Apparatus and Method for Generating Ultra-high-resolution Local Meteorological Data Using Satellite 3D Digital Surface Models and Physics-Informed Neural Networks
The hybrid use of satellite-based 3D DSM, CFD, and PINN generates ultra-high-resolution weather data in real-time, addressing computational inefficiencies and physical inconsistencies in existing models, enhancing applications in urban environments and construction safety.
Patent Information
- Application Number
- KR1020260044975
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2026-02-04
- Filing Date
- 2026-03-12
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-03-12
AI Technical Summary
Existing weather prediction models struggle to provide ultra-high-resolution weather data in real-time for complex urban and mountainous terrains, as they either require massive computational resources or produce physically unrealistic predictions due to lacking fluid dynamic constraints.
A hybrid approach combining satellite-based 3D Digital Surface Model (DSM), Computational Fluid Dynamics (CFD) simulations, and Physics-Informed Neural Networks (PINN) to generate ultra-high-resolution weather data with a resolution of 5m or less, incorporating mass and momentum conservation equations into the loss function of the PINN.
Ensures real-time generation of physically consistent weather data with improved accuracy and reduced computational time, enabling applications in urban air mobility and construction safety.
Smart Images

Figure 112026030416376-PAT00007_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a weather data generation technology, and more specifically, to an apparatus and method for generating ultra-high-resolution local weather data with a resolution of 5m or less in real time in urban areas and complex terrains through hybrid modeling operations that combine satellite-based 3D Digital Surface Model (DSM) and wide-area numerical weather forecast data, and fuse Computational Fluid Dynamics (CFD) and Physics-Informed Neural Network (PINN). Background Technology
[0002] The numerical weather prediction models currently operated by the Korea Meteorological Administration (LDAPS, RDAPS, etc.) have a spatial resolution of approximately 1.5 km, which can be effectively utilized for national-level disaster response and understanding macroscopic weather patterns. However, there are limitations in precisely predicting localized turbulence, building wind, and gusts occurring in complex terrains, such as densely populated urban areas with high-rise buildings or mountainous terrain, using only existing wide-area models.
[0003] To overcome these limitations, Computational Fluid Dynamics (CFD) simulations have been utilized. However, this method involves directly solving partial differential equations, such as the CFD Navier-Stokes equations, by numerically solving them, which requires massive computational resources and calculation times ranging from tens of minutes to several hours. Consequently, while CFD is suitable for preliminary simulations or design reviews, it presents a problem in that it is practically difficult to apply it immediately to real-time weather forecasting services.
[0004] Another example is the approach of utilizing a deep learning-based pure statistical artificial intelligence (AI) model in real-time weather forecasting services. The approach using the aforementioned AI model has the advantage of rapidly extracting patterns from training data. On the other hand, if fluid dynamic physical constraints, such as the law of conservation of mass and the law of conservation of momentum, are not explicitly reflected in the learning process, it leads to problems such as producing physically unrealistic prediction results or reducing the reliability of predictions at unobserved points.
[0005] Therefore, ultra-high-resolution weather forecasting technology that ensures real-time capability while maintaining physical feasibility is required. Prior art literature
[0006] Korean Registered Patent 10-2390005 (2022. 04. 20) The problem to be solved
[0007] The objective of the present invention is to provide an apparatus and method for generating ultra-high-resolution weather data in 5m grid units in real time while maintaining physical consistency, by utilizing a precise 3D digital surface model (DSM) acquired based on satellite and LiDAR as a boundary condition and an artificial intelligence model learned from high-precision physical data generated by CFD simulation.
[0008] The technical problems of the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below. means of solving the problem
[0009] An ultra-high-resolution local weather data generation device according to an embodiment of the present invention for solving the above technical problem comprises: a data collection unit that receives three-dimensional digital surface model (DSM) data acquired from a satellite and low-resolution weather data from a wide-area numerical weather prediction model; a preprocessing unit that matches urban building and terrain information into a three-dimensional terrain grid based on the three-dimensional digital surface model; a hybrid computation engine that includes a Physics-Informed Neural Network that has learned high-resolution flow field data for various wind direction and wind speed scenarios generated in advance through computational fluid dynamics (CFD) simulation, and receives the low-resolution weather data and the three-dimensional grid information as input to calculate a high-resolution weather vector with a resolution of several meters or less in real time; and an output unit that provides the calculated high-resolution weather vector field to an external system. The physics-informed neural network includes, and is characterized by being trained to include the residuals of the mass conservation equation and the momentum conservation equation in the physics-informed loss function of the physics-informed neural network.
