feature optimization by using original interaction variables, so as to obtain a meteorological prediction reference vector; the cross-modal space mapping is used for meteorological element derivation and meteorological trend deduction; the meteorological prediction reference vector is subjected to discriminant analysis through the initial wind field prediction processing network, so as to obtain a wind field prediction training result and obtain a prediction confidence feature of the wind field prediction training result; a data assimilation thermal feature set is generated from the original interaction variables and the prediction confidence feature; a target interaction variable is located from the data assimilation thermal feature set; and an initial wind field prediction processing network including the target interaction variable is determined as a target wind field prediction processing network; the target wind field prediction processing network is used to generate a wind field prediction result according to current data integration indication information.

[0005] The embodiment of the present application provides a wind field prediction data processing system, which comprises at least one processor and a memory; the memory stores computer execution instructions; and the at least one processor executes the computer execution instructions stored in the memory, so that the at least one processor executes the method described above.

[0006] The embodiment of the present application provides a readable storage medium, in which a program or instructions are stored; when the program or instructions are executed by a processor, the steps of the method described above are implemented.

[0007] The present application can make full use of various meteorological data resources and effectively process according to integration indications. The past observation element embedding vector is obtained through observation element mining, and feature pooling and vector mining can deeply extract meteorological observation information features. Meteorological element derivation and trend deduction are performed based on the past cross-modal space mapping vector, which expands the meteorological information connotation. The meteorological prediction reference vector generated subsequently improves the vector accuracy. Discriminant analysis obtains the wind field prediction training result and obtains the prediction confidence feature, which increases the credibility of the prediction result. The data assimilation thermal feature set locates the target interaction variable to determine the target wind field prediction processing network, which finally can accurately generate the wind field prediction result according to the current information, and overall improves the accuracy, reliability and practicability of the wind field prediction. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 A flowchart of a wind field forecast data processing method based on deep learning provided by an embodiment of the present application.

[0009] Figure 2 A structural schematic diagram of a wind field forecast data processing system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0010] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0011] The terms “first”, “second”, and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by “first”, “second”, and the like are generally of a kind and do not limit the number of objects, for example, the first object can be one or more. In addition, “and / or” in the present application means at least one of the connected objects, and the character “ / ” generally represents a “or” relationship between the front and rear associated objects.

[0012] Figure 1 A wind field forecast data processing method based on deep learning is shown, which is applied to a wind field forecast data processing system, and the method comprises the following steps 202-208.

[0013] Step 202: obtaining past multi-source meteorological observation information and past data integration indication information.

[0014] Step 204: in the initial wind field forecast processing network, using the past data integration indication information to mine observation elements from the past multi-source meteorological observation information to obtain a past observation element embedding vector; the observation element mining is used for feature pooling and vector mining; and using the past data integration indication information to perform cross-modal space mapping on the past observation element embedding vector to obtain a past cross-modal space mapping vector, using the original interaction variable to perform vector interaction and feature optimization on the past observation element embedding vector and the past cross-modal space mapping vector to obtain a meteorological forecast reference vector; the cross-modal space mapping is used for meteorological element derivation and meteorological trend deduction.

[0015] Step 206: discriminant analysis is performed on the meteorological forecast reference vector by the initial wind field forecast processing network to obtain a wind field forecast training result, and a forecast confidence feature of the wind field forecast training result is obtained.

[0016] Step 208: a data assimilation thermal feature set is generated from the original interaction variable and the forecast confidence feature, a target interaction variable is located from the data assimilation thermal feature set, and an initial wind field forecast processing network including the target interaction variable is determined as a target wind field forecast processing network; the target wind field forecast processing network is used to generate a wind field forecast result according to current data integration indication information.

[0017] In this application, the wind field forecast data processing system plays an important role in building a high-resolution mesoscale numerical weather prediction system. First, in step 202, the wind field forecast data processing system will obtain past multi-source meteorological observation information and past data integration indication information. Past multi-source meteorological observation information contains meteorological data from multiple sources, such as wind speed, wind direction, temperature, pressure, and other data observed by ground meteorological stations, cloud information obtained by satellite remote sensing, and atmospheric vertical structure information obtained by high-altitude sounding balloons, etc. These data sources are diverse, and their formats and accuracies may vary. Past data integration indication information is used to guide the subsequent processing of multi-source meteorological observation information. It can contain information about data types, data quality identification, relationships between different data sources, etc. For example, some specific indication information may indicate that the data from a certain high-precision meteorological station should have a higher weight in wind field forecasting, or the fusion method of different source data under certain weather conditions, etc.

[0018] Then go to step 204, a series of complex operations are carried out in the initial wind field forecast processing network. The past multi-source meteorological observation information is subjected to observation element mining using the past data integration indication information. Taking wind speed data as an example, feature pooling operation may summarize data with similar wind speed characteristics at different times and different locations. Vector mining further mines the hidden relationships behind these wind speed data, such as the potential relationship between the trend of wind speed change over time in a certain geographical area and the surrounding environmental factors, to obtain past observation element embedding vectors. The observation element embedding vector is a deep feature representation of the wind field related observation elements in the multi-source meteorological observation information.

[0019] Then the past data integration indication information is used to carry out cross-modal space mapping on the past observation element embedding vector. In terms of meteorological element derivation, for example, original wind speed and wind direction data, other meteorological elements related to the wind field, such as wind energy density, can be derived through cross-modal space mapping. For meteorological trend deduction, taking the monsoon in a certain region as an example, according to long-term wind speed and wind direction observation data, the development trend of the monsoon, whether it is gradually increasing or weakening, can be deduced through cross-modal space mapping, so as to obtain a past cross-modal space mapping vector. Then the original interaction variable is used to carry out vector interaction and feature optimization between the past observation element embedding vector and the past cross-modal space mapping vector. The original interaction variable can promote information exchange between the two. For example, the original interaction variable can adjust the weight of the derived elements such as wind energy density in the entire vector according to the wind speed and wind direction data, perform feature optimization, and obtain a meteorological forecast reference vector. This meteorological forecast reference vector is the result of deep mining, mapping, interaction and optimization of multi-source meteorological observation information, and is an important basis for subsequent wind field prediction.

[0020] Subsequently, in step 206, the meteorological forecast reference vector is subjected to discriminant analysis by an initial wind field prediction processing network. The discriminant analysis evaluates the meteorological forecast reference vector according to some pre-set rules and models. For example, for wind field prediction, the discriminant analysis may, according to the wind speed, wind direction and related derived elements in the meteorological forecast reference vector, in combination with the wind field situation under similar meteorological conditions in history, judge the accuracy of the current prediction. Through this discriminant analysis, a wind field prediction training result is obtained, which can include the prediction of wind field in different time periods and different regions. At the same time, the prediction confidence feature of the wind field prediction training result is obtained. The prediction confidence feature reflects the reliability degree of the wind field prediction training result. For example, if in a certain region, according to historical data and current meteorological conditions, the confidence of wind speed prediction is high, while the confidence of wind direction prediction is relatively low, then the prediction confidence feature will clearly reflect this difference.

[0021] Finally, in step 208, a data assimilation thermal feature set is generated from the original interaction variables and the forecast confidence features. The data assimilation thermal feature set is a collection of various information. For example, it can include information about the relationships between different meteorological elements in the original interaction variables, and information about the reliability of wind field forecasts in the forecast confidence features. A target interaction variable is located from the data assimilation thermal feature set. The target interaction variable is a very important part of the entire wind field forecast data processing system, which determines how the initial wind field forecast processing network generates wind field forecast results according to the current data integration indication information. For example, the target interaction variable can be a variable closely related to wind speed and wind direction under specific regional and meteorological conditions. The initial wind field forecast processing network including the target interaction variable is determined as the target wind field forecast processing network. This target wind field forecast processing network can generate accurate wind field forecast results according to the current data integration indication information, such as the type and quality of current meteorological observation data.

[0022] Further in-depth exploration of some specific cases in the initial wind field forecast processing network. This network can use a specific deep learning neural network architecture, such as some variants of convolutional neural network (CNN) or recurrent neural network (RNN). Taking the convolutional neural network as an example, its convolutional layer can effectively extract local features from multi-source meteorological observation information, and can extract local features of wind speed, wind direction, etc. at different spatial positions in wind field forecasting. The pooling layer can compress and simplify these features, reducing the amount of data while retaining key information. When performing observation element mining, the convolutional neural network can adaptively learn the complex relationships between different meteorological observation elements by continuously adjusting the parameters of the convolution kernel, thereby obtaining more accurate past observation element embedding vectors.

[0023] In the cross-modal space mapping process, for example, an attention mechanism-based method is used. The attention mechanism can focus on the most important parts of meteorological element derivation and meteorological trend inference. For example, when performing meteorological element derivation, for wind-related meteorological elements, the attention mechanism will pay more attention to the interaction between wind speed, wind direction, and atmospheric vertical structure, etc. key factors, so as to more accurately derive elements such as wind energy density. In terms of meteorological trend inference, the attention mechanism can focus on factors that have a greater impact on wind field trends, such as terrain, land-sea distribution, etc. according to historical data and current meteorological conditions, so as to better infer the development trend of the wind field.

[0024] In the discriminant analysis link, a decision tree-based model can be used. The decision tree model classifies and evaluates the weather forecast reference vector through a series of conditional judgments. For example, taking wind speed prediction as an example, the root node of the decision tree can be the air pressure value under the current weather condition, and the data is divided into different subsets according to the size of the air pressure value, and then in each subset, further judgment is made according to other meteorological elements such as temperature, humidity, etc., and finally the accuracy evaluation of the wind field prediction training result and the prediction confidence feature are obtained.

[0025] For the generation of the data assimilation thermal feature set, a statistical analysis-based method can also be involved. For example, statistical analysis is performed on the data in the original interaction variable and the prediction confidence feature, and the correlation, variance, etc. between them are calculated. According to these statistical indicators, the data assimilation thermal feature set is constructed, so that this feature set can comprehensively reflect various information in the wind field prediction. In positioning the target interaction variable, a clustering analysis-based method can be used. The variables in the data assimilation thermal feature set are clustered according to their similarity, and then in each cluster, the variable that is most relevant to the wind field prediction and contributes most to improving the prediction accuracy is found as the target interaction variable.

[0026] In practical application, taking the wind field prediction of a coastal area as an example. Past multi-source weather observation information includes wind speed, wind direction, temperature, humidity, etc. data of multiple weather stations in the coastal area, as well as ocean surface temperature, cloud map, etc. information obtained by satellite remote sensing. The past data integration indication information shows that in this coastal area, the ocean surface temperature has a greater impact on the wind field, and the data from the high-precision weather station has a higher weight. In the initial wind field prediction processing network, through observation element mining, the wind speed data of different weather stations is feature-pooled and vector-mined according to its geographical distribution and time sequence to obtain the past observation element embedding vector. In the cross-modal space mapping, according to the ocean surface temperature and wind speed, wind direction data, meteorological elements such as ocean wind energy resource distribution are derived through the method based on attention mechanism, and the development trend of sea wind is deduced to obtain the past cross-modal space mapping vector. The original interaction variable is used to interact and optimize the two vectors to obtain the weather forecast reference vector. Through the discriminant analysis based on the decision tree, the wind field prediction training result and the prediction confidence feature are obtained. For example, the prediction result shows that the wind speed of the sea wind in a certain time period is in the range of 5-10 meters / second, and the wind direction is southeast, while the prediction confidence feature shows that the confidence of the wind speed prediction is 80%, and the confidence of the wind direction prediction is 70%. Then, through the statistical analysis-based method, the data assimilation thermal feature set is generated, and then the target interaction variable is located based on clustering analysis, such as the variable closely related to the ocean surface temperature and sea wind speed, and finally the target wind field prediction processing network is determined, so as to generate more accurate wind field prediction results, and provide accurate wind field prediction support for wind energy development, marine shipping, etc. activities in the coastal area.

