Low-altitude flow field intelligent deduction method based on physical AI

Abstract

The invention provides a physical AI-based low-altitude flow field intelligent deduction method, and relates to the technical field of low-altitude flow field deduction, and the method comprises the steps: obtaining scene geometric information, obstacle information, boundary conditions and sparse observation data, and constructing input features; historical scene geometric information, historical obstacle information, historical boundary conditions and historical sparse observation data are obtained; determining a historical actual wind speed flow field state and a historical actual wind direction flow field state; obtaining a historical sample wind speed flow field state and a historical sample wind direction flow field state; determining a training loss function of the low-altitude flow field intelligent deduction model; obtaining a trained low-altitude flow field intelligent deduction model; obtaining an intelligent deduction result of the low-altitude flow field; and issuing, caching or periodically updating the low-altitude flow field intelligent deduction result through a low-delay output and distribution mechanism, and integrating a rollback strategy to process an abnormal data stream. According to the invention, the accuracy of low-altitude flow field intelligent deduction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of low-altitude flow field deduction, and particularly relates to a low-altitude flow field intelligent deduction method based on physical AI. BACKGROUND

[0002] In the related art, low-altitude flow field intelligent deduction can be performed based on a traditional numerical weather model, but the related art cannot ensure that the deduced flow field strictly follows physical laws, and cannot guarantee the timeliness of the deduction result.

[0003] The information disclosed in the background section of this application is only intended to deepen the understanding of the general background of the application and should not be regarded as acknowledging or implying in any form that this information constitutes prior art known to those skilled in the art. SUMMARY

[0004] The present application provides a low-altitude flow field intelligent deduction method based on physical AI, which can solve the technical problem that the related art cannot ensure that the deduced flow field strictly follows physical laws, and cannot guarantee the timeliness of the deduction result.

[0005] According to a first aspect of the present application, a low-altitude flow field intelligent deduction method based on physical AI is provided, comprising:

[0006] Obtaining scene geometry information, obstacle information, boundary conditions and sparse observation data at multiple time points of a current time period, and constructing input features through a sliding window or prior knowledge;

[0007] Obtaining historical scene geometry information, historical obstacle information, historical boundary conditions and historical sparse observation data at multiple time points of multiple historical time periods;

[0008] Determining historical actual wind speed flow field states and historical actual wind direction flow field states at multiple historical prediction grid points of multiple historical prediction time points corresponding to the multiple historical time periods, wherein the historical actual wind speed flow field states and the historical actual wind direction flow field states are represented as continuous vectors or discrete classification distributions;

[0009] Processing the historical scene geometry information, the historical obstacle information, the historical boundary conditions and the historical sparse observation data through a low-altitude flow field intelligent deduction model to obtain historical sample wind speed flow field states and historical sample wind direction flow field states at multiple historical prediction grid points of multiple historical prediction time points;

[0010] determine a training loss function of the low-altitude flow field intelligent deduction model according to the physical consistency loss term, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state, and the historical sample wind direction flow field state, wherein the physical consistency loss term includes but is not limited to observation consistency, boundary consistency, geometric consistency, or constraints based on fluid mechanics laws;

[0011] train the low-altitude flow field intelligent deduction model according to the training loss function of the low-altitude flow field intelligent deduction model, and obtain a trained low-altitude flow field intelligent deduction model;

[0012] process the input feature according to the trained low-altitude flow field intelligent deduction model, and obtain a low-altitude flow field intelligent deduction result, wherein the low-altitude flow field intelligent deduction result includes a three-dimensional vector wind field and supports continuous value output or classification distribution output;

[0013] publish, cache, or periodically update the low-altitude flow field intelligent deduction result through a low-latency output and distribution mechanism, and integrate a fallback strategy to process abnormal data flow.

[0014] According to the present application, the low-altitude flow field intelligent deduction model adopts hierarchical reasoning, including:

[0015] generate an initial wind field according to the sparse observation data and the boundary condition;

[0016] generate a global high-resolution wind field according to the initial wind field, the scene geometric information, and the obstacle information.

[0017] According to the present application, generating an initial wind field according to the sparse observation data and the boundary condition includes:

[0018] process the sparse observation data according to a multi-source observation data assimilation technology to obtain processed sparse observation data;

[0019] generate an initial wind field according to the processed sparse observation data and the boundary condition.

[0020] According to the present application, determining the historical actual wind speed flow field state and the historical actual wind direction flow field state at a plurality of historical prediction grid points corresponding to a plurality of historical prediction time points of a plurality of historical time periods includes:

[0021] obtain historical multi-source observation data of a historical prediction time point corresponding to a historical time period;

[0022] generate a historical actual three-dimensional flow field of the historical prediction time point according to the historical multi-source observation data;

[0023] According to the historical actual three-dimensional flow field, a historical actual wind speed flow field state and a historical actual wind direction flow field state are determined.

[0024] According to the physical consistency loss term, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state, and the historical sample wind direction flow field state, a training loss function of the low-altitude flow field intelligent inference model is determined.

[0025] Obtaining inference task information and historical prediction grid point information of historical prediction grid points;

[0026] According to the inference task information and the historical prediction grid point information, a first weight is determined.

[0027] According to the physical consistency loss term, the first weight, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state, and the historical sample wind direction flow field state, a training loss function of the low-altitude flow field intelligent inference model is determined.

[0028] According to the inference task information and the historical prediction grid point information, a first weight is determined.

[0029] According to the inference task information, a spatial importance coefficient of each historical prediction grid point is determined.

[0030] According to the historical prediction grid point information, an observation point identification result and a historical grid point height standard deviation are determined.

[0031] According to the historical prediction grid point information, a historical prediction point signal-to-noise ratio of observation information of each historical prediction grid point is determined.

[0032] The historical observation point distance of the nearest observation point of each historical prediction grid point is obtained.

[0033] According to the historical grid point height standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio, and the historical observation point distance, the first weight is determined.

[0034] According to the historical grid point height standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio, and the historical observation point distance, the first weight is determined. ,

[0035] The first weight of the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period is determined , wherein if is a conditional function, 、 、 and is a first preset weight value, is a spatial importance coefficient of a jth historical prediction grid point corresponding to a kth historical prediction moment of an ith historical time period, is an observation point identification result of a jth historical prediction grid point corresponding to a kth historical prediction moment of an ith historical time period, when the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period belongs to an observation point, 1, when the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period does not belong to an observation point, 0, is a historical prediction point signal-to-noise ratio of a jth historical prediction grid point corresponding to a kth historical prediction moment of an ith historical time period is a preset signal-to-noise ratio threshold value, is a historical observation point distance from a jth historical prediction grid point corresponding to a kth historical prediction moment of an ith historical time period to a nearest observation point, is a historical grid point elevation standard deviation of a jth historical prediction grid point corresponding to a kth historical prediction moment of an ith historical time period.