[0010] According to the present embodiment, the three-dimensional numerical surface model data includes data generated from satellite imagery, aerial LiDAR, or aerial orthophotos, and the preprocessing unit can generate a three-dimensional geospatial tensor including building height, surface elevation, and surface roughness information using the data.
[0011] According to the present embodiment, the preprocessing unit can set fluid non-transparency conditions and no-slip boundary conditions for the three-dimensional terrain grid.
[0012] According to the present embodiment, the hybrid computation engine can be trained by applying the boundary condition such that the velocity vector converges to 0 for the building surface and the ground surface grid.
[0013] The device according to the present embodiment may further include a real-time data correction unit that receives real-time wind speed and wind direction data from a weather sensor installed in an urban area, calculates the difference between the predicted value calculated by the hybrid computation engine and the actual value, and fine-tunes the weights of the physical information-based neural network or corrects the output value.
[0014] According to the present embodiment, the loss function of the physical information-based neural network comprises a data loss term (L data ) and physical loss term (L physics It is composed of a weighted sum of ), and the physical loss term may include a mass conservation constraint term based on the continuity equation (∇·u = 0); and a momentum conservation constraint term based on the residual of the Navier-Stokes equation.
[0015] A method for generating ultra-high-resolution local weather data according to another embodiment of the present invention comprises: a collection step of collecting three-dimensional numerical surface model data and low-resolution weather data obtained from a satellite; a preprocessing step of generating a three-dimensional terrain grid tensor including building height and terrain information using the three-dimensional numerical surface model, interpolating the low-resolution weather data into micro-grid units, and expanding the altitude-specific weather profile in the vertical direction to generate an initial weather field; a generation step of inputting the initial weather field and the three-dimensional terrain grid tensor into a physical information-based neural network to generate a high-resolution weather vector field with a resolution of several meters or less that satisfies mass conservation and momentum conservation constraints; and a providing step of providing the generated high-resolution weather vector field to an external system.
[0016] According to the present embodiment, the method may further include the step of receiving a measured value from a weather sensor, calculating an error between a predicted value generated in the generation step and the measured value, and correcting the predicted value.
[0017] According to the present embodiment, the physical information-based neural network can be pre-trained using the results of computational fluid dynamics simulations performed based on the Navier-Stokes equations as training data.
[0018] According to the present embodiment, the learning of the physical information-based neural network can be performed using a physical-based loss function comprising a data loss term and a physical loss term that satisfies the conservation of mass or momentum of the fluid.
[0019] According to the present embodiment, the ultra-high resolution weather data can be generated in the form of a weather vector field with a spatial resolution of 5 meters or less for urban areas or complex terrain areas. Effects of the invention
[0020] According to the present invention, unlike simple super-resolution models, physical consistency can be ensured by including physical laws (continuity equation and Navier-Stokes equation) in the loss function.
[0021] According to the present invention, automatic boundary condition generation is possible without manual modeling by directly utilizing a satellite 3D numerical surface model.
[0022] According to the present invention, real-time inference within 1 second is possible compared to conventional computational fluid dynamics (CFD).
[0023] According to the present invention, continuous improvement in accuracy is possible through IoT data assimilation.
[0024] According to the present invention, it is applicable to various industrial fields such as UAM, smart cities, and construction safety. Brief explanation of the drawing
[0025] FIG. 1 is an overall configuration diagram of an ultra-high resolution urban weather data generation device according to a preferred embodiment of the present invention. Figure 2 is a diagram illustrating the learning mechanism of a neural network model (PINN) running inside the hybrid computation engine of Figure 1. FIG. 3 is a diagram showing the spatial alignment process of macroscopic weather data and microscopic topographic information hierarchically according to an embodiment of the present invention. FIG. 4 is a diagram illustrating a specific example of a micro-weather vector field calculated when an embodiment of the present invention is applied to an actual complex urban environment. FIG. 5 is a diagram illustrating the data assimilation and feedback loop process that improves the prediction accuracy of the device of the present invention in real time. FIG. 6 is a flowchart illustrating a method for generating ultra-high resolution local weather data according to an embodiment of the present invention. Specific details for implementing the invention
[0026] The present invention is capable of various modifications and may have various embodiments, and specific embodiments are illustrated in the drawings and described in detail. However, this is not intended to limit the invention to specific embodiments, and it should be understood that it includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention. In describing the invention, detailed descriptions of related prior art are omitted if it is determined that such detailed descriptions may obscure the essence of the invention.