[0027] As can be seen, the wind field forecast data processing system, through the above steps, makes full use of past multi-source meteorological observation information and integrated indication information, and finally determines the target wind field forecast processing network to generate the wind field forecast result by means of various technical means in the initial wind field forecast processing network, thereby providing a more accurate and effective wind field forecast solution for the meteorological forecasting field.

[0028] It should be noted that in the wind field forecast data processing system, the past observation element embedding vector, the past cross-modal space mapping vector, the meteorological forecast reference vector, the original interaction variable, and the target interaction variable are the key elements for constructing an accurate wind field forecast.

[0029] I. Past observation element embedding vector

[0030] The past observation element embedding vector is a vector representation obtained after in-depth mining of past multi-source meteorological observation information. Past multi-source meteorological observation information is widely sourced, covering ground meteorological stations, satellites, high-altitude detection equipment, etc., and includes wind speed, wind direction, temperature, air pressure, and other meteorological elements.

[0031] Taking a certain region as an example, the region is distributed with multiple ground meteorological stations. At a certain time, the meteorological element values measured by each meteorological station are as follows: wind speed ranges between 2-15 meters / second, wind direction covers 0°-360°, temperature is between -10℃ and 30℃, and air pressure is between 950-1050 hundred pascals.

[0032] When performing observation element mining, the feature pooling operation groups wind speed data according to geographical regions. For example, the wind speed of meteorological stations in the valley area is generally low, such as meteorological stations A, B, and C, with wind speed ranging from 2-5 meters / second, which are grouped together; the wind speed of meteorological stations D, E, and F in the open plain area is relatively high, ranging from 8-15 meters / second, which are grouped together.

[0033] Vector mining further mines the relationship behind the grouped data. For example, in the low wind speed group (meteorological stations A, B, and C), when the temperature is between -5℃-0℃ and the air pressure is between 1030-1050 hundred pascals, the wind speed tends to maintain at 2-3 meters / second, which can be represented in vector form as [2-3, -5-0, 1030-1050], where the first component represents the wind speed range, the second component represents the temperature range, and the third component represents the air pressure range; when the temperature rises to 0℃-5℃ and the air pressure slightly decreases to 1020-1030 hundred pascals, the wind speed may increase to 3-5 meters / second, which can be represented as [3-5, 0-5, 1020-1030]. For the high wind speed group (meteorological stations D, E, and F), there are similar relationships, such as when the temperature is between 15℃-20℃ and the air pressure is between 980-1000 hundred pascals, the wind speed is between 10-12 meters / second, which can be represented as [10-12, 15-20, 980-1000].

[0034] The relationship information of each meteorological element under different conditions is integrated to obtain a past observation element embedding vector. This vector can include multiple sub-vectors similar to the above, comprehensively representing the wind speed characteristics of different meteorological stations under different meteorological conditions and the relationship between meteorological elements, laying a foundation for subsequent processing.

[0035] II. Past cross-modal space mapping vector

[0036] The past cross-modal space mapping vector is obtained based on the past observation element embedding vector through cross-modal space mapping, which is of great significance in meteorological element derivation and meteorological trend deduction.

[0037] (I) Meteorological element derivation

[0038] Taking the wind field and temperature field in a certain region as an example, the wind speed data in the past observation element embedding vector is 8 meters per second, the wind direction is southwest, and the temperature is between 15-20℃. According to the principles of aerodynamics and meteorology, meteorological element derivation is carried out. Given that the air density is, for example, 1.2 kg / m3, according to the kinetic energy formula, the kinetic energy value when the wind speed is 8 meters per second is 0.5 x 1.2 x 8^2 = 38.4. This derived kinetic energy value is related to the original observation elements such as wind speed and temperature, and the numerical representation in the past cross-modal space mapping vector can be [8, 15-20, 38.4], where the first component represents the wind speed, the second component represents the temperature range, and the third component represents the kinetic energy value.

[0039] (II) Meteorological trend deduction

[0040] Consider the past month's meteorological data sequence of a coastal area. At the beginning of the month, the wind speed is 3-5 meters per second, the wind direction is north, and the temperature is about 10℃; at the end of the month, the wind speed reaches 8-10 meters per second, the wind direction is south, and the temperature rises to 15-20℃. The past observation element embedding vector contains the wind speed, wind direction, and temperature information at different times.

[0041] Through cross-modal space mapping for meteorological trend deduction, it is found that the increase in temperature increases the sea-land temperature difference, leading to the strengthening of sea breeze and the change of wind direction. For example, if the temperature continues to rise in the future, according to the past trend relationship, the wind speed may continue to increase to 10-12 meters per second, and the wind direction may further shift to the south to be closer to the south. In the past cross-modal space mapping vector, this trend information can be represented as [3-5-8-10-10-12, north-south-more south, 10-15-20], where the first component represents the wind speed trend, the second component represents the wind direction trend, and the third component represents the temperature trend.

[0042] The past cross-modal space mapping vector not only contains new elements derived from the original meteorological elements, but also covers meteorological trend prediction information, which is a more comprehensive and in-depth vector representation of meteorological phenomena, providing more basis for weather forecasting.

[0043] III. Meteorological forecasting reference vector

[0044] The meteorological forecasting reference vector is obtained by vector interaction and feature optimization of the original interaction variables based on the past observation element embedding vector and the past cross-modal space mapping vector.

[0045] The original interaction variable plays a key role in connecting and adjusting the relationship between different vectors and elements in the process. For example, in wind field forecasting, the weight of wind speed is 0.4, the weight of wind direction is 0.3, the weight of temperature is 0.2, and the weight of air pressure is 0.1.

[0046] Taking a certain area as an example, the wind speed data in the past observation element embedding vector is 6 meters per second on average, and the wind speed kinetic energy data in the past cross-modal space mapping vector is 30 (for example, according to the specific numerical value calculated earlier). The original interaction variable adjusts the interaction relationship between the two according to the weight. Assuming that there is a quantitative relationship between wind speed and wind speed kinetic energy, if the wind speed kinetic energy is proportional to the wind speed, the adjusted wind speed related value is 6x0.4+30x0.4=14.4.

[0047] In terms of feature optimization, if there is an abnormal wind speed value of 18 meters per second (obviously deviating from the normal range) in the past observation element embedding vector of a certain meteorological station due to instrument failure, the original interaction variable can be corrected according to the wind speed data of other normal meteorological stations and the meteorological trend information in the past cross-modal space mapping vector. For example, according to the wind speed and meteorological trend of the surrounding meteorological stations, it is judged that the normal wind speed of this area should be 4-6 meters per second, and the original interaction variable corrects this abnormal value to 6 meters per second (or reduces its influence in the vector).

[0048] After such vector interaction and feature optimization, the meteorological forecasting reference vector combines the advantages of the past observation element embedding vector and the past cross-modal space mapping vector. For example, the meteorological forecasting reference vector may be [14.4 (adjusted wind speed related value), a certain direction (adjusted wind direction related value), a certain temperature value (adjusted temperature related value), and a certain air pressure value (adjusted air pressure related value)], which more accurately reflects the relationship between meteorological elements and the nature of meteorological phenomena, providing reliable basis for wind field forecasting discriminant analysis.

[0049] IV. Original interaction variable

[0050] The original interaction variable is an important part of the wind field forecasting data processing system, which connects and adjusts the relationship between different vectors and elements in the forecasting process.

[0051] (1) Wind speed and wind direction relationship based on geographical area

[0052] The values and definitions of the original interaction variables are based on in-depth understanding of the meteorological system and a large amount of data statistical analysis. For example, in a mountainous area, the wind direction has a significant impact on the wind speed. When the wind direction is perpendicular to the mountain direction, the wind speed can be significantly increased; when the wind direction is parallel to the mountain direction, the wind speed is reduced.

[0053] For example, in a mountainous area, when the wind direction is perpendicular to the mountain direction, the impact coefficient of wind direction on wind speed is set to 1.5; when the wind direction is parallel to the mountain direction, the impact coefficient is 0.5. If the angle between the wind direction and the mountain direction at a certain time is 30° (for example, the relationship between the angle and the impact coefficient can be quantified), the impact coefficient of wind direction on wind speed at this time is calculated to be about 1.3 according to the trigonometric function relationship. This 1.3 is the value of the original interaction variable that reflects the special relationship between wind direction and wind speed.

[0054] (2) Interaction variables related to spatial and temporal scales

[0055] The original interaction variables are related to the spatial and temporal scales of meteorological elements. For short-term (hourly) wind field prediction, in urban environments, building height and density affect wind speed and wind direction. For example, in a city, there are high-rise buildings with a height of 50-300 meters and a density of 10-50 buildings per square kilometer. The original interaction variable sets the interaction variable value related to wind speed and wind direction according to these values.

[0056] For example, when the building height is 200 meters and the density is 30 buildings per square kilometer, the attenuation coefficient of wind speed near the ground is 0.6 (this 0.6 is the value contained in the original interaction variable) after analysis and statistics, which accurately reflects the impact of urban environment on wind field.

[0057] (3) Original interaction variables in the interaction of different meteorological elements

[0058] For the interaction of wind field and temperature field, the original interaction variable contains the heat transfer coefficient and other values related to heat exchange. In a coastal area, the sea water temperature is 20°C and the air temperature on land is 25°C, and the wind blows from the land to the sea. If the heat transfer coefficient is large (for example, 0.8, indicating fast heat exchange), the temperature changes faster during the wind blowing to the sea, which in turn affects the change of wind speed and wind direction. The original interaction variable accurately adjusts the interaction of each element in the wind field prediction by containing these different meteorological element-related values.

[0059] V. Target interaction variables

[0060] The target interaction variables are located from the data assimilation thermal characteristics set, which is crucial for determining the target wind field prediction processing network.

[0061] The assimilated thermodynamic feature set is generated from the original interaction variables and the prediction confidence features, and contains the meteorological element relationship information in the original interaction variables and the prediction reliability information in the prediction confidence features. For example, when the wind speed prediction confidence of a certain area is 80%, the wind speed and wind direction, temperature, pressure, and other meteorological element relationship values are included, which are the embodiment of the original interaction variables under a specific prediction confidence.

[0062] Taking the wind field prediction of a large wind farm as an example, the stability of the wind field and the accuracy of the wind speed in this area are crucial to the power generation efficiency. There can be multiple variables related to the wind field in the assimilated thermodynamic feature set, such as wind speed and terrain relationship variables, wind speed and atmospheric stability relationship variables, wind direction and seasonal relationship variables, etc.