[0036] According to a second aspect of the present application, a low-altitude flow field intelligent deduction system based on physical AI is provided, comprising:

[0037] a real-time data module configured to acquire scene geometry information, obstacle information, boundary conditions and sparse observation data at multiple moments of a current time period, and construct input features through a sliding window or prior knowledge;

[0038] a historical data module configured to acquire historical scene geometry information, historical obstacle information, historical boundary conditions and historical sparse observation data at multiple moments of multiple historical time periods;

[0039] an actual state module configured to determine historical actual wind speed flow field states and historical actual wind direction flow field states at multiple historical prediction grid points corresponding to multiple historical prediction moments of the multiple historical time periods, wherein the historical actual wind speed flow field states and the historical actual wind direction flow field states are represented as continuous vectors or discrete classification distributions;

[0040] a sample state module configured to process the historical scene geometry information, the historical obstacle information, the historical boundary conditions and the historical sparse observation data through a low-altitude flow field intelligent deduction model to obtain historical sample wind speed flow field states and historical sample wind direction flow field states at multiple historical prediction grid points corresponding to multiple historical prediction moments;

[0041] a loss function module configured to determine a training loss function of the low-altitude flow field intelligent inference model according to a physical consistency loss term, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state, and the historical sample wind direction flow field state, wherein the physical consistency loss term includes but is not limited to observation consistency, boundary consistency, geometric consistency, or constraints based on fluid mechanics laws;

[0042] a model training module configured to train the low-altitude flow field intelligent inference model according to the training loss function of the low-altitude flow field intelligent inference model, and obtain a trained low-altitude flow field intelligent inference model;

[0043] an inference result module configured to process the input feature according to the trained low-altitude flow field intelligent inference model, and obtain a low-altitude flow field intelligent inference result, wherein the low-altitude flow field intelligent inference result includes a three-dimensional vector wind field, and supports continuous value output or classification distribution output;

[0044] an inference service module configured to publish, cache, or periodically update the low-altitude flow field intelligent inference result through a low-latency output and distribution mechanism, and integrate a fallback strategy to process abnormal data flow.

[0045] Technical effects: According to the application, the historical actual wind speed flow field state and the historical actual wind direction flow field state can be determined, and the historical sample wind speed flow field state and the historical sample wind direction flow field state can be obtained through the low-altitude flow field intelligent deduction model. Further, the training loss function of the low-altitude flow field intelligent deduction model is determined according to the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state, and the low-altitude flow field intelligent deduction model is trained according to the training loss function to obtain a trained low-altitude flow field intelligent deduction model. The trained low-altitude flow field intelligent deduction model is used to complete the deduction task, thereby improving the accuracy of low-altitude flow field intelligent deduction. When determining the first weight, the first weight can be determined according to the historical grid point elevation standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio and the historical observation point distance. In the calculation process, the first weight can be determined according to the spatial importance coefficient, the data quality condition, the data sparsity condition and the terrain complexity condition, thereby improving the rationality of the first weight setting. When determining the training loss function of the low-altitude flow field intelligent deduction model, the training loss function of the low-altitude flow field intelligent deduction model can be determined according to the first weight, the deduced wind speed physical data, the deduced wind direction physical data, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state. In the calculation process, the weight of the loss value of the wind speed distribution can be determined according to the law of conservation of mass, the kinetic energy intensity and the law of conservation of energy. The weight of the loss value of the wind direction distribution can be determined according to the vorticity and the vorticity dynamics law. The training loss function of the low-altitude flow field intelligent deduction model is determined by combining the first weight, the basic cross-entropy loss value of the wind direction distribution and the wind speed distribution, thereby improving the physical law of the deduction result of the low-altitude flow field intelligent deduction model. The low-altitude flow field intelligent deduction model pays more attention to the observation data sparse area, the observation data low quality area, the deduction task critical area, the complex terrain area, the high kinetic energy area and the high vorticity area, thereby improving the accuracy of the low-altitude flow field intelligent deduction model.

[0046] It should be understood that the above general description and the following detailed description are exemplary and explanatory, but not limiting the present application. Other features and aspects of the present application will become more apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other embodiments from these drawings without creative labor;

[0048] Figure 1 Fig. 1 shows a flowchart of a method for low-altitude flow field intelligent deduction based on physical AI according to an embodiment of the present application;

[0049] Figure 2 Fig. 2 shows a schematic diagram for determining a historical actual wind speed flow field state and a historical actual wind direction flow field state according to an embodiment of the present application;

[0050] Figure 3 Fig. 3 shows a schematic diagram for determining a training loss function of a low-altitude flow field intelligent deduction model according to an embodiment of the present application;

[0051] Figure 4 Fig. 4 shows a block diagram of a low-altitude flow field intelligent deduction system based on physical AI according to an embodiment of the present application. DETAILED DESCRIPTION

[0052] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0053] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and some embodiments may not be described again for the same or similar concepts or processes.

[0054] Figure 1 Fig. 1 shows a flowchart of a method for low-altitude flow field intelligent deduction based on physical AI according to an embodiment of the present application, the method comprising:

[0055] Step S1, acquiring scene geometry information, obstacle information, boundary conditions and sparse observation data at multiple time points of a current time period, and constructing input features through a sliding window or prior knowledge;

[0056] Step S2, acquiring historical scene geometry information, historical obstacle information, historical boundary conditions and historical sparse observation data at multiple time points of multiple historical time periods;

[0057] Step S3, determining a historical actual wind speed flow field state and a historical actual wind direction flow field state at multiple historical prediction grid points of multiple historical prediction time points corresponding to the multiple historical time periods, wherein the historical actual wind speed flow field state and the historical actual wind direction flow field state are represented as continuous vectors or discrete classification distributions;

[0058] In step S4, the historical scene geometry information, the historical obstacle information, the historical boundary condition and the historical sparse observation data are processed by the low-altitude flow field intelligent deduction model to obtain historical sample wind speed flow field states and historical sample wind direction flow field states at a plurality of historical prediction grid points at a plurality of historical prediction moments.

[0059] In step S5, a training loss function of the low-altitude flow field intelligent deduction model is determined according to a physical consistency loss term, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state, wherein the physical consistency loss term includes but is not limited to observation consistency, boundary consistency, geometric consistency or constraint based on fluid mechanics law.