[0027] Terms such as "first," "second," etc., may be used to describe various components, but said components should not be limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.
[0028] The terms used in this invention are used merely to describe specific embodiments and are not intended to limit the invention. The singular expression includes the plural expression unless the context clearly indicates otherwise. In this application, terms such as "comprising" or "having" are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.
[0029] Spatially relative terms such as "below," "beneath," "lower," and "above" or "upper" may be used to facilitate the description of the relationship between one element or component and another, as illustrated in the drawings. Spatially relative terms should be understood as encompassing different orientations of the element during use or operation, in addition to the orientations illustrated in the drawings. For example, if an element illustrated in the drawings is flipped, an element described as being "below" or "beneath" another element may be placed "above" or "upper" of that other element. Therefore, the exemplary term "below" may encompass both the lower and upper directions. Elements may also be oriented in other directions, and accordingly, spatially relative terms may be interpreted based on the orientation.
[0030] Expressions indicating a part, such as “part” or “component,” used in this invention mean that the corresponding component may represent a device capable of including a specific function, software capable of including a specific function, or a combination of a device and software capable of including a specific function; however, it is not necessarily limited to the expressed function. This is provided merely to aid in a more comprehensive understanding of the invention, and a person with ordinary knowledge in the field to which this invention pertains can make various modifications and variations from this description.
[0031] Accordingly, the scope of the present invention is not limited to the described embodiments, and all things equivalent to or having equivalent variations to the claims set forth below, as well as the claims themselves, shall be considered to fall within the scope of the concept of the present invention.
[0033] The present invention will be described in more detail below based on the embodiments illustrated in the drawings.
[0034] The present invention is not merely a simple image resolution enhancement technology, but features a physics engine-embedded artificial intelligence structure that produces prediction results that do not violate physical laws by internalizing the governing equations of fluid dynamics into the loss function.
[0036] FIG. 1 is an overall configuration diagram of an ultra-high resolution urban weather data generation device according to a preferred embodiment of the present invention.
[0037] Referring to FIG. 1, the device (100) of the present invention comprises a data collection unit (110), a 3D terrain preprocessing unit (120), a hybrid computation engine (130), a real-time data correction unit (140), and a service interface (150).
[0038] In the embodiment, the data collection unit (110) can collect static data and dynamic data. The static data may utilize, for example, satellite imagery and a LiDAR-based 3D digital surface model (DSM), and may include spatial information essential for determining boundary conditions of the weather flow field, such as building heights in the city center, terrain elevation, and land cover information. The dynamic data may include, for example, weather data received from the Korea Meteorological Administration's wide-area numerical weather forecast models (LDAPS, RDAPS, GFS, etc.). The weather data includes time-varying weather factors such as wind direction, wind speed, temperature, atmospheric pressure, and humidity, which are collected in real time.
[0039] In the embodiment, the 3D terrain preprocessing unit (120) can convert the DSM data into a 3D grid tensor format that can be processed by a neural network. Specifically, satellite latitude and longitude coordinates expressed in the WGS84 coordinate system are converted into the UTM rectangular coordinate system to ensure accuracy in distance calculation, and the initial field is established by linearly interpolating wide-area grid data with a resolution of 1.5 km into a target resolution grid domain of 5 m. In addition, by applying vertical profiling based on atmospheric boundary layer theory to expand the single altitude wind speed data of LDAPS into an altitude-specific flow velocity distribution according to the Power Law or Log Law, 3D boundary conditions reflecting building height are automatically configured.