[0063] The power plant is located in the channel between the valley and the mountain, and the terrain has a significant impact on the wind field. The terrain slope is between 5°-15°, and the wind speed enhancement coefficient is 1.2-1.5 (according to actual measurement and statistics). Through analysis, it is found that the terrain factor is the most critical, and the variables related to wind speed and terrain, such as the wind speed enhancement coefficient of 1.3 corresponding to a terrain slope of 10°, will be positioned as target interaction variables.

[0064] Once the target interaction variables are determined, the initial wind field prediction processing network containing the variables is determined as the target wind field prediction processing network. This network can integrate current data integration indication information (such as meteorological observation data, data quality information, etc.), focus on considering the meteorological element relationships represented by the target interaction variables, and generate more accurate wind field prediction results. For example, in the wind field prediction of a wind farm, the target wind field prediction processing network can more accurately predict wind speed and wind direction based on the terrain wind speed enhancement coefficient of 1.3 and the current meteorological conditions, providing reliable wind field information for the operation and management of the wind farm, and ensuring the maximization of power generation efficiency.

[0065] As can be seen, the past observation element embedding vector, the past cross-modal space mapping vector, the meteorological prediction reference vector, the original interaction variable, and the target interaction variable are interrelated and interact in the wind field prediction data processing system, and together build a complete wind field prediction technical solution, providing a more accurate and effective wind field prediction method for the meteorological prediction field. This technical solution can make full use of various meteorological observation data and information, deeply mine the complex relationships between meteorological elements, and through continuous adjustment and optimization, improve the accuracy and reliability of wind field prediction, and meet the needs of different fields (such as wind power generation, aerospace, navigation, etc.) for wind field prediction.

[0066] In an alternative technical approach, the obtaining of the past multi-source meteorological observation information and the past data integration indication information includes: obtaining an initial wind field prediction processing network, obtaining target prediction training labels and a deep learning network debugging set corresponding to the initial wind field prediction processing network; the initial wind field prediction processing network is a deep learning network that meets a pre-debugging condition; past multi-source meteorological observation information containing observation elements corresponding to the target prediction training labels is obtained from the deep learning network debugging set, and past data integration indication information is generated according to the target prediction training labels.

[0067] In this alternative technical approach, a specific way of obtaining past multi-source meteorological observation information and past data integration indication information is involved.

[0068] First, an initial wind field prediction processing network is obtained. This initial wind field prediction processing network is a deep learning network that meets a pre-debugging condition. Deep learning networks have powerful data analysis and processing capabilities, and in wind field prediction, they can perform deep mining and analysis on complex meteorological data. For example, in an actual wind field prediction scenario, this deep learning network may have been pre-trained to a certain extent, and its internal neuron structure and parameters have been preliminarily adjusted to a state where it can perform basic processing on meteorological data.

[0069] Next, target prediction training labels and a deep learning network debugging set corresponding to the initial wind field prediction processing network are obtained. Target prediction training labels are very important data that provide clear target guidance for the entire wind field prediction process. For example, target prediction training labels may include ideal prediction results of wind fields in a specific region, such as accurate numerical ranges of wind speed and precise directions of wind direction in a certain geographical region within a period of time. These labeled data provide a reference for subsequent processing. The deep learning network debugging set is a collection of various data that can be used to debug and optimize the deep learning network.

[0070] Then, past multi-source meteorological observation information containing observation elements corresponding to the target prediction training label is obtained from the deep learning network debugging set. For example, if the target prediction training label focuses on the wind speed and wind direction of a certain area, the past multi-source meteorological observation information containing data related to the two elements is selected from the deep learning network debugging set. Past multi-source meteorological observation information comes from a wide range of sources, including ground meteorological stations, satellites, upper air sounding equipment, etc. For example, the wind speed measured by a ground meteorological station at a certain time is between 3-10 meters per second, and the wind direction is between 0°-180°. These values are part of the past multi-source meteorological observation information. Satellites can provide meteorological element information in a more macro area, and upper air sounding equipment can provide information related to the vertical structure of the atmosphere. These different sources of data together constitute rich past multi-source meteorological observation information.

[0071] Finally, past data integration indication information is generated according to the target prediction training label. Taking wind speed prediction as an example, if the target prediction training label has a high accuracy requirement for wind speed, such as an error of ±1 meter per second, the past data integration indication information will integrate the data from different sources according to this requirement. For ground meteorological station data with high accuracy, a higher weight may be given, while satellite data may need to be further processed or adjusted in weight to meet the accuracy requirement of wind speed prediction in the target prediction training label. If the target prediction training label indicates that wind direction is more critical for wind field prediction under certain meteorological conditions (such as air pressure between 980-1000 hectopascals), the past data integration indication information will emphasize the integration method of wind direction related data, such as special feature extraction of wind direction data or specific association analysis with other meteorological elements during data processing.

[0072] By applying the above embodiment, first, the target prediction training label provides an accurate target for wind field prediction, making the entire prediction process more targeted. For example, in the prediction of wind speed and wind direction, data can be accurately processed according to the label requirements to improve the accuracy of the prediction. Second, past multi-source meteorological observation information is obtained from the deep learning network debugging set, which can make full use of existing data resources. These data contain rich meteorological information, and by selecting observation elements related to the target prediction training label, interference from irrelevant data can be avoided. Finally, past data integration indication information is generated according to the target prediction training label, which can reasonably integrate data from different sources according to different prediction requirements, and make reasonable weight allocation and processing method selection for different data, thereby improving the performance of the entire wind field prediction system and providing more reliable wind field prediction results for meteorological prediction.

[0073] On the basis of the technical idea, the past multi-source meteorological observation information containing the observation elements corresponding to the target prediction training label is obtained from the deep learning network debugging set in any one of the following manners.

[0074] Manner 1: Identify the past prediction training labels of the multi-source observation data examples included in the deep learning network debugging set, obtain the label feature commonality value between the past prediction training labels of the multi-source observation data examples and the target prediction training label, and determine the multi-source observation data examples corresponding to the past prediction training labels with a label feature commonality value not less than a prediction label commonality threshold as the past multi-source meteorological observation information.

[0075] Manner 2: Determine a selected wind field to which a wind field prediction event corresponds according to the target prediction training label, analyze the selected wind field from the multi-source observation data examples included in the deep learning network debugging set, and determine the multi-source observation data examples of the selected wind field obtained by the analysis as the past multi-source meteorological observation information.

[0076] In the wind field prediction data processing system, based on the technical idea mentioned above, there are two ways to obtain past multi-source meteorological observation information containing observation elements corresponding to the target prediction training label from the deep learning network debugging set, which will be described in detail as follows.

[0077] First, let's look at manner 1. The deep learning network debugging set contains a large number of multi-source observation data examples, each of which has its corresponding past prediction training label. Identifying the past prediction training labels of these multi-source observation data examples is the first step of this manner. For example, in the deep learning network debugging set, there is a multi-source observation data example whose past prediction training label contains information such as the wind speed in a certain region being in the range of 5-10 meters per second, the wind direction being in a certain range of angles, the temperature being in the range of 15-20℃, etc. The target prediction training label may contain more accurate range of wind speed in the region such as 6-8 meters per second, more detailed range of wind direction and similar temperature range, etc.

[0078] Then the annotation feature commonality value between the past prediction training annotation and the target prediction training annotation of the multi-source observation data example is obtained. The determination of this annotation feature commonality value needs detailed comparative analysis of each meteorological element. Taking wind speed as an example, if the wind speed range in the past prediction training annotation of the multi-source observation data example has a large overlapping part with the wind speed range in the target prediction training annotation, then it has a higher annotation feature commonality value on the wind speed element. For example, the wind speed range in the past prediction training annotation of the multi-source observation data example is 5-10 m / s, and the wind speed range in the target prediction training annotation is 6-8 m / s, and they have a certain overlapping part. Similar comparative analysis is also carried out on other meteorological elements such as wind direction and temperature, and finally the annotation feature commonality value is obtained.

[0079] For example, the prediction annotation commonality threshold is set to 0.6 (this value is determined according to the requirements of the system for data and past experience), and the multi-source observation data example corresponding to the past prediction training annotation with an annotation feature commonality value not less than the prediction annotation commonality threshold is determined as past multi-source meteorological observation information. If a multi-source observation data example has a commonality value of 0.7 on meteorological elements such as wind speed, wind direction and temperature after calculation, which is greater than the prediction annotation commonality threshold of 0.6, then this multi-source observation data example will be determined as past multi-source meteorological observation information. This means that the meteorological observation data in this multi-source observation data example has a high correlation with the target prediction training annotation and can provide valuable data for wind field prediction.

[0080] Look at way 2. According to the target prediction training annotation, the selected wind field to which the wind field prediction event is directed is determined. For example, the target prediction training annotation clearly indicates that the wind field in a certain coastal area is predicted, and this coastal area is the selected wind field. This selected wind field has its unique geographical features, meteorological environment and other factors.

[0081] The selected wind field is analyzed from the multi-source observation data examples included in the deep learning network debugging set. The multi-source observation data examples in the deep learning network debugging set contain meteorological observation data from different data sources, such as data from ground meteorological stations, satellites, etc. In the analysis process, data related to the selected wind field is found from these large amounts of data. For example, in the data of the ground meteorological station, the wind speed, wind direction, temperature, pressure, etc. observed by the meteorological station located in the coastal area and its surrounding area within a certain range are selected. In the satellite data, the cloud information covering the coastal area, the atmospheric temperature distribution and other data related to the wind field are extracted. The multi-source observation data examples of the selected wind field obtained by analysis are determined as past multi-source meteorological observation information. These determined past multi-source meteorological observation information are closely related to the selected wind field to which the target prediction training annotation is directed, and can provide accurate data basis for wind field prediction.

[0082] Further, the method 1 can accurately screen out the multi-source observation data examples highly related to the target prediction training label as the past multi-source meteorological observation information by comparing the annotated feature commonality value with the predicted annotation commonality threshold. For example, by comparing the wind speed, wind direction and other meteorological elements to determine the commonality value, irrelevant data is avoided, and the data is more targeted and effective. The method 2 can focus on a specific wind field area and extract data related to the wind field from numerous multi-source observation data examples by analyzing the selected wind field to determine the past multi-source meteorological observation information. For example, in the coastal area wind field prediction, the meteorological observation data of the coastal area is accurately obtained, thereby providing accurate and targeted data support for the wind field prediction, and improving the accuracy and reliability of the wind field prediction.

[0083] On the basis of the above technical ideas, the past multi-source meteorological observation information containing the observation elements corresponding to the target prediction training label is obtained from the deep learning network debugging set, which further includes any one of the following methods.

[0084] Method 3: Obtain the original meteorological observation information example containing the observation elements corresponding to the target prediction training label from the deep learning network debugging set, and determine the original meteorological observation information example as the past multi-source meteorological observation information.

[0085] Method 4: Obtain the original meteorological observation information example containing the observation elements corresponding to the target prediction training label from the deep learning network debugging set, and obtain the past multi-source meteorological observation information by preprocessing the original meteorological observation information example.

[0086] Method 5: Obtain the original meteorological observation information example containing the observation elements corresponding to the target prediction training label from the deep learning network debugging set, obtain the observation state label of the original meteorological observation information example, and determine the past multi-source meteorological observation information from the original meteorological observation information example according to the observation state label.