[0060] In step S6, the low-altitude flow field intelligent deduction model is trained according to the training loss function of the low-altitude flow field intelligent deduction model to obtain a trained low-altitude flow field intelligent deduction model.

[0061] In step S7, the input feature is processed according to the trained low-altitude flow field intelligent deduction model to obtain a low-altitude flow field intelligent deduction result, wherein the low-altitude flow field intelligent deduction result includes a three-dimensional vector wind field and supports continuous value output or classification distribution output.

[0062] In step S8, the low-altitude flow field intelligent deduction result is published, cached or periodically updated by a low-latency output and distribution mechanism, and a fallback strategy is integrated to process abnormal data flow.

[0063] The low-altitude flow field intelligent deduction method based on physical AI can determine the historical actual wind speed flow field state and the historical actual wind direction flow field state, and can obtain the historical sample wind speed flow field state and the historical sample wind direction flow field state through the low-altitude flow field intelligent deduction model. Further, the training loss function of the low-altitude flow field intelligent deduction model is determined according to the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state, and the low-altitude flow field intelligent deduction model is trained according to the training loss function to obtain a trained low-altitude flow field intelligent deduction model. The trained low-altitude flow field intelligent deduction model is used to complete the deduction task, and the accuracy of the low-altitude flow field intelligent deduction is improved.

[0064] According to an embodiment of the present application, in step S1, the scene geometry information, the obstacle information, the boundary condition and the sparse observation data at a plurality of moments in a current time period are obtained, and the input feature is constructed by a sliding window or prior knowledge.

[0065] For example, the sparse observation data refers to the discontinuous and non-dense measured wind field and related meteorological data in the deduction area in the spatial distribution, which can be directly measured by physical sensors, and specifically includes basic wind field variables (such as east-west wind speed, north-south wind speed), thermodynamic variables (such as temperature gradient, air pressure), and turbulence and stratification parameters (such as turbulent kinetic energy, Richardson number), scene geometry information and obstacle information refer to the terrain or obstacle information in the deduction area, such as terrain elevation, building group geometry information, boundary conditions refer to the values or constraint conditions of physical quantities required to be set on the boundary of the physical model or the deduction system, mainly divided into spatial boundary conditions and physical boundary conditions, usually derived based on larger scale data or physical laws, such as large scale numerical weather prediction model, geographic information system and physical law, if the current deduction task is to deduce the low-altitude flow field at 00:30 on August 8 of this month, the preset time window before August 8 is set as the current time period, the interval between each time of the time period is the preset sampling interval, the scene geometry information, obstacle information, boundary condition and sparse observation data of the deduction area at multiple times of the current time period are obtained through meteorological software or meteorological collection equipment, and the input features are constructed through sliding window or priori knowledge, wherein, "constructing input features through sliding window or priori knowledge" refers to preprocessing and structuring the original, discrete observation data before inputting the data into the low-altitude flow field intelligent deduction model, so as to organize the data into meaningful features that can be effectively recognized and learned by the model, the sliding window mainly processes the data in the time dimension, emphasizes the real-time and dynamic evolution ability of the model, and the priori knowledge mainly integrates physical laws and historical experience, emphasizes the physical rationality and inference ability of the model in the case of sparse data.

[0066] According to one embodiment of the present application, in step S2, the historical scene geometry information, historical obstacle information, historical boundary condition and historical sparse observation data of multiple times of multiple historical time periods are obtained.

[0067] For example, the length of the historical time period is the length of the preset time window, and the historical scene geometry information, historical obstacle information, historical boundary condition and historical sparse observation data of multiple times of multiple historical time periods are obtained.

[0068] According to one embodiment of the present application, in step S3, the historical actual wind speed flow field state and the historical actual wind direction flow field state at the historical prediction grid points of the multiple historical prediction times corresponding to the multiple historical time periods are determined, wherein the historical actual wind speed flow field state and the historical actual wind direction flow field state are represented as continuous vectors or discrete classification distributions.

[0069] For example, the first historical time period is from June 6th to July 7th last month, the historical prediction time corresponding to the first historical time period is 12:30 a.m. on August 8th last month, and the historical prediction grid point is a point in the deduction area. The historical actual wind speed flow field state and the historical actual wind direction flow field state at the historical prediction time corresponding to the plurality of historical time periods are determined.

[0070] Figure 2 An exemplary schematic diagram of determining the historical actual wind speed flow field state and the historical actual wind direction flow field state according to an embodiment of the present application is shown.

[0071] According to an embodiment of the present application, step S3 comprises:

[0072] Step S31, obtaining historical multi-source observation data of a historical prediction time corresponding to a historical time period;

[0073] Step S32, generating a historical actual three-dimensional flow field at the historical prediction time according to the historical multi-source observation data;

[0074] Step S33, determining the historical actual wind speed flow field state and the historical actual wind direction flow field state according to the historical actual three-dimensional flow field.

[0075] For example, historical multi-source observation data corresponding to a historical prediction time period is acquired, such as through ground observation stations (directly measuring wind speed and direction at a specific point), wind profile radars (measuring wind speed and direction at different heights), Doppler lidar (providing higher-precision three-dimensional wind field information), weather balloons (directly measuring meteorological elements on a vertical profile), and satellite remote sensing (providing large-scale wind field information) to acquire historical multi-source observation data; a historical actual three-dimensional flow field at the historical prediction time is generated according to the historical multi-source observation data, such as data assimilation and fusion of the historical multi-source observation data, using a physical model (such as a fluid equation) as a constraint, and calculating a complete, physically consistent, continuous analysis field on the entire region by a mathematical algorithm (such as a variational method or ensemble Kalman filtering), that is, the historical actual three-dimensional flow field; a historical actual wind speed flow field state and a historical actual wind direction flow field state are determined according to the historical actual three-dimensional flow field, wherein the historical actual wind speed flow field state and the historical actual wind direction flow field state are represented as continuous vectors or discrete classification distributions, such as determining historical actual wind speed and historical actual wind direction of a plurality of historical prediction grid points at the historical prediction time according to the historical actual three-dimensional flow field, wherein the historical actual wind direction is a continuous wind direction value at the historical prediction grid point; a wind speed classification rule and a wind direction classification rule are defined, such as setting wind speed of 0-3 m / s as a first wind speed, setting 3-6 m / s as a second wind speed, setting 6-9 m / s as a third wind speed, and so on, to divide wind speed into five categories, and defining a wind direction classification rule, such as classifying wind direction into north wind, northeast wind, east wind, southeast wind, south wind, southwest wind, west wind, and northwest wind, with north wind as a first wind direction, northeast wind as a second wind direction, and so on, to divide wind direction into eight categories; a historical actual wind speed flow field state is determined according to the historical actual wind speed and the wind speed classification rule, such as using one-hot encoding to represent the historical actual wind speed flow field state of the first historical prediction grid point, when the historical actual wind speed of the first historical prediction grid point is 8 m / s, the historical actual wind speed belongs to the third wind speed, and the historical actual wind speed flow field state of the historical actual wind speed of the first historical prediction grid point is a [0, 0, 1, 0, 0] vector, indicating that the wind speed at the historical prediction grid point is 100% of the third wind speed; a historical actual wind direction flow field state is determined according to the historical actual wind direction and the wind direction classification rule, such as when the historical actual wind direction of the first historical prediction grid point is northeast wind, the historical actual wind direction flow field state of the first historical prediction grid point is a [0, 1, 0, 0, 0, 0, 0, 0] vector, indicating that the wind direction at the historical prediction grid point is 100% of the second wind direction.