[0040] In the embodiment, the hybrid computation engine (130) may be configured as a Teacher-Student architecture that combines computational fluid dynamics (CFD)-based precision physics simulation and artificial intelligence (AI)-based high-speed inference capabilities. That is, it is a configuration that includes a CFD simulator (132) and a neural network model (210). In the training phase, the CFD simulator (132) generates a high-precision flow field dataset for various wind direction and wind speed scenarios in advance using numerical analysis software such as OpenFOAM based on the RANS (Reynolds-Averaged Navier-Stokes) or LES (Large Eddy Simulation) equations, and trains the neural network model using this as the ground truth label. In the inference phase, a 5m-class weather vector field can be generated within 1 second based on the LDAPS data and DSM information input in real time, using only the AI model without heavy CFD computation.
[0041] That is, the hybrid computation engine (130) performs hybrid modeling by combining physical consistency and the real-time nature of artificial intelligence (AI), wherein the hybrid modeling involves applying a physics-informed loss function to train the AI so that it does not violate physical laws.
[0042] In the embodiment, the real-time data correction unit (140) continuously analyzes the residual error between the predicted value calculated by the hybrid computation engine (130) and the actual measurement data received from small weather stations (AWS) and IoT sensors deployed within the city, and performs a closed-loop correction process to minimize the residual error. An Ensemble Kalman Filter (EnKF) is applied as the correction algorithm, thereby allowing the error information measured at a specific observation point to be physically and validly propagated to surrounding grids, so that spatially consistent correction results can be derived. Finally, the 5m-class ultra-high resolution weather information calculated through the service interface (150) is transmitted in the form of an API to various application fields such as a UAM control system, a construction safety system, and a digital twin platform.
[0044] Figure 2 is a diagram illustrating the learning mechanism of a neural network model (PINN) running inside the hybrid computation engine of Figure 1.
[0045] In this embodiment, the neural network model (210) is characterized in that it does not merely learn the statistical correlation between input and output data, but rather learns by internalizing the governing equations of fluid dynamics as constraints of the loss function, thereby performing physically valid predictions even at points where data is insufficient or unobserved.
[0046] To this end, the algorithm used in the neural network model (210) is as follows.
[0047] 1. Governing Equations (CFD-based)
[0048] The Navier-Stokes equation is used to simulate fluid flow, which can be expressed by the following mathematical equation 1.
[0049]
[0050] Here, u represents the velocity vector, p represents pressure, ρ represents density, ν represents the kinematic viscosity coefficient, and f represents the external force.
[0051] This embodiment adopts a strategy of rapidly approximating the equation by projecting the solution onto the weights of a neural network model, instead of iteratively solving the Navier-Stokes equation directly by numerical calculation.
[0052] 2. Training and Inference of Neural Network Models
[0053] The neural network model M is a low-resolution weather field W low and 3D terrain information G 3D Receives as input, high-resolution weather field W high_pred A function that outputs can be defined by the following mathematical formula 2.
[0054]
[0055] The above function M can be implemented based on a CNN (Convolutional Neural Network) or GAN (Generative Adversarial Network) architecture. CNN structures are strong in spatial pattern extraction, while GAN structures are advantageous for generating rich high-frequency spatial details.
[0056] 3. Physics-Informed Loss Function
[0057] The total loss function (L) defined in this embodiment total ) is the data loss term (220, L data ) and physical loss term (230, L physicsIt consists of a weighted sum of ) and can be defined by the following mathematical formula 3.
[0058]
[0059] Here, α and β are weighting coefficients that adjust the relative contribution of each loss term (220, 230).
[0060] Meanwhile, data loss term (220, L data ) is the correct high-resolution meteorological field (W) generated by CFD simulation. high GT) and AI predicted value (W high It can be calculated as the Mean Squared Error (MSE) between pred).
[0061]
[0062] And physical loss term (230, L physics ) is again composed of two constraints, which may include two constraints: a mass conservation law constraint term (231) and a momentum conservation law constraint term (232).
[0063] First, the mass conservation law constraint term (231) is the predicted velocity vector field (u pred The squared value of the divergence for ) can be defined by the following mathematical formula 5.
[0064]
[0065] This may be a discretized expression of the continuity equation (∇·u = 0) which forces the sum of the fluid inflow and outflow to be zero at all grid points within the predicted vector field.