[0087] Under the technical framework of the wind field prediction data processing system, based on the above technical ideas, there are different methods such as method 3, method 4 and method 5 for obtaining the past multi-source meteorological observation information containing the observation elements corresponding to the target prediction training label from the deep learning network debugging set, which will be described in detail as follows.

[0088] The method 3 is relatively direct. Examples of original meteorological observation information containing the observation elements corresponding to the target prediction training label are obtained from the deep learning network debugging set, and then they are determined as the past multi-source meteorological observation information. The deep learning network debugging set contains rich examples of original meteorological observation information, which covers observation data of various meteorological elements. For example, the observation elements corresponding to the target prediction training label are the wind speed and direction of a certain region. In the deep learning network debugging set, there are some examples of original meteorological observation information, which explicitly record the wind speed value range (such as 3-10 m / s) and the specific angle range (such as 0°-180°) of the wind direction of the region. Directly determining these examples of original meteorological observation information containing the observation elements concerned by the target prediction training label as the past multi-source meteorological observation information is a simple and direct way to obtain the data related to the target prediction training label, which provides a basic data source for the subsequent wind field prediction process.

[0089] The method 4 adds a preprocessing step on the basis of the method 3. First, examples of original meteorological observation information containing the observation elements corresponding to the target prediction training label are also obtained from the deep learning network debugging set. For example, the wind speed data of a certain region in the example of original meteorological observation information fluctuates within a certain range, with a wind speed value of 2-12 m / s and a wind direction angle range of 0°-360°, and also contains data of other meteorological elements such as temperature and pressure.

[0090] Then, the examples of original meteorological observation information are preprocessed. The purpose of preprocessing is to improve the data quality and make it more suitable for wind field prediction. For example, for wind speed data, there may be some abnormal values caused by measurement device precision or environmental interference. If the wind speed recorded at a certain time is 15 m / s, which is obviously higher than the normal wind speed range of the region, it can be corrected by preprocessing methods such as data smoothing processing and abnormal value elimination. For example, after processing, the abnormal 15 m / s wind speed is corrected to 8 m / s, which is within the normal range of the region. For wind direction data, if the record is disordered, it can be arranged by certain algorithms (such as angle average algorithm) to make the wind direction angle range more accurate. After such preprocessing, the past multi-source meteorological observation information obtained has been improved in terms of data accuracy and usability, which is more conducive to the subsequent wind field prediction process.

[0091] The method 5 also obtains examples of original meteorological observation information containing the observation elements corresponding to the target prediction training label from the deep learning network debugging set, but then obtains the observation state label of the example of original meteorological observation information, and determines the past multi-source meteorological observation information from the example of original meteorological observation information according to the observation state label.

[0092] After obtaining the original meteorological observation information examples, each example is attached with an observation state label. The observation state label reflects some characteristics of the original meteorological observation information example. For example, the observation state label can contain the quality level of the data (such as high quality, medium quality, and low quality), the reliability of the data source (such as data from a high-precision meteorological station is marked as high reliability, and data from a part of simple observation equipment is marked as low reliability), and the meteorological environment state at the time of observation (such as normal meteorological environment, observation under complex meteorological environment), and the like.

[0093] Taking the wind speed and direction observation of a certain area as an example, the wind speed data in the original meteorological observation information example is between 4-10 meters / second, and the wind direction is between 90°-270°. The observation state label shows that the data is from a meteorological station with medium reliability, and is observed under complex meteorological environment, and the quality level is medium quality. According to these observation state labels, if the wind field forecasting system has a higher requirement for data quality, the original meteorological observation information example can be filtered according to the preset rules (such as preferentially selecting data from a high-reliability source, under normal meteorological environment, and with high quality). If the original meteorological observation information example does not meet the requirements, it can not be determined as past multi-source meteorological observation information. If there is another original meteorological observation information example, the wind speed is between 3-9 meters / second, the wind direction is between 120°-240°, the observation state label shows that the data is from a high-reliability meteorological station, is observed under normal meteorological environment, and has high quality, then this original meteorological observation information example is more likely to be determined as past multi-source meteorological observation information. Through this filtering method according to the observation state label, the determined past multi-source meteorological observation information has high quality and reliability, thereby improving the accuracy of wind field forecasting.

[0094] In combination with the above, the method 3 directly obtains the original meteorological observation information example as the past multi-source meteorological observation information, which simply and efficiently provides a data basis related to the target prediction training label. The method 4 can correct the abnormal values in the data, improve the accuracy and availability of the data, and make the data relied on by subsequent wind field forecasting more reliable. The method 5 determines the past multi-source meteorological observation information from the original meteorological observation information example according to the observation state label, which can filter out data with higher quality and stronger reliability, and avoid the influence of low-quality or unreliable data on wind field forecasting. In combination with the three methods, more high-quality, accurate, and targeted past multi-source meteorological observation information can be obtained from the deep learning network debugging set, which provides a solid data support for wind field forecasting, and improves the precision and reliability of wind field forecasting.

[0095] In a preferred example, the past observation element embedding vector includes an observation element dynamic framework vector, a physical process parameterization vector, and a mesoscale weather model vector. Based on this, the observation element mining of the past multi-source meteorological observation information in the initial wind field prediction processing network based on the past data integration indication information to obtain the past observation element embedding vector includes: in the initial wind field prediction processing network, performing dynamic framework feature mapping on the multi-source meteorological observation element vector corresponding to the past multi-source meteorological observation information to obtain the observation element dynamic framework vector; the dynamic framework feature mapping is used to strengthen the observation element in the dynamic framework state; performing feature quantization on the observation element dynamic framework vector to obtain the physical process parameterization vector; the physical process parameterization vector is used to represent the meteorological physical process state corresponding to the past multi-source meteorological observation information; obtaining a past integration indication vector corresponding to the past data integration indication information, and performing weighted integration on the physical process parameterization vector and the past integration indication vector to obtain the mesoscale weather model vector; the mesoscale weather model vector is used to represent the meteorological joint description feature of the past multi-source meteorological observation information and the past data integration indication information.

[0096] On the basis of the above example, the method further includes: performing convolution operation on the past multi-source meteorological observation information to obtain a past meteorological observation convolution vector; generating a past observation offset feature according to the vector size of the past meteorological observation convolution vector; and performing weighted integration on the past meteorological observation convolution vector and the past observation offset feature to obtain the multi-source meteorological observation element vector.

[0097] In the wind field prediction data processing system, the technical scheme of this preferred example is developed around the construction of the past observation element embedding vector and its related operations, and this process contains multiple complex and interrelated links. First, for the construction of the past observation element embedding vector, it contains three important parts: the observation element dynamic framework vector, the physical process parameterization vector, and the mesoscale weather model vector. The starting step of this construction process involves the processing of past multi-source meteorological observation information.

[0098] The past multi-source meteorological observation information contains a variety of meteorological data, which comes from different observation sources, such as numerous ground meteorological stations, satellite observation equipment, and upper air detection instruments, etc. The data provided by these data sources covers multiple meteorological elements such as wind speed, wind direction, temperature, air pressure, and humidity, and each meteorological element has different numerical records at different observation points and observation times.

[0099] In the initial wind field prediction processing network, to obtain the past observation element embedding vector, the past multi-source meteorological observation information needs to be convolved to obtain the past meteorological observation convolution vector. Convolution operation is a widely used technique in data processing, and it also has important significance in meteorological data processing. Taking the meteorological observation data of a certain area as an example, for example, there are 10 ground meteorological stations in this area, and each meteorological station observes the wind speed, wind direction, temperature, and pressure every 1 hour. At a certain time, the wind speed values observed by these meteorological stations may fluctuate between 0.5-15 meters / second, the wind direction covers all possible directions from 0° to 360°, the temperature varies between -10℃ to 30℃, and the air pressure varies between 950-1050 hundred pascals. When the multi-source meteorological observation information is convolved, the convolution kernel will slide on the data according to certain rules, thereby extracting the local features of the data and obtaining the past meteorological observation convolution vector. This vector integrates the local relationship features between different meteorological stations and different meteorological elements, for example, it may highlight the special correlation between wind speed and temperature in a small area.

[0100] According to the vector size of the past meteorological observation convolution vector, the past observation offset feature is generated. The vector size of the past meteorological observation convolution vector reflects the structural characteristics of the data after convolution operation. For example, if the size of the past meteorological observation convolution vector is [20, 10] (here the values represent the size of the vector in two dimensions), this size information will be used to generate the past observation offset feature. The generated past observation offset feature is related to the vector size, and it may adjust the data in the convolution vector according to a specific algorithm to better adapt to subsequent processing. This adjustment is designed based on a deep understanding of the structure and features of meteorological data, aiming to dig out deeper relationships in meteorological data.

[0101] Then the past meteorological observation convolution vector and the past observation offset feature are weighted and integrated to obtain the multi-source meteorological observation element vector. Weighted integration is a method of considering the information of two vectors comprehensively. For example, the value of an element in the past meteorological observation convolution vector is 5 (this value may represent the wind speed of a meteorological station after convolution operation), and the corresponding element value in the past observation offset feature is 3. If the set weighting coefficients are 0.6 and 0.4 respectively (these weighting coefficients are determined according to the importance of different information to the system), then the corresponding element value in the multi-source meteorological observation element vector obtained after weighted integration is 5×0.6+3×0.4=4.2. In this way, the multi-source meteorological observation element vector integrates the information of the past meteorological observation convolution vector and the past observation offset feature, providing a basis for subsequent construction of the observation element dynamic framework vector.

[0102] Next, the multi-source meteorological observation element vector is mapped to a dynamic framework feature to obtain an observation element dynamic framework vector. The dynamic framework is a basic framework for describing atmospheric movement and changes in meteorology, and the dynamic framework feature mapping is a mapping process of data in the multi-source meteorological observation element vector according to the requirements of the dynamic framework. For example, in the dynamic framework of the atmosphere, there is a complex dynamic relationship between wind speed and wind direction, which is affected by many factors such as terrain, pressure gradient, etc. In the multi-source meteorological observation element vector, the values of wind speed and wind direction may be simply observed values, such as wind speed of 8 meters / second and wind direction of 120°. Through dynamic framework feature mapping, the terrain characteristics of the area (such as whether it is a plain area or a mountainous area) and the pressure gradient situation (such as the pressure gradually decreasing from east to west) are considered, and the wind speed and wind direction are adjusted. After dynamic framework feature mapping, the wind speed in the observation element dynamic framework vector may be 9 meters / second and the wind direction may be 125°. The adjusted values can more accurately reflect the true relationship between wind speed and wind direction in the dynamic framework state, and strengthen the observation elements in the dynamic framework state.

[0103] After that, the observation element dynamic framework vector is quantized to obtain a physical process parameterization vector. Meteorological physical processes are very complex, involving heat exchange, water vapor transport, energy conversion, and many other aspects. The purpose of the physical process parameterization vector is to represent the state of the meteorological physical process corresponding to the past multi-source meteorological observation information in a quantitative form. For example, in the observation element dynamic framework vector, in addition to wind speed and wind direction, there is also temperature of 18°C and pressure of 1010 hPa. In the physical process, temperature and pressure interact with wind speed and wind direction. Through feature quantization, the values of these meteorological elements are converted to a specific quantitative representation. For example, the wind speed after feature quantization is 10 (this value is obtained according to the wind speed of 9 meters / second and the specific quantization rule), the wind direction is 15 (the quantized value corresponding to 125°), the temperature is 15 (the quantized value corresponding to 18°C), and the pressure is 12 (the quantized value corresponding to 1010 hPa). These quantized values constitute the physical process parameterization vector, which can accurately represent the relationship between the elements in the meteorological physical process.