[0076] According to one embodiment of the present application, the low-altitude flow field intelligent deduction model adopts hierarchical reasoning, including:

[0077] An initial wind field is generated according to the sparse observation data and the boundary conditions;

[0078] generate a global high-resolution wind field according to the initial wind field, the scene geometry information and the obstacle information.

[0079] For example, an initial three-dimensional flow field, i.e., an initial wind field, is generated according to sparse observation data and the boundary conditions; a global high-resolution wind field is generated according to the initial wind field, scene geometry information and obstacle information, the scene geometry information and the obstacle information are introduced into the initial wind field, and a wind field optimization is applied by using a physical conservation law to generate the global high-resolution wind field.

[0080] According to one embodiment of the present application, an initial wind field is generated according to the sparse observation data and the boundary conditions, comprising:

[0081] The sparse observation data is processed according to a multi-source observation data assimilation technology to obtain processed sparse observation data;

[0082] An initial wind field is generated according to the processed sparse observation data and the boundary conditions.

[0083] For example, multi-source, heterogeneous, possibly containing noise and missing observation data are standardized, quality controlled, and features recognizable by the model are extracted, i.e., processed sparse observation data; an initial wind field is generated according to the processed sparse observation data and in combination with the boundary conditions, the initial wind field covers the entire three-dimensional spatial domain that needs to be derived,

[0084] According to one embodiment of the present application, in step S4, the historical scene geometry information, the historical obstacle information, the historical boundary conditions and the historical sparse observation data are processed by a low-altitude flow field intelligent derivation model to obtain historical sample wind speed flow field states and historical sample wind direction flow field states at a plurality of historical prediction grid points at a plurality of historical prediction times.

[0085] For example, the low-altitude flow field intelligent derivation model belongs to a kind of neural network model, comprising: data preprocessing and input layer, feature extraction layer, feature fusion layer and decision and output layer, the low-altitude flow field intelligent derivation model is trained by historical data, so that the low-altitude flow field intelligent derivation model can derive historical sample wind speed flow field states and historical sample wind direction flow field states at a plurality of historical prediction grid points, the historical scene geometry information, the historical obstacle information, the historical boundary conditions and the historical sparse observation data are processed by the low-altitude flow field intelligent derivation model, to determine the global high-resolution wind field at the historical prediction time, according to the global high-resolution wind field at the historical prediction time, the historical sample wind speed flow field states and the historical sample wind direction flow field states at a plurality of historical prediction grid points are determined, wherein the determination mode of the historical sample wind speed flow field state and the historical sample wind direction flow field state is similar to that of the historical actual wind speed flow field state and the historical actual wind direction flow field state, which will not be described here.

[0086] According to one embodiment of the present application, in step S5, a training loss function of the low-altitude flow field intelligent deduction model is determined according to the physical consistency loss term, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state, wherein the physical consistency loss term includes but is not limited to observation consistency, boundary consistency, geometric consistency or constraints based on fluid mechanics laws.

[0087] According to one embodiment of the present application, step S5 includes:

[0088] Step S51, obtaining deduction task information and historical prediction grid point information of historical prediction grid points;

[0089] Step S52, determining a first weight according to the deduction task information and the historical prediction grid point information;

[0090] Step S53, determining a training loss function of the low-altitude flow field intelligent deduction model according to the physical consistency loss term, the first weight, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state.

[0091] For example, the deduction task information (such as deduction area information and deduction target information) and the historical prediction grid point information (such as the quality of the data collected at the historical prediction grid points) are obtained; the first weight is determined according to the deduction task information and the historical prediction grid point information; and the training loss function of the low-altitude flow field intelligent deduction model is determined according to the physical consistency loss term, the first weight, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state, wherein the physical consistency loss term includes but is not limited to observation consistency (calculated based on the deviation between observation data and deduction results), boundary consistency (to ensure that the flow field boundary conditions meet physical constraints), geometric consistency (to incorporate terrain or obstacle geometric information) or constraints based on fluid mechanics laws (such as mass conservation (divergence), energy conservation (kinetic energy) or vorticity dynamics laws).

[0092] According to one embodiment of the present application, step S52 includes:

[0093] Step S521, determining a spatial importance coefficient of each historical prediction grid point according to the deduction task information;

[0094] Step S522, determining an observation point identification result and a historical grid point elevation standard deviation according to the historical prediction grid point information;

[0095] Step S523, determining a historical prediction point signal-to-noise ratio of observation information of each historical prediction grid point according to the historical prediction grid point information;

[0096] Step S524, obtaining a historical observation point distance of a nearest observation point of each historical prediction grid point;

[0097] Step S525, determining a first weight according to the historical grid point height standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio and the historical observation point distance.

