Cold season air temperature intelligent forecasting method
By using multi-source collaborative observation of '1 station and 2 networks' and an AI-powered intelligent temperature forecasting model based on the Transformer architecture, the problem of low accuracy in cold season temperature forecasts in Yunnan Province has been solved, improving the accuracy of temperature forecasts in eastern Yunnan and supporting refined weather warnings.
Patent Information
- Application Number
- CN202511145597.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-15
- Publication Date
- 2026-01-20
AI Technical Summary
In the current technology, the accuracy of cold season temperature forecasts in Yunnan Province is low, especially the accuracy of operational forecasts for minimum and maximum temperatures.
By constructing a YGQF collaborative observation dataset based on multi-source collaborative observation of '1 station and 2 networks', calculating long-sequence YGQF location and intensity data, constructing a YGQF label dataset, and integrating feature parameters and topographic parameters, a temperature intelligent forecasting model based on AI technology was developed using the Transformer architecture, and a temperature forecasting module was constructed and connected to the test platform.
It has improved the accuracy of cold season temperature forecasts in eastern Yunnan by 3-10%, and enhanced the ability of refined and intelligent weather forecasting and early warning.
Smart Images
Figure CN121364514A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of temperature forecasting technology, and in particular to a smart method for forecasting temperatures during the cold season. Background Technology
[0002] YGQF is an abbreviation for the Yunnan-Guizhou Quasi-stationary Front, which is an important weather system that appears in the Yunnan-Guizhou Plateau and its adjacent areas during the winter half of the year (mainly late autumn, winter and early spring, from about November to April of the following year).
[0003] Currently, according to data from the Yunnan Provincial Meteorological Observatory's "Yunnan Numerical Model Interpretation and Verification System," the average accuracy rate of Yunnan's minimum temperature operational forecast (≤1℃) (subjective 08_model 08 start date) from November 1, 2023 to April 30, 2024 is 0.42, and the average accuracy rate of the maximum temperature operational forecast (≤1℃) (subjective 08_model 08 start date) is 0.34, indicating low forecast accuracy. Summary of the Invention
[0004] Embodiments of the present invention provide a method for intelligent temperature forecasting in the cold season, aiming to solve the problem of accurate temperature forecasting. The problem of low efficiency.
[0005] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions: A method for intelligent temperature forecasting during the cold season includes the following steps: S1: Conduct YGQF collaborative observations based on "1 station and 2 networks" and construct a collaborative observation dataset; S2: Calculate the YGQF location and intensity data of the long sequence using the collaborative observation dataset, and construct a YGQF label dataset; S3: Match the feature parameters and terrain parameters with the YGQF label dataset to construct the full YGQF dataset; S4: Using the full YGQF dataset, develop an AI-based intelligent temperature forecasting model based on the Transformer architecture; S5: Using the aforementioned intelligent temperature forecasting model, construct a temperature forecasting module based on AI technology; S6: Connect the temperature forecast module to the test platform for trial use.
[0006] Furthermore, in S1, the "1 station, 2 networks" includes: The Mile YGQF core test base and data processing center, the local topographic impact observation network built along the main path of cold air activity, and the joint observation network in eastern Yunnan.
[0007] Furthermore, in S1, the YGQF cooperative observation includes: After data quality control and processing, the data sets of ground routine weather observation, new vertical observation, valley gradient observation, mountain profile observation, aircraft onboard equipment and weather radar collaborative observation are formed.
[0008] Further, in S2, the YGQF position and intensity data of the long sequence include: The longitude and latitude of each time front line grid and the potential temperature gradient in 1 latitude distance, the daily average temperature, the daily maximum temperature and the daily minimum temperature of the matching front line grid position (positive sample) and the east and west two sides of the grid points (negative sample) within 2 latitude distances.
[0009] Further, in S3, the characteristic parameters include: weather background, terrain, underlying surface, obvious diurnal variation, intra-seasonal variation and inter-annual variation characteristics.
[0010] Further, in S3, the terrain parameters include: the birth and death of the front, the terrain, the ground pressure, the low-level wind field, the intensity of the cold and warm air masses on both sides of the front, the ground radiation and the cloud amount.
[0011] Further, in S4, the Transformer architecture includes an attention mechanism (Attention Mechanism) and a feedforward neural network, and the attention mechanism (Attention Mechanism) calculates the overall correlation of T and the characteristic quantity sequence through inner product, and the formula is:
[0012] The attention mechanism (Attention Mechanism) calculates the spatio-temporal correlation of T and the characteristic quantity through outer product, and the formula is:
[0013] T is the spatio-temporal sequence of air temperature, Δp is the sequence of characteristic quantities, and the one inner product and outer product result obtains one attention, and the multiple inner product and outer product results obtain multiple attentions, that is, multi-head attention mechanism (Multi-head attention).