[0104] The past data integration indication information includes various indication information for integrating the past multi-source meteorological observation information, such as weight distribution of data from different data sources, importance identification of different meteorological elements in the integration process, and the like. The past integration indication vector is to represent the indication information in the form of a vector. For example, the weights of wind speed, wind direction, temperature, and pressure in the past integration indication vector are respectively set to 0.3, 0.2, 0.3, and 0.2 (these weight values are determined according to the demand of meteorological forecast and the evaluation of the importance of different meteorological elements). The wind speed quantization value in the physical process parameterization vector is 10, the wind direction is 15, the temperature is 15, and the pressure is 12. In the weighted integration, the corresponding component values in the mesoscale weather mode vector are calculated according to the weights, for example, the wind speed component is 10*0.3=3, the wind direction component is 15*0.2=3, the temperature component is 15*0.3=4.5, and the pressure component is 12*0.2=2.4. The mesoscale weather mode vector is used to represent the meteorological joint description characteristics of the past multi-source meteorological observation information and the past data integration indication information, which integrates meteorological observation data and integration indication information, and provides more comprehensive and accurate feature description for wind field prediction.

[0105] Therefore, the whole technical solution constructs the past observation element embedding vector through a series of rigorous operations, and provides a comprehensive and accurate basis for wind field prediction. Starting from the initial convolution operation, the local features of the original meteorological observation data are gradually integrated, so that the data can better reflect the internal relationship between meteorological elements in subsequent processing. The dynamic framework feature mapping strengthens the relationship between meteorological elements in the dynamic framework state, so that the relationship between elements such as wind speed and wind direction is more in line with the actual situation of atmospheric movement, and the description accuracy of meteorological dynamic process is improved. The physical process parameterization vector obtained by feature quantization can accurately represent the state of meteorological physical process, which is helpful to deeply understand the interaction mechanism between meteorological elements. Finally, the mesoscale weather mode vector obtained by weighted integration combines observation information and integration indication information, which provides more comprehensive and detailed information for wind field prediction, and helps to improve the accuracy and reliability of wind field prediction, so as to better meet the demand of meteorological forecast and other related fields for wind field prediction.

[0106] In an optional embodiment, the past data integration indication information is used to map the past observation element embedding vector in the cross-modal space to obtain a past cross-modal space mapping vector, and the past observation element embedding vector and the past cross-modal space mapping vector are interacted by using the original interaction variable and optimized in feature to obtain a meteorological forecast reference vector, comprising:

[0107] (1) When u is a positive integer greater than 1 and not greater than X, in the u-th spatial mapping core of the initial wind field prediction processing network, the (u-1)-th meteorological prediction reference vector is cross-modal spatially mapped by using the past data integration indication information to obtain the u-th past cross-modal spatial mapping vector, the X-u+1-th past observed element embedding vector generated by the X-u+1-th element embedding core is obtained, the u-th original interaction variable in the u-th spatial mapping core is used to interact the X-u+1-th past observed element embedding vector and the u-th past cross-modal spatial mapping vector to obtain the u-th past interaction vector, if u is X, the X-th past interaction vector is optimized in feature to obtain a meteorological prediction reference vector; X is the number of spatial mapping cores included in the initial wind field prediction processing network;

[0108] (2) When u is 1, in the first spatial mapping core of the initial wind field prediction processing network, the X-th past observed element embedding vector is cross-modal spatially mapped by using the past data integration indication information to obtain the first past cross-modal spatial mapping vector, and the first original interaction variable is used to interact the X-th past observed element embedding vector and the first past cross-modal spatial mapping vector to obtain the first past interaction vector.

[0109] In this optional embodiment, the technical solution is developed around the use of past data integration indication information and original interaction variables to process vectors to obtain meteorological prediction reference vectors. First of all, it needs to be clear that the entire processing process is carried out under the framework of the initial wind field prediction processing network, which contains X spatial mapping cores, and X is the key parameter for determining the number of spatial mapping cores. Each spatial mapping core has a unique and indispensable role in the entire process.

[0110] When u is a positive integer greater than 1 and not greater than X, for example, X is 10, when u=3, a series of operations begin in the 3rd spatial mapping core of the initial wind field prediction processing network.

[0111] In this space mapping kernel, first, the second meteorological forecast reference vector is cross-modally space mapped using the past data integration indication information. The past data integration indication information is a rich information set that covers a variety of integration-related instruction information for meteorological data. For example, for the wind speed meteorological element, the past data integration indication information may specify the integration weight of the wind speed data from different observation sources (such as ground meteorological stations, satellite observations, etc.) in a specific geographical area or under specific meteorological conditions. Similar specifications exist for other meteorological elements such as wind direction, temperature, and pressure. When cross-modally space mapping, the information in the second meteorological forecast reference vector is converted according to these indication information. For example, the part of the second meteorological forecast reference vector that contains the wind speed feature is represented as the wind speed being roughly in the range of 5-8 meters / second in a certain area, and the wind direction being in a specific angle range. After cross-modally space mapping, the third past cross-modally space mapping vector may establish a new relationship between this wind speed range and other meteorological elements in different modalities (such as different spatial dimensions or different meteorological physical process-related modalities). For example, it may establish a relationship between wind speed and the temperature gradient in the vertical direction of the atmosphere, forming a new vector representation in which there is a specific correspondence between the wind speed range and the temperature gradient value. This relationship is discovered in the cross-modally space mapping process.

[0112] At the same time, the X-u+1th past observation element embedding vector generated by the X-u+1th element embedding kernel is obtained. When X=10 and u=3, the 8th past observation element embedding vector generated by the 8th element embedding kernel is obtained. The past observation element embedding vector is the result of processing by the element embedding kernel and contains the feature information of meteorological observation data after deep mining. Taking wind speed as an example, the wind speed-related features in this past observation element embedding vector may not only be a simple numerical range, but may also include wind speed trend information at different time series, and correlation information between wind speed and other meteorological elements (such as temperature, pressure, etc.) observed by surrounding meteorological stations. For example, in this past observation element embedding vector, the wind speed changes in a certain trend of first increasing and then decreasing at several specific times, and this trend is associated with the temperature change of the surrounding meteorological station in a certain way, which is represented in the vector by a specific numerical or coding method.

[0113] Then the u-th original interaction variable in the u-th spatial mapping kernel is used to perform vector interaction between the X-u+1-th past observation element embedding vector and the u-th past cross-modal spatial mapping vector. The original interaction variable plays a crucial role in this process as a bridge and adjustment. The original interaction variable is based on in-depth understanding of the meteorological system and a large amount of statistical analysis of meteorological data. For example, for wind speed and wind direction, the 3rd original interaction variable in the 3rd spatial mapping kernel may specify the interaction coefficient between wind speed and wind direction under a certain meteorological environment (such as a certain pressure range and temperature range). For example, under this meteorological environment, the pressure is between 980-1000 hundred pascal, the temperature is between 15-20℃, and the interaction coefficient between wind speed and wind direction is specified as 0.6 (this value is obtained from a large number of historical data statistical analysis, indicating that the influence degree of wind direction on wind speed under this meteorological condition is 0.6). When performing vector interaction, if the wind speed related eigenvalue in the 8th past observation element embedding vector is a certain value, and the wind direction related eigenvalue is a certain angle, the wind speed and wind direction related eigenvalues in the 3rd past cross-modal spatial mapping vector are converted, and the original interaction variable will perform weighted sum or other forms of interaction operation on the wind speed and wind direction related features in the two vectors according to the interaction coefficient, thereby obtaining the 3rd past interaction vector. This past interaction vector integrates the information of the two vectors, and accurately reflects the relationship between meteorological elements through the original interaction variable, so that the relationship between meteorological elements is more accurately reflected in the vector.

[0114] When u is X, taking X=10 as an example, when u=10, the 10th past interaction vector is optimized for feature to obtain a meteorological forecast reference vector. In this feature optimization process, a comprehensive check and adjustment of the 10th past interaction vector is involved. For example, the 10th past interaction vector may contain feature representations of multiple meteorological elements, and there may be some conditions between the eigenvalues of wind speed, wind direction, temperature, pressure and other elements that do not conform to the basic principles of meteorology or historical data statistical rules. For example, in this vector, the wind speed corresponding to the wind speed eigenvalue in a certain region is very high, but at the same time there is no corresponding significant change in the pressure eigenvalue, which is contrary to the general relationship between pressure and wind speed in meteorology. In the feature optimization process, these eigenvalues will be adjusted according to the principles of meteorology and historical data statistical rules. It may refer to the normal range of pressure values under similar wind speed conditions in history to adjust the pressure eigenvalue to a more reasonable value. At the same time, the relationship between other meteorological elements such as wind direction and temperature will also be checked and adjusted. Through such a feature optimization process, the final meteorological forecast reference vector can more accurately reflect the true relationship between meteorological elements and be more suitable for meteorological forecasting.

[0115] When u is 1, in the first spatial mapping core of the initial wind field prediction processing network, the past observation element embedding vector of the Xth past observation is cross-modally spatially mapped using the past data integration indication information. When X = 10, the 10th past observation element embedding vector is cross-modally spatially mapped to obtain the first past cross-modal spatial mapping vector. This process is also based on the past data integration indication information. The 10th past observation element embedding vector contains rich meteorological observation element information, and in the cross-modal spatial mapping process, the past data integration indication information guides the conversion of this information to different modalities to mine new relationships between meteorological elements. For example, the wind speed and direction observed on the ground and the satellite-observed atmospheric cloud information (related to the wind speed and direction through some conversion) can be combined to obtain the first past cross-modal spatial mapping vector, and the information in this vector is no longer limited to traditional meteorological element values, but contains more cross-modal associated information. Then, the first past cross-modal spatial mapping vector and the Xth past observation element embedding vector are vector-interacted using the first original interaction variable to obtain the first past interaction vector. The first original interaction variable plays an initial adjustment role in this process. For example, the first original interaction variable can perform preliminary weighting or adjustment operations on the feature values of wind speed and wind direction in the two vectors, so that the first past interaction vector can accurately reflect the interaction relationship of meteorological observation elements in the initial stage, laying a foundation for subsequent operations in other spatial mapping cores.

[0116] As can be seen, by performing a series of complex operations in a specific order in multiple spatial mapping cores in the initial wind field prediction processing network, the information of the past observation element embedding vector and the past cross-modal spatial mapping vector is effectively integrated. The operations in each spatial mapping core are closely dependent on the past data integration indication information and the original interaction variable, and this design enables the relationships between meteorological elements to be deeply mined, accurately adjusted and optimized. From the operation process of different spatial mapping cores, when u is different, the operations are sequentially performed, gradually integrating and optimizing the relationships between meteorological elements in different modalities, which can better adapt to the complexity and diversity of meteorological data. The meteorological prediction reference vector is obtained by optimizing the features of the past interaction vector, improving the accuracy and reliability of this vector, thereby providing a more accurate basis for wind field prediction, which helps to improve the accuracy, stability and response capability to complex weather conditions of wind field prediction.