[0098] For example, according to the deduction task information, the importance of each historical observation grid point in the position to the deduction task is determined, for example, when the deduction task is “aviation safety and airport scheduling”, when the historical observation grid point is in a key position (for example, an airport runway, a taxiway and a parking apron), the spatial importance coefficient of the historical prediction grid point is 5, when the historical observation grid point is in a secondary key position (for example, an approach surface, a take-off climb area, and a waiting airspace), the spatial importance coefficient of the historical prediction grid point is 2, and when the historical observation grid point is in other non-key positions, the spatial importance coefficient of the historical prediction grid point is 1; according to the historical prediction grid point information, the observation point identification result is determined, for example, when the first historical prediction grid point is provided with a sensor or can directly observe data, it is indicated that the historical prediction grid point belongs to an observation point, the observation point identification result of the first historical prediction grid point is 1, otherwise, the observation point identification result of the first historical prediction grid point is 0, indicating that the historical prediction grid point does not belong to an observation point, according to the historical prediction grid point information, the historical grid point height standard deviation is determined, for example, by a digital elevation model, the height standard deviation of a preset area (the preset area size is 3m*3m) centered on the projection of each historical prediction grid point on the ground is calculated, that is, the historical grid point height standard deviation; according to the historical prediction grid point information, the historical prediction point signal-to-noise ratio of observation information of each historical prediction grid point (the average signal-to-noise ratio of observation information collected by the historical prediction grid point in a historical time period) is determined; the historical observation point distance (the distance between the historical prediction grid point and the nearest observation point) of the nearest observation point (the position closest to the historical prediction grid point and provided with a sensor or capable of directly observing data) of each historical prediction grid point is obtained; and the first weight is determined according to the historical grid point height standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio and the historical observation point distance.

[0099] According to one embodiment of the present application, step S525 comprises: determining the first weight of the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period according to formula (1) , (1),

[0100] wherein if is a conditional function, , , and is a first preset weight value, is a spatial importance coefficient of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is an observation point identification result of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, when the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period belongs to an observation point, 1, when the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period does not belong to an observation point, 0, is a historical prediction point signal-to-noise ratio of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period is a preset signal-to-noise ratio threshold value, is a historical observation point distance from the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period to the nearest observation point, is a historical grid point elevation standard deviation of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period.

[0101] According to one embodiment of the present application, is a spatial importance coefficient of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, when the historical prediction grid point is more important for the position to the inference task, the greater the value of is, the greater the value of the first weight is, and thus when the low-altitude flow field intelligent inference model is trained using the training loss function set according to the first weight, the inference accuracy of the low-altitude flow field intelligent inference model for the key historical prediction grid point can be improved. is a historical observation point distance from the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period to the nearest observation point, the greater the value of is, the greater the first weight is, so that in the area with higher data sparsity or higher data uncertainty, the first weight is greater, guiding the low-altitude flow field intelligent inference model to still maintain physical reasonableness in these areas. is a historical grid point elevation standard deviation of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, The greater the value of the first weight is, the more complex the terrain at the ground projection position of the jth historical prediction grid point is, and the higher the prediction accuracy of the low-altitude flow field intelligent deduction model for complex terrain is when the low-altitude flow field intelligent deduction model is trained using the training loss function set according to the first weight.

[0102] According to one embodiment of the present application, in formula (1), the conditional function includes the following two cases: when the condition is met, it means that the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period belongs to an observation point, observation information can be directly collected at the historical prediction grid point, the data quality at the historical prediction grid point needs to be evaluated, and the value of the conditional function is , wherein is the historical prediction point signal-to-noise ratio of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, and the smaller is, the worse the data quality at the historical prediction grid point is, The greater the value of is, the greater the first weight is, so that in the area with poor data quality, the first weight is greater, forcing the model to not completely rely on unreliable data, but to more comply with physical constraints to deduce reasonable values, thereby improving the physical logic of the low-altitude flow field intelligent deduction model, and when the condition is not met, it means that the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period does not belong to an observation point, and observation information cannot be directly collected at the historical prediction grid point. A preset signal-to-noise ratio threshold (such as a regional average signal-to-noise ratio) is set to ensure that the low-altitude flow field intelligent deduction model remains cautious in the data missing area, and the value of the conditional function is , and similarly, represents the data quality condition when observation information cannot be directly collected at the historical prediction grid point.

[0103] According to one embodiment of the present application represents that the first weight is determined according to the spatial importance coefficient, the data quality condition, the data sparsity condition, and the terrain complexity condition. , , and are first preset weight values, which can be set to 1, 0.9, 0.8, and 0.7, respectively.

[0104] In this way, the first weight can be determined according to the historical grid point elevation standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio and the historical observation point distance. In the calculation process, the first weight can be determined according to the spatial importance coefficient, the data quality condition, the data sparsity condition and the terrain complexity condition, thereby improving the rationality of the first weight setting.

[0105] According to one embodiment of the present application, the training loss function of the low-altitude flow field intelligent deduction model is determined according to the physical consistency loss term, the first weight, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state, and specifically comprises: when the training loss function is set according to the constraint based on the fluid mechanics law, the training loss function can be set as (2), wherein, , , , and are preset parameters, is the first weight of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the historical actual wind speed flow field state of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the historical sample wind speed flow field state of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the historical actual wind direction flow field state of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the historical sample wind direction flow field state of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the flow velocity field divergence of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the unit volume kinetic energy density of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the kinetic energy local change rate of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the kinetic energy flux divergence of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, is the pressure work of the jth historical prediction grid point at the kth historical prediction moment corresponding to the ith historical time period, the vorticity of the jth historical prediction grid point at the kth historical prediction time of the ith historical time period, the local variation rate of the vorticity of the jth historical prediction grid point at the kth historical prediction time of the ith historical time period, the advection term of the vorticity of the jth historical prediction grid point at the kth historical prediction time of the ith historical time period, the stretching and twisting term of the vorticity of the jth historical prediction grid point at the kth historical prediction time of the ith historical time period, n is the number of historical time periods, i≤n, K is the number of historical prediction times, k≤K, m is the number of historical prediction grid points, j≤m, and i, n, k, K, j, and m are all positive integers.

[0106] According to one embodiment of the present application, the cross-entropy loss value of the basis wind speed distribution, the cross-entropy loss value of the basis wind direction distribution, which is calculated as follows, for example, when is [0, 0, 1, 0, 0], when is [0.1, 0.1, 0.6, 0.15, 0.05], then If the low-altitude flow field intelligent deduction model is completely correct, the cross-entropy loss value is 0, and if the low-altitude flow field intelligent deduction model is completely incorrect, the cross-entropy loss value tends to infinity.