[0014] Further, the multi-head attention mechanism (Multi-head attention) calculates the correlation of the characteristic parameters in the YGQF full data set and the T sequence, and obtains an intelligent air temperature prediction model through a Transformer fully connected feedforward neural network.
[0015] Beneficial effects: The present application carries out multi-source collaborative observation through "1 station 2 networks", obtains comprehensive and high-quality observation data, and provides a solid data foundation for subsequent model training and prediction. By constructing the YGQF label data set and the YGQF full-quantity data set after integrating the characteristic parameters and the terrain parameters, the full-quantity data set contains rich factors affecting the change of air temperature, which helps to improve the model's ability to capture the change rule of air temperature.
[0016] The AI air temperature intelligent prediction model based on the Transformer architecture can effectively calculate the correlation between the characteristic parameters and the air temperature spatiotemporal sequence through the multi-head attention mechanism, fully excavate the complex relationships in the data, and improve the cold-season air temperature prediction accuracy in the eastern region of Yunnan by 3-10% on the basis of the existing technology, thereby supporting the improvement of the fine and intelligent weather prediction and early warning capability in Yunnan. BRIEF DESCRIPTION OF DRAWINGS
[0017] Figure 1 It is a flowchart of the present application; Figure 2 It is a mountain valley terrain gradient observation station distribution map of the present application; Figure 3 It is a mountain terrain profile observation station distribution of the present application; Figure 4 It is a Yunnan radar ground 1-kilometer coverage map of the present application; Figure 5 It is a schematic diagram of the attention mechanism in the Transformer architecture; DETAILED DESCRIPTION
[0018] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.
[0019] In the description of the present application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the devices or elements indicated must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.
[0020] The terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "multiple" is two or more.
[0021] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection, it can be mechanical connection, or electrical connection, it can be direct connection, or indirect connection through intermediate medium, or internal communication of two elements, the specific meaning of the above terms in the present application can be understood according to the specific circumstances by those skilled in the art.
[0022] The present application provides a cold season air temperature intelligent prediction method, comprising the following steps: S1: based on "1 station 2 network", YGQF collaborative observation is carried out, and a collaborative observation data set is constructed; S2: the long sequence YGQF position and intensity data are calculated through the collaborative observation data set, and the YGQF label data set is constructed; S3: the characteristic parameters and terrain parameters are matched with the YGQF label data set, and the YGQF full data set is constructed; S4: using the YGQF full data set, an air temperature intelligent prediction model based on AI technology is developed based on the Transformer architecture; S5: using the air temperature intelligent prediction model, an air temperature prediction module based on AI technology is constructed; S6: the air temperature prediction module is connected to the test platform for trial use.
[0023] The present application provides a cold season air temperature intelligent prediction method, the present application carries out multi-source collaborative observation through "1 station 2 network", obtains comprehensive and high-quality observation data, and provides a solid data foundation for subsequent model training and prediction. Through the construction of YGQF label data set and YGQF full data set after the fusion of characteristic parameters and terrain parameters, the full data set contains rich factors affecting the change of air temperature, which helps to improve the model's ability to capture the change rule of air temperature. The AI air temperature intelligent prediction model based on the Transformer architecture can effectively calculate the correlation between the characteristic parameters and the air temperature spatiotemporal sequence through the multi-head attention mechanism, fully excavate the complex relationship in the data, and the cold season air temperature prediction accuracy in the eastern region of Yunnan is improved by 3-10% on the basis of the existing, supporting the improvement of the fine intelligent weather prediction and warning capability in Yunnan.
[0024] In the embodiments of the present application, in order to deeply study the evolution law of YGQF at different time scales and its mechanism, establish an objective prediction method and system, and improve the prediction accuracy, the "1 station 2 network" includes: the Melle YGQF core test base and data processing center, the local terrain influence observation network constructed on the main path of cold air activity, and the joint observation network in the east of Yunnan. The Melle scientific test base has been built into a comprehensive meteorological station integrating observation test, business scientific research and meteorological popularization, and was approved by the joint document of Yunnan Meteorological Bureau and the Institute of Atmospheric Sciences in April 2023. The base is built according to the standard of national benchmark meteorological station, and new detection equipment such as microwave radiometer, millimeter wave cloud radar, wind profile radar and meteorological multi-element long-haul unmanned aerial vehicle is added according to the requirements of long-term observation, as well as large-scale evaporation, radiation, soil moisture, atmospheric electric field and other observation projects. In addition, the mountain valley gradient observation system( Figure 2 ) and the mountain profile observation system( Figure 3 ) located in the east of Melle have also been completed, which can monitor the changes of various meteorological elements with high temporal and spatial resolution when the cold air and YGQF advance and retreat in the typical local terrain area. At present, more than 4200 ground conventional meteorological observation stations with an average spacing of 9.5 kilometers, 5 sounding stations, 12 C-band weather radars, 16 X-band weather radars and 1 S-band weather radar( Figure 4 ) have been built, and the yellow represents C-band radar, the green represents S-band radar, and the blue represents X-band dual-polarization radar, and the multi-band dual-polarization weather radar network of the whole province has been realized.