[0117] In a possible technical solution, the acquiring the prediction confidence feature of the wind field prediction training result comprises: acquiring a relationship influence coefficient between the wind field prediction training result and the past data integration indication information, acquiring a prediction inversion label of the wind field prediction training result; performing prediction viewpoint evaluation on the wind field prediction training result by using a prediction viewpoint evaluation network to obtain a prediction accuracy feature of the wind field prediction training result; and combining the relationship influence coefficient, the prediction inversion label, and the prediction accuracy feature to obtain the prediction confidence feature of the wind field prediction training result.

[0118] In addition, the method further comprises: acquiring a quantitative discrimination interval, acquiring a network debugging expectation index, acquiring a global discrimination score from the quantitative discrimination interval according to the network debugging expectation index, determining a target discrimination score according to the global discrimination score, and converting the target discrimination score into an inversion matching accuracy score and a prediction accuracy score. The combining the relationship influence coefficient, the prediction inversion label, and the prediction accuracy feature to obtain the prediction confidence feature of the wind field prediction training result comprises: combining the global discrimination score, the inversion matching accuracy score, and the prediction accuracy score, and combining the relationship influence coefficient, the prediction inversion label, and the prediction accuracy feature to obtain the prediction confidence feature of the wind field prediction training result.

[0119] In the wind field prediction data processing system, the above technical solution relates to the acquisition of the prediction confidence feature of the wind field prediction training result and operations related thereto. First, for the process of acquiring the prediction confidence feature of the wind field prediction training result, a plurality of key operations are involved. The relationship influence coefficient between the wind field prediction training result and the past data integration indication information is first acquired. The past data integration indication information contains various instructions and guiding information for the integration of past multi-source meteorological observation information. The wind field prediction training result is the result of wind field prediction obtained through a series of processing (such as the various vector operations in the initial wind field prediction processing network mentioned above). The relationship influence coefficient reflects the degree of association between the two. For example, the wind speed prediction part in the wind field prediction training result may have a close relationship with the weight distribution of the wind speed data source and the integration method of different regional wind speed data in the past data integration indication information. If a high-precision wind speed observation source is given a higher weight in the past data integration indication information, and the wind speed prediction in the wind field prediction training result is highly consistent with the data of the high-precision observation source, then the relationship influence coefficient may be relatively high, such as 0.8.

[0120] Then the prediction inversion label of the wind field prediction training result is obtained. The prediction inversion label is a kind of identification of the wind field prediction training result, which can contain various attribute information about the prediction result. For example, the prediction inversion label can identify that the wind field prediction training result is obtained under specific meteorological conditions (such as specific pressure range, temperature range, etc.), or is a prediction result for a specific area (such as the wind field of a certain geographical area). For example, the prediction inversion label indicates that the wind field prediction training result is for a coastal area, and is obtained under the meteorological conditions of pressure 980-1000 hPa and temperature 15-20°C.

[0121] Then the prediction viewpoint evaluation network is used to evaluate the prediction viewpoint of the wind field prediction training result, and the prediction accuracy feature of the wind field prediction training result is obtained. The prediction viewpoint evaluation network is a network structure specially used to evaluate the prediction accuracy of the wind field, which is constructed based on a large amount of meteorological data, meteorological principles and historical prediction experience. When performing prediction viewpoint evaluation, it will comprehensively consider the prediction of each meteorological element in the wind field prediction training result. For example, for wind speed prediction, if the wind field prediction training result predicts that the wind speed of a certain area is between 5-8 m / s, and the actual meteorological observation data (which can be accurate data in the validation data set) shows that the wind speed of the area is between 6-7 m / s, then the prediction accuracy feature will give a corresponding evaluation result according to the difference. For example, this evaluation result is in the range of 0-1, and according to the accuracy of the above wind speed prediction, the prediction accuracy feature obtained may be 0.8 (indicating that the prediction is relatively accurate).

[0122] Then the prediction confidence feature of the wind field prediction training result is obtained by combining the relationship influence coefficient, the prediction inversion label and the prediction accuracy feature. This process is a comprehensive consideration of the influence of multiple factors on the prediction confidence. The relationship influence coefficient reflects the close relationship between the wind field prediction training result and the past data integration indication information, the prediction inversion label provides background information about the prediction result, and the prediction accuracy feature directly reflects the accuracy of the prediction. For example, if the relationship influence coefficient is 0.8, the prediction inversion label indicates that the prediction result is obtained under relatively stable meteorological conditions (in this case, the reliability of the prediction is usually high), and the prediction accuracy feature is 0.8, then the prediction confidence feature obtained by comprehensively considering these factors may be a comprehensive numerical value, such as 0.85 (here, the calculation is a kind of comprehensive consideration of each factor, and is not a simple weighted average formula calculation).

[0123] In addition, the technical solution also relates to other related operations. First, a quantitative discrimination interval is obtained, which is a numerical range interval for discriminating and evaluating the wind field prediction result. For example, the interval can be set as [0, 1], where 0 represents that the prediction is completely inaccurate, and 1 represents that the prediction is completely accurate. At the same time, a network debugging expectation index is obtained, which is an index determined based on the target requirements of the wind field prediction system and debugging experience. The global discrimination score is obtained from the quantitative discrimination interval according to the network debugging expectation index. For example, the network debugging expectation index requires a high prediction accuracy, and according to this index, the global discrimination score that can be obtained from the quantitative discrimination interval [0, 1] is 0.7.

[0124] The target discrimination score is determined according to the global discrimination score, which is a more targeted score further determined on the basis of the global discrimination score. For example, according to certain specific rules in the wind field prediction system or the requirements of certain meteorological elements, the global discrimination score 0.7 is further adjusted to the target discrimination score 0.65. Then, the target discrimination score is converted into the inversion matching score and the prediction accuracy score. This conversion process is based on specific conversion rules, which are developed according to meteorological principles and the characteristics of the wind field prediction system. For example, the target discrimination score 0.65 can be converted into the inversion matching score 0.3 and the prediction accuracy score 0.35 according to the conversion rules (the specific conversion rules will vary depending on the complexity of the system).

[0125] Finally, the global discrimination score, the inversion matching score, and the prediction accuracy score are used when obtaining the prediction confidence feature of the wind field prediction training result in combination with the relationship influence coefficient, the prediction inversion label, and the prediction accuracy feature. For example, in consideration of the relationship influence coefficient 0.8 mentioned above, the prediction inversion label (indicating the prediction under stable meteorological conditions), the prediction accuracy feature 0.8, and the global discrimination score 0.7, the inversion matching score 0.3, and the prediction accuracy score 0.35, etc., the final prediction confidence feature is obtained through a complex comprehensive mechanism (this mechanism is established based on a comprehensive understanding of the wind field prediction system and a large amount of data verification, and is not a simple linear combination), for example, the final prediction confidence feature is 0.82.

[0126] It can be seen that by obtaining the relationship influence coefficient between the wind field prediction training result and the past data integration indication information, the correlation between the prediction result and the data integration indication information can be considered, and the scientificity of the prediction can be ensured. The prediction inversion label provides background information of the prediction result, which is helpful for comprehensively evaluating the reliability of the prediction. The prediction accuracy feature obtained by using the prediction view evaluation network directly reflects the accuracy degree of the prediction. The prediction confidence feature obtained by comprehensively considering these factors and the global discrimination score, the inversion matching score and the prediction accuracy score can comprehensively and accurately measure the confidence of the wind field prediction from multiple angles, thereby improving the reliability and usability of the wind field prediction result, and providing a more reliable basis for meteorological prediction decision.

[0127] On the basis of steps 202-208, the method further comprises: obtaining a first feature enhancement variable interval and a second feature enhancement variable interval, generating Y first original enhancement variables in the first feature enhancement variable interval at random, and generating Y second original enhancement variables in the second feature enhancement variable interval at random; Y is a positive integer, the first original enhancement variable is used for feature enhancement of the past observation element embedding vector, and the second original enhancement variable is used for feature enhancement of the past cross-modal space mapping vector; Y original interaction variables are obtained by using the Y first original enhancement variables and the Y second original enhancement variables; each original interaction variable comprises a first original enhancement variable and a second original enhancement variable.

[0128] In the present application, based on the technical framework constructed in steps 202-208, this embodiment further introduces new operations to obtain original interaction variables, and this process contains multiple key links and is closely related to each other. First, the first feature enhancement variable interval and the second feature enhancement variable interval are obtained. These two intervals are key elements for determining the value range of the enhancement variable, and their determination is based on the demand of the wind field prediction data processing system for feature enhancement of the past observation element embedding vector and the past cross-modal space mapping vector. The first feature enhancement variable interval is a value range set for generating a variable for feature enhancement of the past observation element embedding vector, and the second feature enhancement variable interval is a feature enhancement variable for the past cross-modal space mapping vector. For example, the first feature enhancement variable interval can be set to [1, 10], and this interval is set considering the range of the characteristic values of the meteorological elements in the past observation element embedding vector and the degree of enhancement required in subsequent processing. For the second feature enhancement variable interval, for example, it is set to [5, 15], which is determined according to the characteristics and enhancement requirements of the past cross-modal space mapping vector.

[0129] Then, Y first original reinforcement variables are randomly generated in the first characteristic reinforcement variable interval, and Y second original reinforcement variables are randomly generated in the second characteristic reinforcement variable interval, where Y is a positive integer. Taking Y = 3 as an example, 3 first original reinforcement variables are generated in the first characteristic reinforcement variable interval [1, 10], for example, the values of the 3 variables are 3, 6, and 9 respectively. The generation of these first original reinforcement variables is random, but within the specified interval, they will be used for feature reinforcement of the past observation element embedding vector. For the second characteristic reinforcement variable interval [5, 15], 3 second original reinforcement variables are also generated, for example, the values are 7, 10, and 13. These second original reinforcement variables will act on the feature reinforcement of the past cross-modal space mapping vector.

[0130] Then, Y first original reinforcement variables are randomly generated in the first characteristic reinforcement variable interval, and Y second original reinforcement variables are randomly generated in the second characteristic reinforcement variable interval, where Y is a positive integer. Taking Y = 3 as an example, 3 first original reinforcement variables are generated in the first characteristic reinforcement variable interval [1, 10], for example, the values of the 3 variables are 3, 6, and 9 respectively. The generation of these first original reinforcement variables is random, but within the specified interval, they will be used for feature reinforcement of the past observation element embedding vector. For the second characteristic reinforcement variable interval [5, 15], 3 second original reinforcement variables are also generated, for example, the values are 7, 10, and 13. These second original reinforcement variables will act on the feature reinforcement of the past cross-modal space mapping vector.

[0131] In the wind field prediction data processing system, the past observation element embedding vector contains rich meteorological observation element information after deep processing, such as the relationship between wind speed, wind direction, temperature, pressure and other meteorological elements and their feature representation under different conditions. The first original reinforcement variable performs feature reinforcement on it, which can adjust the representation method of the relationship between these meteorological elements or highlight some key features. For example, in the past observation element embedding vector, there is a certain relationship representation between wind speed and wind direction. When the first original reinforcement variable acts on the vector, it may change the weight of this relationship representation or enhance its relevance, so that the real relationship between wind speed and wind direction can be more accurately reflected in the subsequent prediction processing.