[0107] According to one embodiment of the present application, the divergence of the flow velocity field of the jth historical prediction grid point at the kth historical prediction time of the ith historical time period, for low-altitude incompressible air flow, the divergence should be close to zero, which indicates that the air is a continuous medium and cannot be created out of nothing or disappear into thin air at a certain point, represents the degree to which the divergence of the flow velocity field of the jth historical prediction grid point violates the law of conservation of mass, The greater the value of, the higher the degree to which the divergence of the flow velocity field of the jth historical prediction grid point violates the law of conservation of mass, and the greater the loss value of the wind speed distribution, so that the low-altitude flow field intelligent deduction model must prioritize adjusting its parameters during the training process to make the deduced future flow field as much as possible to satisfy the law of conservation of mass, the unit volume kinetic energy density of the jth historical prediction grid point at the kth historical prediction time of the ith historical time period, The greater the value of, the greater the kinetic energy at the jth historical prediction grid point, and the more important the influence of high kinetic energy regions (such as the center of the jet stream) on the energy balance and dynamic processes of the overall flow field, The greater the value of, the greater the loss value of the wind speed distribution, so that the low-altitude flow field intelligent deduction model pays more attention to high kinetic energy regions during the training process, The sum of the kinetic energy local change rate, the kinetic energy flux divergence and the pressure work of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period is 0 under ideal conditions due to energy conservation, The sum of the kinetic energy local change rate, the kinetic energy flux divergence and the pressure work of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period is 0 under ideal conditions due to energy conservation, The greater the value of the sum of the kinetic energy local change rate, the kinetic energy flux divergence and the pressure work of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period, the higher the severity of violation of the energy conservation law by the flow field deduced by the low-altitude flow field intelligent deduction model at the jth historical prediction grid point, The greater the value of the sum of the kinetic energy local change rate, the kinetic energy flux divergence and the pressure work of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period, the greater the loss value of the wind speed distribution, so that the low-altitude flow field intelligent deduction model must prioritize adjusting its parameters during the training process to make the deduced future flow field satisfy the energy conservation law as much as possible. The weight of the loss value of the wind speed distribution determined according to the mass conservation law, the kinetic energy intensity and the energy conservation law, 、 and may be set to 1, 0.5 and 0.3, respectively.

[0108] According to one embodiment of the present application, The vorticity of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period, The intensity of the vorticity of the jth historical prediction grid point, The greater the value of the vorticity of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period, the greater the loss value of the wind direction distribution, so that the low-altitude flow field intelligent deduction model pays more attention to the high-vorticity area during the training process, The value of the vorticity local change rate plus the vorticity advection term minus the vorticity stretching and twisting term of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period is 0 under ideal conditions, The value of the vorticity local change rate plus the vorticity advection term minus the vorticity stretching and twisting term of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period is 0 under ideal conditions, The greater the value of the vorticity local change rate plus the vorticity advection term minus the vorticity stretching and twisting term of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period, the higher the severity of violation of the vorticity dynamics law by the flow field deduced by the low-altitude flow field intelligent deduction model at the jth historical prediction grid point, The greater the value of the vorticity local change rate plus the vorticity advection term minus the vorticity stretching and twisting term of the jth historical prediction grid point corresponding to the kth historical prediction moment of the ith historical time period, the greater the loss value of the wind direction distribution, so that the low-altitude flow field intelligent deduction model improves the rationality of the evolution (such as movement, strengthening and weakening) of the vortex during the training process, The weight of the loss value of the wind direction distribution determined according to the vorticity and the vorticity dynamics law, and may be initially set to 1 and 0.4, respectively.

[0109] In this way, the training loss function of the low-altitude flow field intelligent deduction model can be determined according to the first weight, the deduced wind speed physical data, the deduced wind direction physical data, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state. In the calculation process, the weight of the loss value of the wind speed distribution can be determined according to the law of conservation of mass, the kinetic energy intensity and the law of conservation of energy, the weight of the loss value of the wind direction distribution can be determined according to the vorticity and the vorticity dynamics law, the training loss function of the low-altitude flow field intelligent deduction model can be determined by combining the first weight, the basic cross-entropy loss value of the wind direction distribution and the wind speed distribution, the physical law of the deduction result of the low-altitude flow field intelligent deduction model is improved, the low-altitude flow field intelligent deduction model pays more attention to the observation data sparse area, the observation data low quality area, the deduction task key area, the complex terrain area, the high kinetic energy area and the high vorticity area, and the accuracy of the low-altitude flow field intelligent deduction model is improved.

[0110] According to one embodiment of the present application, the training loss function of the low-altitude flow field intelligent deduction model is determined according to the physical consistency loss term, the first weight, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state, and specifically includes: when the training loss function is set according to the observation consistency, the boundary consistency and the geometric consistency, the training loss function can be set as follows: 1, when the training loss function is set according to the observation consistency, the mean square error of the predicted value (such as wind direction and wind speed) of the low-altitude flow field intelligent deduction model at the historical prediction grid point and the actual observation value at the historical prediction grid point is taken as a loss term, so that the prediction result of the low-altitude flow field intelligent deduction model is as consistent as possible with the actual sparse observation data, the low-altitude flow field intelligent deduction model is ensured to output “anchored” on the real data, and the solution is prevented from deviating from reality. 2, when the training loss function is set according to the boundary consistency, a series of points can be sampled on the boundary, the difference between the model prediction value and the given boundary condition is calculated, and the physical reasonableness of the solution on the boundary is ensured, which is a prerequisite for obtaining a correct solution in the whole domain. 3, when the training loss function is set according to the geometric consistency, the complex calculation domain geometry (such as terrain, buildings) can be integrated into the constraint, such as applying physical constraints in the domain (this is the most important way, the residual loss term of the physical equation (such as Navier-Stokes equation) will naturally take effect in the whole calculation domain including the geometric boundary, and the low-altitude flow field intelligent deduction model will learn the profound influence of the geometric shape on the fluid motion during training), the information describing the geometry (such as the signed distance function SDF to the boundary, the terrain elevation) is taken as an additional input feature and is transmitted to the network, so that the low-altitude flow field intelligent deduction model learns the relationship between the geometry and the flow field.

[0111] According to one embodiment of the present application, in step S6, the low-altitude flow field intelligent deduction model is trained according to the training loss function of the low-altitude flow field intelligent deduction model, and a trained low-altitude flow field intelligent deduction model is obtained.

[0112] For example, the low-altitude flow field intelligent deduction model is trained according to the training loss function of the low-altitude flow field intelligent deduction model, the deduction accuracy of the low-altitude flow field intelligent deduction model for the low-altitude flow field is improved, and a trained low-altitude flow field intelligent deduction model is obtained.

[0113] According to one embodiment of the present application, in step S7, the input features are processed according to the trained low-altitude flow field intelligent deduction model, and a low-altitude flow field intelligent deduction result is obtained, wherein the low-altitude flow field intelligent deduction result includes a three-dimensional vector wind field and supports continuous value output or classification distribution output.

[0114] For example, the basic wind field variables, thermodynamic variables, and turbulence and stratification parameters are processed according to the trained low-altitude flow field intelligent deduction model, and a low-altitude flow field intelligent deduction result can be obtained, such as a future flow velocity field, a future wind direction field, a series of derived fields with important significance calculated in real time based on the predicted flow velocity field, the overall structure of the flow field, and key phenomena.