[0025] In the embodiments of the present application, the YGQF cooperative observation includes: after data quality control and arrangement, the data set of ground conventional meteorological observation, new vertical observation, mountain valley gradient observation, mountain profile observation, aircraft onboard equipment and weather radar cooperative observation is formed. High temporal and spatial resolution and fine meteorological observation data in the process of frontogenesis and frontolysis and advance and retreat are obtained.
[0026] In the embodiments of the present application, the long sequence of YGQF position and intensity data includes: the longitude and latitude of each time front line grid point and the potential temperature gradient in 1 latitude distance, the daily average temperature, the daily maximum temperature and the daily minimum temperature of the matching front line grid point position (positive sample) and the 2 latitude points on the east and west sides of the grid point (negative sample).
[0027] YGQF is located below 700 hPa and is a shallow circulation system with significant multi-time scale variation characteristics. In the embodiments of the present application, the characteristic parameters include: weather background, terrain, underlying surface, obvious diurnal variation, intraseasonal variation and interannual variation characteristics.
[0028] In the embodiments of the present application, the terrain parameters include the birth and death of the front, the terrain, the ground pressure, the low-altitude wind field, the intensity of the cold and warm air masses on both sides of the front, the ground radiation and the cloud amount.
[0029] With the advent of the era of large models, the popularity of the Transformer architecture underlying large models has exceeded that of CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network). The Transformer architecture discards traditional CNN and RNN neural networks. In the embodiments of the present application, the Transformer architecture includes an attention mechanism (Attention Mechanism) and a feedforward neural network. The attention mechanism (Attention Mechanism) calculates the overall correlation of T and the feature quantity sequence through an inner product, and the formula is as follows:
[0030] The attention mechanism (Attention Mechanism) calculates the spatio-temporal correlation of T and the feature quantity through an outer product, and the formula is as follows:
[0031] T is a spatio-temporal sequence of air temperature, and Δp is a sequence of feature quantities. One inner product and outer product result obtains one attention, and multiple inner product and outer product results obtain multiple attentions, i.e., a multi-head attention mechanism (Multi-head attention). The attention mechanism (Attention Mechanism) improves the efficiency and accuracy of the model by calculating a set of importance weights to allocate limited computing resources to more important information.
[0032] In the embodiments of the present application, the multi-head attention mechanism (Multi-head attention) calculates the correlation of the feature parameters in the YGQF full data set and the T sequence, and obtains the air temperature intelligent prediction model (Y) through a Transformer full connection feedforward neural network. Figure 5 ) through training.
[0033] The specific steps of the present application are as follows: Step S1: Construct a collaborative observation data set. The observation system of '1 station 2 networks' is formed by the YGQF core test base and data processing center, the local terrain influence observation network constructed along the main path of cold air activity, and the joint observation network in the east of Yunnan. Through the observation system, the YGQF coordinated observation is carried out, which specifically includes collecting conventional meteorological data such as temperature, humidity, air pressure, and wind speed by ground conventional meteorological observation instruments; obtaining vertical distribution data of meteorological elements at different heights by using new vertical observation equipment; collecting meteorological data at different altitudes by setting gradient observation stations in valley areas, forming mountain valley gradient observation data; obtaining meteorological information at different positions along the mountain profile, forming mountain profile observation data; at the same time, air observation is carried out by using airborne equipment, and combined with weather radar data, multi-dimensional and multi-device coordinated observation is realized. All observation data are strictly quality controlled, and abnormal values and missing values are removed and filled, and after arrangement, the coordinated observation data set is formed.
[0034] Step S2: Constructing YGQF label data set; Based on the coordinated observation data set obtained in step S1, the long sequence YGQF position and intensity data are calculated. For each time, the longitude and latitude of the frontal grid point are determined, and the potential temperature gradient in 1 latitude-longitude distance is calculated. The frontal grid point position is selected as the positive sample, and the grid points within 2 latitudes on both sides are selected as the negative sample. The daily average temperature, daily maximum temperature and daily minimum temperature of these samples are matched to construct the YGQF label data set.