[0132] The past cross-modal spatial mapping vector contains the relationship information of meteorological elements in different modalities (such as different spatial dimensions or different meteorological physical process related modalities). The second original reinforcement variable performs feature reinforcement on it, which can further excavate and highlight the key information in these cross-modal relationships. For example, the past cross-modal spatial mapping vector contains the cross-modal relationship between the wind speed in the vertical direction of the atmosphere and the temperature gradient in the horizontal direction. After the action of the second original reinforcement variable, some features in this relationship may be strengthened, so that this cross-modal relationship information can be more accurately transmitted when interacting with the past observation element embedding vector and other operations.

[0133] The original interaction variable combines the two different function reinforcement variables together, which is important in the entire wind field prediction processing flow. It can accurately adjust the interaction relationship between the past observation element embedding vector and the past cross-modal spatial mapping vector in subsequent vector interaction, feature optimization and other operations according to the characteristics of the first original reinforcement variable and the second original reinforcement variable in the combination, and then affect the generation of the meteorological prediction reference vector, and finally affect the wind field prediction result. For example, in the vector interaction process, if the first original reinforcement variable in the original interaction variable has performed specific reinforcement on the wind speed feature in the past observation element embedding vector, and the second original reinforcement variable has also performed reinforcement on the wind speed related cross-modal relationship in the past cross-modal spatial mapping vector, then in the interaction process, the two pieces of information about wind speed can be more accurately integrated, thereby improving the accuracy of the entire wind field prediction system for wind speed prediction.

[0134] In this way, by obtaining the first original reinforcement variable and the second original reinforcement variable in a specific interval to construct the original interaction variable, the past observation element embedding vector and the past cross-modal spatial mapping vector can be more accurately feature reinforced. This way considers the characteristics and reinforcement needs of the two vectors, combines their reinforcement variables into the original interaction variable, so that in the wind field prediction processing process, the information of the two vectors can be better integrated, and the interaction relationship between them can be accurately adjusted. Thus, the accuracy of the meteorological prediction reference vector generation is improved, and the accuracy and reliability of the wind field prediction result are further improved, providing more effective data support for meteorological prediction.

[0135] In an extensible embodiment, the number of the original interaction variables is Y, the number of the past data integration indication information is Z, Y is a positive integer, and Z is a positive integer; the number of the wind field forecast training results is Y×Z; the locating of the target interaction variable from the data assimilation thermal feature set generated by the original interaction variables and the forecast confidence feature includes: determining the initial feature enhancement thermal value corresponding to each of the Y original interaction variables based on the record information of the forecast confidence feature corresponding to each of the Z past data integration indication information under each original interaction variable; grouping the Y original interaction variables and the initial feature enhancement thermal values corresponding to the Y original interaction variables to form Y feature enhancement binary groups; each feature enhancement binary group includes an original interaction variable and the initial feature enhancement thermal value corresponding to the original interaction variable; generating a data assimilation thermal feature set according to the Y feature enhancement binary groups; if the target feature enhancement thermal value in the data assimilation thermal feature set meets the stability requirement, determining the interaction variable corresponding to the target feature enhancement thermal value in the data assimilation thermal feature set as the target interaction variable; if the target feature enhancement thermal value in the data assimilation thermal feature set does not meet the stability requirement, obtaining the interaction adjustment variable associated with the target feature enhancement thermal value from the data assimilation thermal feature set, and obtaining the forecast confidence feature of the Z past data integration indication information under the interaction adjustment variable.

[0136] Based on the extensible embodiment, the present application is developed around the complex relationship among the original interaction variable, the past data integration indication information, the wind field forecast training result, the data assimilation thermal feature set, and the target interaction variable.

[0137] First, it is clear that the number of original interaction variables is Y (Y is a positive integer), the number of past data integration indication information is Z (Z is a positive integer), and the number of wind field forecast training results is Y×Z. This quantity relationship is based on the logical architecture of the entire wind field forecast data processing system and reflects the internal relationship between different elements.

[0138] Next, the process of generating the data assimilation thermal feature set from the original interaction variables and the forecast confidence features and locating the target interaction variable from the data assimilation thermal feature set is described. Based on the record information of the forecast confidence features corresponding to the Z past data integration indication information under each original interaction variable, the initial feature enhancement thermal value corresponding to each of the Y original interaction variables is determined. For example, Y=3 and Z=4. For the first original interaction variable, the record information of the forecast confidence features obtained under the 4 past data integration indication information can be some specific numerical values or state identifiers. According to these record information, the initial feature enhancement thermal value corresponding to the first original interaction variable is determined according to a specific algorithm or logical relationship, for example, 10. Similarly, for the other two original interaction variables, their initial feature enhancement thermal values are also determined according to the record information of the forecast confidence features under the respective past data integration indication information, for example, 12 and 8 respectively.

[0139] The Y original interaction variables and the initial feature enhancement thermal values corresponding to the Y original interaction variables form Y feature enhancement binary tuples. Each feature enhancement binary tuple includes an original interaction variable and the initial feature enhancement thermal value corresponding to the original interaction variable. Taking the above example, the first feature enhancement binary tuple is (original interaction variable 1, 10), the second is (original interaction variable 2, 12), and the third is (original interaction variable 3, 8). Then, the data assimilation thermal feature set is generated according to the Y feature enhancement binary tuples. The data assimilation thermal feature set is a collection of original interaction variables and their related initial feature enhancement thermal values, which contains rich information reflecting the correlation between the original interaction variables and the forecast confidence features, and is an important basis for subsequent positioning of the target interaction variable.

[0140] After that, two cases are considered to locate the target interaction variable. If the target feature enhancement thermal value in the data assimilation thermal feature set meets the stability requirement, the interaction variable corresponding to the target feature enhancement thermal value in the data assimilation thermal feature set is determined as the target interaction variable. The stability requirement is a judgment standard set based on the overall stability and accuracy requirements of the wind field forecasting system. For example, the target feature enhancement thermal value is 12 (for example, this value is the initial feature enhancement thermal value corresponding to an element in the data assimilation thermal feature set), and if the stability requirement is set to be greater than or equal to 10, then since 12 meets this requirement, the corresponding interaction variable is determined as the target interaction variable. This target interaction variable has special importance in the entire wind field forecasting system, and it will directly affect the subsequent wind field forecasting results.

[0141] If the target feature reinforcement thermal value in the data assimilation thermal feature set does not meet the stability requirement, the interactive adjustment variable associated with the target feature reinforcement thermal value in the data assimilation thermal feature set is obtained, and the prediction confidence features of the Z past data integration indication information under the interactive adjustment variable are obtained. For example, the target feature reinforcement thermal value is 8 (which does not meet the stability requirement of, for example, greater than or equal to 10), and then the interactive adjustment variable associated with this target feature reinforcement thermal value is found in the data assimilation thermal feature set (this association is based on the internal structure and logical relationship of the data assimilation thermal feature set). Then the prediction confidence features of the Z past data integration indication information (Z = 4 in the foregoing example) under this interactive adjustment variable are obtained. These prediction confidence features will provide more information basis for further adjusting and determining the target interactive variable, so as to ensure that the finally determined target interactive variable can meet the accuracy and stability requirements of the wind field prediction system.

[0142] It can be seen that by determining the initial feature reinforcement thermal value based on the prediction confidence features corresponding to the original interactive variable and the past data integration indication information, and then generating the data assimilation thermal feature set, the key information in the system is comprehensively integrated. In positioning the target interactive variable, different operations are performed according to whether the target feature reinforcement thermal value meets the stability requirement, which can more accurately screen out suitable target interactive variables. When the stability requirement is met, the target interactive variable is directly determined, and when it is not met, the associated interactive adjustment variable and its corresponding prediction confidence features are further adjusted, which helps to improve the accuracy of the determination of the target interactive variable, thereby improving the accuracy and stability of the entire wind field prediction system and providing a more reliable basis for wind field prediction.

[0143] In summary, the present application improves the performance of the entire wind field prediction system by means of the mesoscale weather model WRF and the WRFDA data assimilation. The WRFDA data assimilation system can assimilate various types of observation data, and this process is a key link in integrating multi-source meteorological observation information. Just like a precise filter and integrator, it filters and fuses meteorological observation data from different sources (such as ground meteorological stations, satellites, etc.), and integrates data from multiple sources into a unified framework, thereby improving the accuracy of the initial field. This lays a solid foundation for subsequent numerical simulation calculations based on the WRF model.

[0144] In the wind field prediction processing network, the integration of past data indication information and the collaborative processing of multi-source meteorological observation information have unique advantages. Through observation element mining of past multi-source meteorological observation information, past observation element embedding vectors are obtained. The feature pooling and vector mining operations in them are like precisely mining treasures in a vast meteorological data treasure trove. Feature pooling can summarize similar feature data, and vector mining can deeply mine the potential relationship between meteorological elements, which helps better adapt to the advanced dynamic framework and physical process parameterization scheme of the WRF model, thereby improving the accuracy of the description of the wind field state.

[0145] The use of past data integration indication information for cross-modal spatial mapping to obtain past cross-modal spatial mapping vectors adds rich depth information to wind field prediction. This is similar to the atmospheric simulation environment constructed by the WRF model, where not only can the current meteorological element state be seen, but also the derived changes of meteorological elements and future development trends can be predicted based on the physical process parameterization scheme. For example, under the WRF model, through meteorological element derivation, key information such as energy changes of the wind field under different physical processes can be obtained, and meteorological trend extrapolation can more accurately predict the development trend of the wind field in the mesoscale range, which plays an irreplaceable role in improving the accuracy of wind field prediction and the timeliness of meteorological disaster warning.

[0146] The process of the original interaction variable interacting with the past observation element embedding vector and the past cross-modal spatial mapping vector and optimizing the features to obtain the meteorological prediction reference vector is a further integration of the advantages of the WRF model and WRFDA data assimilation. The original interaction variable is like an intelligent adjustment hub that adjusts the vector based on the physical process relationship in the WRF model and the characteristics of the multi-source data after WRFDA data assimilation. This adjustment ensures that the meteorological prediction reference vector can fully reflect the true state of the wind field, maximally reduces the prediction error caused by complex data sources or variable meteorological processes, and makes the prediction result more consistent with the actual atmospheric conditions simulated by the WRF model.

[0147] Through discriminant analysis of the meteorological prediction reference vector by the initial wind field prediction processing network, the wind field prediction training result is obtained, and the prediction confidence feature is obtained. This step is similar to the final quality inspection process based on WRF model simulation. The prediction confidence feature provides a reliable reliability evaluation for the wind field prediction result, which is crucial in meteorological prediction and disaster warning application scenarios. The reliability of the high-resolution mesoscale weather prediction result based on the WRF model output depends on it, and decision-makers can use the prediction information more scientifically based on this reliability evaluation, thereby avoiding making wrong decisions due to unreliable predictions.