[0115] According to one embodiment of the present application, in step S8, the low-altitude flow field intelligent deduction result is published, cached, or periodically updated through a low-latency output and distribution mechanism, and a fallback strategy is integrated to process abnormal data flow.

[0116] For example, after the low-altitude flow field intelligent deduction model deduces the low-altitude flow field intelligent deduction result, the message queue / publish-subscribe model is used to publish the low-altitude flow field intelligent deduction result to all clients (such as flight control systems and visualization platforms) interested in the result, and the latest or most commonly used low-altitude flow field intelligent deduction result is stored in a high-speed in-memory database. When multiple clients request the same data, there is no need to re-run the model or access the slow disk, and the data is directly read from the memory, which greatly reduces the delay. The system automatically starts a complete reasoning process at a preset update period, and when important new observation data (such as sudden emergency weather observation) or external trigger signals are received, the current period is immediately interrupted, and a new deduction is started.

[0117] The low-altitude flow field intelligent deduction method based on physical AI according to the embodiment of the application can determine the historical actual wind speed flow field state and the historical actual wind direction flow field state, and can obtain the historical sample wind speed flow field state and the historical sample wind direction flow field state through the low-altitude flow field intelligent deduction model. Further, the training loss function of the low-altitude flow field intelligent deduction model is determined according to the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state, and the low-altitude flow field intelligent deduction model is trained according to the training loss function to obtain the trained low-altitude flow field intelligent deduction model. The trained low-altitude flow field intelligent deduction model is used to complete the deduction task, thereby improving the accuracy of the low-altitude flow field intelligent deduction. When the first weight is determined, the first weight can be determined according to the historical grid point elevation standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio and the historical observation point distance. In the calculation process, the first weight can be determined according to the spatial importance coefficient, the data quality condition, the data sparsity condition and the terrain complexity condition, thereby improving the rationality of the first weight setting. When the training loss function of the low-altitude flow field intelligent deduction model is determined, the training loss function of the low-altitude flow field intelligent deduction model can be determined according to the first weight, the deduced wind speed physical data, the deduced wind direction physical data, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state and the historical sample wind direction flow field state. In the calculation process, the weight of the loss value of the wind speed distribution can be determined according to the law of conservation of mass, the kinetic energy intensity and the law of conservation of energy. The weight of the loss value of the wind direction distribution can be determined according to the vorticity and the vorticity dynamics law. The training loss function of the low-altitude flow field intelligent deduction model is determined by combining the first weight, the basic cross-entropy loss value of the wind direction distribution and the wind speed distribution, thereby improving the physical law of the deduction result of the low-altitude flow field intelligent deduction model. The low-altitude flow field intelligent deduction model pays more attention to the observation data sparse area, the observation data low quality area, the deduction task key area, the complex terrain area, the high kinetic energy area and the high vorticity area, thereby improving the accuracy of the low-altitude flow field intelligent deduction model.

[0118] Figure 4 An exemplary block diagram of a low-altitude flow field intelligent deduction system based on physical AI according to an embodiment of the application is shown. The system includes:

[0119] A real-time data module is configured to acquire scene geometry information, obstacle information, boundary conditions and sparse observation data at multiple time points in a current time period, and construct input features through a sliding window or prior knowledge.

[0120] A historical data module is configured to acquire historical scene geometry information, historical obstacle information, historical boundary conditions and historical sparse observation data at multiple time points in multiple historical time periods.

[0121] An actual state module is configured to determine historical actual wind speed flow field states and historical actual wind direction flow field states at historical prediction grid points corresponding to a plurality of historical prediction moments of a plurality of historical time periods, wherein the historical actual wind speed flow field states and the historical actual wind direction flow field states are represented as continuous vectors or discrete classification distributions;

[0122] A sample state module is configured to process the historical scene geometric information, the historical obstacle information, the historical boundary condition and the historical sparse observation data by a low-altitude flow field intelligent deduction model to obtain historical sample wind speed flow field states and historical sample wind direction flow field states at historical prediction grid points of a plurality of historical prediction moments;

[0123] A loss function module is configured to determine a training loss function of the low-altitude flow field intelligent deduction model according to a physical consistency loss term, the historical actual wind speed flow field states, the historical actual wind direction flow field states, the historical sample wind speed flow field states and the historical sample wind direction flow field states, wherein the physical consistency loss term includes but is not limited to observation consistency, boundary consistency, geometric consistency or constraints based on fluid mechanics laws;

[0124] A model training module is configured to train the low-altitude flow field intelligent deduction model according to the training loss function of the low-altitude flow field intelligent deduction model to obtain a trained low-altitude flow field intelligent deduction model;

[0125] A deduction result module is configured to process the input features according to the trained low-altitude flow field intelligent deduction model to obtain a low-altitude flow field intelligent deduction result, wherein the low-altitude flow field intelligent deduction result includes a three-dimensional vector wind field and supports continuous value output or classification distribution output;

[0126] An inference service module is configured to publish, cache or periodically update the low-altitude flow field intelligent deduction result by a low-latency output and distribution mechanism, and integrate a fallback strategy to process abnormal data flow.

[0127] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for performing various aspects of the present application.

[0128] Those skilled in the art will understand that the embodiments of the application shown in the above description and drawings are only examples and do not limit the application. The purpose of the application has been fully and effectively achieved. The function and structural principle of the application has been shown and explained in the embodiments, and the embodiments of the application can be modified or changed without departing from the principle.

Claims

1. A low-altitude flow field intelligent deduction method based on physical AI, characterized in that, The method comprises the following steps: obtaining scene geometry information, obstacle information, boundary conditions and sparse observation data at multiple time points of a current time period, and constructing input features through a sliding window or prior knowledge; obtaining historical scene geometry information, historical obstacle information, historical boundary conditions and historical sparse observation data at multiple time points of multiple historical time periods; determining historical actual wind speed flow field states and historical actual wind direction flow field states at multiple historical prediction grid points of multiple historical prediction time points corresponding to the multiple historical time periods, wherein the historical actual wind speed flow field states and the historical actual wind direction flow field states are represented as continuous vectors or discrete classification distributions; processing the historical scene geometry information, the historical obstacle information, the historical boundary conditions and the historical sparse observation data through a low-altitude flow field intelligent inference model to obtain historical sample wind speed flow field states and historical sample wind direction flow field states at multiple historical prediction grid points of multiple historical prediction time points; determining a training loss function of the low-altitude flow field intelligent inference model according to a physical consistency loss term, the historical actual wind speed flow field states, the historical actual wind direction flow field states, the historical sample wind speed flow field states and the historical sample wind direction flow field states, wherein the physical consistency loss term includes but is not limited to observation consistency, boundary consistency, geometric consistency or constraints based on fluid mechanics laws; training the low-altitude flow field intelligent inference model according to the training loss function of the low-altitude flow field intelligent inference model to obtain a trained low-altitude flow field intelligent inference model; processing the input features according to the trained low-altitude flow field intelligent inference model to obtain a low-altitude flow field intelligent inference result, wherein the low-altitude flow field intelligent inference result includes a three-dimensional vector wind field and supports continuous value output or classification distribution output; publishing, caching or periodically updating the low-altitude flow field intelligent inference result through a low-latency output and distribution mechanism, and integrating a fallback strategy to process abnormal data streams.