[0035] Step S3: Constructing YGQF full-amount data set; Collect feature parameters such as weather background (such as atmospheric circulation situation), terrain (such as mountain direction, altitude, etc.), underlying surface (such as vegetation type, soil property, etc.), obvious daily variation, intraseasonal variation and interannual variation characteristics, etc. At the same time, collect the relationship between the birth, death, advance and retreat of the front and the terrain, ground pressure, low-level wind field data, the intensity of cold and warm air masses on both sides of the front, ground radiation and cloud amount, etc. The above feature parameters and terrain parameters are matched with the YGQF label data set constructed in step S2 to ensure that each label corresponds to the corresponding feature and terrain information, forming the YGQF full-amount data set.
[0036] Step S4: Developing an intelligent air temperature prediction model based on AI technology; Using the YGQF full data set obtained in step S3, a model is developed based on the Transformer architecture. Set T as the spatiotemporal sequence of air temperature, and Δp as the characteristic quantity sequence. A plurality of attentions are obtained by calculating the inner product and the outer product to form a multi-head attention mechanism. The correlation between the characteristic parameters in the full data set and the T sequence is calculated using the multi-head attention mechanism. The results are input into the Transformer fully connected feedforward neural network. Appropriate training parameters (such as learning rate, number of iterations, etc.) are set, and the model is trained to obtain an air temperature intelligent prediction model.
[0037] Step S5: Building an air temperature prediction module based on AI technology; Based on the air temperature intelligent prediction model developed in step S4, an air temperature prediction module based on AI technology is built. This module can receive input data and output corresponding air temperature prediction results.
[0038] Step S6: Accessing a test platform for trial use; The air temperature prediction module built in step S5 is connected to the meteorological test platform for trial use in the actual application environment. The prediction results and actual observation data during the trial use are collected to evaluate and optimize the performance of the module, continuously improving its prediction accuracy and stability.
[0039] In the description of the present specification, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.
[0040] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. A cold season air temperature intelligent forecasting method, characterized in that, It comprises the following steps: S1: YGQF collaborative observation is carried out based on "1 station 2 networks", and a collaborative observation dataset is constructed; S2: long sequence YGQF position and intensity data are calculated through the collaborative observation dataset, and a YGQF label dataset is constructed; S3: feature parameters and terrain parameters are matched with the YGQF label dataset, and a YGQF full quantity dataset is constructed; S4: an AI technology-based air temperature intelligent prediction model is developed based on the Transformer architecture using the YGQF full quantity dataset; S5: an AI technology-based air temperature prediction module is constructed using the air temperature intelligent prediction model; S6: the air temperature prediction module is connected to a test platform for trial use.
2. A cold season air temperature intelligent forecasting method according to claim 1, characterized in that, In S1, the "1 station 2 networks" comprise: the Mile YGQF core test base and data processing center, and the local terrain influence observation network and the eastern Yunnan joint observation network constructed along the main path of cold air activity.
3. The cold-season air temperature intelligent forecasting method of claim 1, wherein, In S1, the YGQF collaborative observation comprises: After data quality control and arrangement, a dataset of ground routine weather observation, new vertical observation, valley gradient observation, mountain profile observation, aircraft onboard equipment and weather radar collaborative observation is formed.
4. The cold-season air temperature intelligent forecasting method of claim 1, wherein, In S2, the long sequence YGQF position and intensity data comprise: the longitude and latitude of each time front line grid and the potential temperature gradient in 1 latitude distance, matching the daily average temperature, daily maximum temperature and daily minimum temperature of the front line grid position (positive sample) and the east and west two sides of each 2 latitude grid points (negative sample).
5. The method of claim 1, wherein, In S3, the feature parameters comprise: weather background, terrain, underlying surface, obvious daily variation, intraseasonal variation and interannual variation characteristics.
6. The method of claim 1, wherein, In S3, the terrain parameters comprise: the birth, death, advance and retreat of the front, ground pressure, low-level wind field, cold and warm air mass intensity on both sides of the front, ground radiation and cloud cover.
7. The method of claim 1, wherein, In S4, the Transformer architecture comprises an attention mechanism (Attention Mechanism) and a feedforward neural network, the attention mechanism (Attention Mechanism) calculates the overall correlation of T and the feature quantity sequence through inner product, the formula is: the attention mechanism (Attention Mechanism) calculates the spatio-temporal correlation of T and the feature quantity through outer product, the formula is: T is the spatio-temporal sequence of air temperature, Δp is the sequence of feature quantities, the results of the inner product and outer product are obtained, and one attention is obtained, the results of multiple inner products and outer products are obtained, and multiple attentions are obtained, that is, a multi-head attention mechanism (Multi-head attention).
8. A cold-season air temperature intelligent forecasting method according to claim 7, characterized in that, The multi-head attention mechanism (Multi-head attention) calculates the correlation of the feature parameters in the YGQF full quantity dataset and the T sequence, and the air temperature intelligent prediction model is obtained through the Transformer full connection feedforward neural network.