[0148] Finally, the wind field prediction result is generated by generating the assimilation thermal feature set through the original interaction variable and the prediction confidence feature generation data, and positioning the target interaction variable to determine the target wind field prediction processing network. This process is closely combined with the WRF and WRFDA data assimilation system. In the high-resolution simulation environment constructed by the WRF model, the generation of the data assimilation thermal feature set can make full use of the multi-source data information after the data assimilation of WRFDA, and the positioning of the target interaction variable is to determine the most suitable prediction processing network for the current weather condition according to the dynamic framework and physical process characteristics in the WRF model. This way helps to further improve the accuracy and timeliness of wind field prediction, and can output more accurate and detailed wind field prediction results in the mesoscale range, thereby providing more powerful meteorological data support for meteorological related fields such as meteorological prediction and disaster warning, and meeting the urgent needs of these fields for high-precision meteorological prediction.

[0149] Further, Figure 2 A structural schematic diagram of a wind field prediction data processing system 200 provided by an embodiment of the present application is shown in FIG. 1. Figure 2 The wind field prediction data processing system 200 shown in FIG. 1 includes a processor 210. The processor 210 can call and run a computer program from a memory to implement the method in the embodiment of the present application.

[0150] Optionally, as shown in FIG. 1, the wind field prediction data processing system 200 can further include a memory 230. The processor 210 can call and run a computer program from the memory 230 to implement the method in the embodiment of the present application. Figure 2 The memory 230 can be a separate device independent of the processor 210, or can be integrated in the processor 210.

[0151] Optionally, as shown in FIG. 1, the wind field prediction data processing system 200 can further include a transceiver 220. The processor 210 can control the transceiver 220 to interact with other devices, specifically, can send information or data to other devices, or receive information or data sent by other devices.

[0152] Figure 2 Optionally, the wind field prediction data processing system 200 can implement the corresponding processes of the storage engine or the components (such as processing modules) in the storage engine or the devices deployed with the storage engine in the various methods of the embodiments of the present application. For the sake of brevity, they will not be described here.

[0153] It should be understood that the processor of the embodiments of the present application can be an integrated circuit chip with signal processing capability.

[0154] It should be understood that the processor of the embodiments of the present application can be an integrated circuit chip with signal processing capability.

[0155] ​It is to be understood that the memory in the embodiments of the present application can be a volatile memory or a nonvolatile memory, or can include both volatile and nonvolatile memory. It should be noted that the memory of the system and method described herein is intended to include, but not be limited to, the appropriate type of memory suitable for the described use.

[0156] On the basis of the above, a readable storage medium is provided, the readable storage medium stores a program or instructions, the program or instructions are executed by a processor to realize the steps of the above method.

[0157] It should be noted that in this paper, the term "including", "containing" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or device. Without more limitations, the element defined by the sentence "including a…" does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiments of the present application is not limited to the order of functions shown or discussed, but can also include functions performed in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method can be performed in an order different from that described, and various steps can also be added, omitted or combined. In addition, the features described with reference to certain examples can be combined in other examples.

[0158] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), including a plurality of instructions for making a terminal (which can be a mobile phone, computer, server, or network equipment, etc.) execute the method described in each embodiment of the present application.

[0159] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above specific embodiments, the above specific embodiments are only illustrative, not limiting, and those skilled in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope of the present application.

Claims

1. A deep learning-based method for processing wind field forecast data, characterized in that, The method is applied to a wind field forecast data processing system, and the method includes: Acquire past multi-source meteorological observation information and integrate past data to provide guidance; In the initial wind field forecasting processing network, the historical data integration indication information is used to mine the observation elements of the historical multi-source meteorological observation information to obtain the historical observation element embedding vector; the observation element mining is used for feature pooling and vector mining. By integrating the information from the past data, the embedding vectors of the past observation elements are mapped across modal space to obtain the past cross-modal space mapping vector. The original interaction variables are used to perform vector interaction and feature optimization on the embedding vectors of the past observation elements and the past cross-modal space mapping vector to obtain the weather forecast reference vector. include: When u is a positive integer greater than 1 and not greater than X, in the u-th spatial mapping kernel of the initial wind field forecasting processing network, the past data integration indication information is used to perform cross-modal spatial mapping on the (u-1)-th meteorological forecasting reference vector to obtain the u-th past cross-modal spatial mapping vector. The (X-u+1)-th past observation element embedding vector generated by the (X-u+1)-th element embedding kernel is obtained. Using the u-th original interaction variable in the u-th spatial mapping kernel, the (X-u+1)-th past observation element embedding vector and the u-th past cross-modal spatial mapping vector are used to perform vector interaction to obtain the u-th past interaction vector. If u is X, the X-th past interaction vector is feature optimized to obtain the meteorological forecasting reference vector; X is the number of spatial mapping kernels included in the initial wind field forecasting processing network. When u is 1, in the first spatial mapping kernel of the initial wind field forecasting processing network, the cross-modal spatial mapping of the Xth past observation element embedding vector is performed using the past data integration indication information to obtain the first past cross-modal spatial mapping vector. The first original interaction variable is used to perform vector interaction between the Xth past observation element embedding vector and the first past cross-modal spatial mapping vector to obtain the first past interaction vector. The cross-modal spatial mapping is used for meteorological element derivation and meteorological trend inference; the initial wind field forecast processing network is a convolutional neural network with adjustable kernel parameters, which includes convolutional layers for extracting local features from multi-source meteorological observation information and pooling layers for compression and simplification. The initial wind field forecasting processing network is used to perform discriminant analysis on the meteorological forecasting reference vector to obtain the wind field forecasting training results and to acquire the forecast confidence features of the wind field forecasting training results. A data assimilation thermodynamic feature set is generated by combining the original interaction variables with the forecast confidence features. The target interaction variable is located from the data assimilation thermodynamic feature set, and the initial wind field forecasting processing network including the target interaction variable is determined as the target wind field forecasting processing network. The target wind field forecasting processing network is used to generate wind field forecast results based on the current data integration indication information.

2. The method as described in claim 1, characterized in that, The acquisition of past multi-source meteorological observation information and past data integration indication information includes: Obtain the initial wind field forecast processing network, obtain the target forecast training annotations and the deep learning network debugging set corresponding to the initial wind field forecast processing network; the initial wind field forecast processing network is a deep learning network that has reached the pre-debugging conditions; The past multi-source meteorological observation information containing the observation elements corresponding to the target forecast training labels is obtained from the deep learning network debugging set, and the past data integration indication information is generated based on the target forecast training labels.

3. The method as described in claim 2, characterized in that, The step of obtaining past multi-source meteorological observation information containing the observation elements corresponding to the target forecast training annotations from the deep learning network debugging set includes: Identify the past forecast training labels of the multi-source observation data examples included in the deep learning network debug set, obtain the label feature common value between the past forecast training labels of the multi-source observation data examples and the target forecast training labels, and determine the multi-source observation data examples corresponding to the past forecast training labels whose label feature common value is not less than the forecast label common threshold as past multi-source meteorological observation information; Alternatively, based on the target forecast training annotation, the selected wind field to which the wind field forecast event is targeted is determined, and the selected wind field is parsed from the multi-source observation data examples included in the deep learning network debug set. The parsed multi-source observation data examples of the selected wind field are then determined as the past multi-source meteorological observation information.

4. The method as described in claim 2, characterized in that, The step of obtaining past multi-source meteorological observation information containing the observation elements corresponding to the target forecast training annotations from the deep learning network debugging set includes: Obtain examples of original meteorological observation information containing the observation elements corresponding to the target forecast training annotations from the deep learning network debugging set, and identify the examples of original meteorological observation information as past multi-source meteorological observation information; Alternatively, obtain examples of original meteorological observation information containing the observation elements corresponding to the target forecast training annotations from the deep learning network debugging set, preprocess the examples of original meteorological observation information, and obtain the past multi-source meteorological observation information; Alternatively, obtain examples of original meteorological observation information containing the observation elements corresponding to the target forecast training labels from the deep learning network debugging set, obtain the observation status labels of the original meteorological observation information examples, and determine the past multi-source meteorological observation information from the original meteorological observation information examples based on the observation status labels.

5. The method as described in claim 1, characterized in that, The embedded vector of the past observation elements includes the dynamic framework vector of the observation elements, the parameterized vector of the physical process, and the vector of the mesoscale weather model. In the initial wind field forecasting processing network, the historical data integration indication information is used to mine the observation elements of the historical multi-source meteorological observation information, resulting in the embedded vector of historical observation elements, including: In the initial wind field forecasting processing network, dynamic frame feature mapping is performed on the multi-source meteorological observation element vectors corresponding to the past multi-source meteorological observation information to obtain the dynamic frame vectors of the observation elements; the dynamic frame feature mapping is used to enhance the observation elements under the dynamic frame state. The dynamic framework vector of the observed elements is quantized to obtain the parameterized vector of the physical process; the parameterized vector of the physical process is used to represent the state of the meteorological physical process corresponding to the past multi-source meteorological observation information. Obtain the historical integration indication vector corresponding to the historical data integration indication information, and perform weighted integration of the physical process parameterization vector and the historical integration indication vector to obtain the mesoscale weather model vector; the mesoscale weather model vector is used to represent the meteorological joint description features of the historical multi-source meteorological observation information and the historical data integration indication information.

6. The method as described in claim 5, characterized in that, The method further includes: The past multi-source meteorological observation information is convolved to obtain the past meteorological observation convolution vector. Based on the vector size of the convolution vector of the past meteorological observations, generate past observation offset features; The convolution vector of the past meteorological observations and the offset features of the past observations are weighted and integrated to obtain the vector of the multi-source meteorological observation elements.

7. The method as described in claim 1, characterized in that, The method of obtaining the forecast confidence features of the wind field forecast training results includes: Obtain the relationship influence coefficient between the wind field forecast training results and the historical data integration indication information, and obtain the forecast inversion label of the wind field forecast training results; The forecast viewpoint evaluation network is used to evaluate the wind field forecast training results to obtain the forecast accuracy characteristics of the wind field forecast training results; By combining the relationship influence coefficient, the forecast inversion label, and the forecast accuracy feature, the forecast confidence feature of the wind field forecast training result is obtained; The method further includes: Obtain the quantization discrimination interval, obtain the network debugging expectation index, and obtain the global discrimination score from the quantization discrimination interval based on the network debugging expectation index; Based on the global discrimination score, a target discrimination score is determined, and the target discrimination score is converted into an inversion matching score and a prediction accuracy score. The prediction confidence features of the wind field prediction training results are obtained by combining the relationship influence coefficient, the prediction inversion label, and the prediction accuracy features, including: By utilizing the global discrimination score, the inversion matching score, and the forecast accuracy score, combined with the relationship influence coefficient, the forecast inversion label, and the forecast accuracy feature, the forecast confidence feature of the wind field forecast training result is obtained.

8. The method as described in claim 1, characterized in that, The method further includes: Obtain a first feature enhancement variable interval and a second feature enhancement variable interval. Generate Y first original enhancement variables arbitrarily within the first feature enhancement variable interval, and generate Y second original enhancement variables arbitrarily within the second feature enhancement variable interval; Y is a positive integer. The first original enhancement variables are used to enhance the features of the embedded vector of the past observation elements, and the second original enhancement variables are used to enhance the features of the past cross-modal space mapping vector. Using the Y first original reinforcement variables and the Y second original reinforcement variables, Y original interaction variables are obtained; each original interaction variable includes one first original reinforcement variable and one second original reinforcement variable.

9. A wind field forecast data processing system, characterized in that, The method includes at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, causing the at least one processor to perform the method according to any one of claims 1-8.

Images