2. The physical AI-based low-altitude flow field intelligent deduction method according to claim 1, characterized in that, The low-altitude flow field intelligent inference model adopts hierarchical reasoning, which comprises: generating an initial wind field according to the sparse observation data and the boundary conditions; generating a global high-resolution wind field according to the initial wind field, the scene geometry information and the obstacle information.

3. The physical AI-based low-altitude flow field intelligent deduction method according to claim 2, characterized in that, Generating an initial wind field according to the sparse observation data and the boundary conditions comprises: processing the sparse observation data according to a multi-source observation data assimilation technology to obtain processed sparse observation data; generating an initial wind field according to the processed sparse observation data and the boundary conditions.

4. The physical AI-based low-altitude flow field intelligent deduction method according to claim 1, wherein, Determining historical actual wind speed flow field states and historical actual wind direction flow field states at multiple historical prediction grid points of multiple historical prediction time points corresponding to multiple historical time periods comprises: obtaining historical multi-source observation data at a historical prediction time point corresponding to a historical time period; generating a historical actual three-dimensional flow field at the historical prediction time point according to the historical multi-source observation data; determining historical actual wind speed flow field states and historical actual wind direction flow field states according to the historical actual three-dimensional flow field.

5. The physical AI-based low-altitude flow field intelligent deduction method according to claim 1, wherein, Determine a training loss function of the low-altitude flow field intelligent inference model according to the physical consistency loss term, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state, and the historical sample wind direction flow field state, including: Obtain inference task information and historical prediction grid point information of a historical prediction grid point; Determine a first weight according to the inference task information and the historical prediction grid point information; Determine a training loss function of the low-altitude flow field intelligent inference model according to the physical consistency loss term, the first weight, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state, and the historical sample wind direction flow field state.

6. The physical AI-based low-altitude flow field intelligent deduction method according to claim 5, characterized in that, Determine a first weight according to the inference task information and the historical prediction grid point information, including: Determine a spatial importance coefficient of each historical prediction grid point according to the inference task information; Determine an observation point identification result and a historical grid point elevation standard deviation according to the historical prediction grid point information; Determine a historical prediction point signal-to-noise ratio of observation information of each historical prediction grid point according to the historical prediction grid point information; Obtain a historical observation point distance of a nearest observation point of each historical prediction grid point; Determine a first weight according to the historical grid point elevation standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio, and the historical observation point distance.

7. The physical AI-based low-altitude flow field intelligent deduction method according to claim 6, characterized in that, Determine a first weight according to the historical grid point elevation standard deviation, the spatial importance coefficient, the observation point identification result, the historical prediction point signal-to-noise ratio, and the historical observation point distance, including: according to the formula: ， a first weight of a jth historical prediction grid point of a kth historical prediction moment corresponding to the ith historical time period wherein if is a conditional function, , , and is a first preset weight value, is a spatial importance coefficient of the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period, is an observation point identification result of the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period, when the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period belongs to an observation point, 1, when the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period does not belong to an observation point, 0, is a historical prediction point signal-to-noise ratio of the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period is a preset signal-to-noise ratio threshold value, is a historical observation point distance from the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period to a nearest observation point, is a historical grid point elevation standard deviation of the jth historical prediction grid point of the kth historical prediction moment corresponding to the ith historical time period.

8. A physical AI-based low-altitude flow field intelligent deduction system, characterized in that, For performing the method of any one of claims 1-7, including: A real-time data module is configured to obtain scene geometry information, obstacle information, boundary conditions, and sparse observation data at multiple time points in a current time period, and construct input features through a sliding window or prior knowledge; A historical data module is configured to obtain historical scene geometry information, historical obstacle information, historical boundary conditions, and historical sparse observation data at multiple time points in multiple historical time periods; An actual state module is configured to determine historical actual wind speed flow field states and historical actual wind direction flow field states at multiple historical prediction grid points at multiple historical prediction time points corresponding to the multiple historical time periods, wherein the historical actual wind speed flow field states and the historical actual wind direction flow field states are represented as continuous vectors or discrete classification distributions; A sample state module is configured to process the historical scene geometry information, the historical obstacle information, the historical boundary conditions, and the historical sparse observation data through a low-altitude flow field intelligent inference model to obtain historical sample wind speed flow field states and historical sample wind direction flow field states at multiple historical prediction grid points at multiple historical prediction time points; and A model training module is configured to determine a training loss function of the low-altitude flow field intelligent inference model according to a physical consistency loss term, the first weight, the historical actual wind speed flow field states, the historical actual wind direction flow field states, the historical sample wind speed flow field states, and the historical sample wind direction flow field states. The loss function module is configured to determine a training loss function of the low-altitude flow field intelligent inference model according to a physical consistency loss term, the historical actual wind speed flow field state, the historical actual wind direction flow field state, the historical sample wind speed flow field state, and the historical sample wind direction flow field state, wherein the physical consistency loss term includes but is not limited to observation consistency, boundary consistency, geometric consistency, or a constraint based on a fluid mechanics law; The model training module is configured to train the low-altitude flow field intelligent inference model according to the training loss function of the low-altitude flow field intelligent inference model, and obtain a trained low-altitude flow field intelligent inference model; The inference result module is configured to process the input feature according to the trained low-altitude flow field intelligent inference model, and obtain a low-altitude flow field intelligent inference result, wherein the low-altitude flow field intelligent inference result includes a three-dimensional vector wind field, and supports continuous value output or classification distribution output; The reasoning service module is configured to publish, cache, or periodically update the low-altitude flow field intelligent inference result through a low-latency output and distribution mechanism, and integrate a fallback strategy to process abnormal data flow.

Images