[0066] Second, the momentum conservation law constraint term (232) can be defined by calculating the residual of the Navier-Stokes equation and then using Equation 6.
[0067]
[0068] Equation 5 enables the reproduction of a physically valid flow field even in complex urban turbulent environments by inducing learning to maintain physical equilibrium relationships between external forces such as fluid viscosity, pressure gradient, and Coriolis force.
[0069] Meanwhile, the neural network model (210) in FIG. 2 may further include a backpropagation optimization unit (240). The backpropagation optimization unit (240) is optimized to derive physically valid flow analysis results even in complex urban areas where data may be insufficient through the numerical constraints mentioned above.
[0070] As shown in FIG. 2, this embodiment is configured to provide a structure in which the AI model learns fluid dynamic characteristics such as the Venturi effect from data while being double-verified by physical laws by applying a physics-based loss function.
[0072] FIG. 3 is a diagram showing the spatial alignment process of macroscopic weather data and microscopic topographic information hierarchically according to an embodiment of the present invention.
[0073] Referring to FIG. 3, the multi-scale data integration process may consist of four layers: a macroscopic weather numerical forecast grid (310), a satellite-based 3D DSM layer (320), a microscopic weather computation grid (330), and a weather vector interpolation layer by vertical elevation (340).
[0074] The macroscopic numeric weather forecast grid (310) may be the top layer that accommodates wide-area weather data with a resolution of about 1.5 km provided by the weather agency LDAPS model. And the satellite-based 3D DSM layer (320) contains three-dimensional building and terrain shape information reconstructed based on satellite imagery or LiDAR data.
[0075] The above macroscopic numeric weather forecast grid (310) and the satellite-based 3D DSM layer (320) are processed by spatially overlaying. Unlike simple resolution enhancement techniques, the present invention performs data projection into a three-dimensional physical space that reflects the terrain curvature information of the DSM, thereby enabling the automatic generation of physically consistent boundary conditions even in areas where building density or terrain slope changes rapidly.
[0076] The microscopic weather computation grid (330) is a final computation grid that subdivides the entire domain into 5m units, and the weather vector interpolation by vertical altitude (340) performs the role of converting upper-level wind speed data from the Korea Meteorological Administration LDAPS into weather vectors by vertical altitude according to atmospheric boundary layer theory and inputting them as boundary conditions of the microscopic grid. This enables the realization of physical-based data projection across the entire three-dimensional space beyond planar grid subdivision.
[0078] FIG. 4 is a diagram illustrating a specific example of a micro-weather vector field calculated when an embodiment of the present invention is applied to an actual complex urban environment.
[0079] Referring to FIG. 4, it can be seen that in the narrow canyon section between two adjacent high-rise building objects (410), a Venturi acceleration zone (420) is clearly formed where the flow converges and the flow velocity increases rapidly. This is a direct result of mass conservation, which states that the flow velocity must increase as the cross-sectional area decreases in the continuity equation, and is a phenomenon that is not captured at all in existing 1.5 km class low-resolution models. Additionally, at the rear of the building, a stagnant zone (430) is formed where the wind speed decreases rapidly and a vortex is formed due to the phenomenon of flow separation.
[0080] The micro-weather vector (440) of 5m units that is finally output through the present invention is visualized as the magnitude of the wind speed (arrow length) and the direction (arrow direction) at each grid point, and the data can be used as a key parameter in various application fields such as urban air mobility (UAM) safe route design, construction site tower crane gust warning, and urban wind path design.
[0081] It can be seen that the microscopic weather vector (440) described in Fig. 4 is impossible to implement in conventional low-resolution weather models, but is precisely reproduced by the 5m resolution prediction technology of the present invention.
[0083] FIG. 5 is a diagram illustrating the data assimilation and feedback loop process that improves the prediction accuracy of the device of the present invention in real time, and FIG. 5 can be said to illustrate the configuration of the real-time data correction unit (140) of FIG. 1 in more detail.
[0084] Referring to FIG. 5, the real-time data correction unit (140) may be configured to include an error analysis unit (530) and an assimilation / correction processor (540).
[0085] The error analysis unit (530) can compare the AI predicted weather field (510) calculated in the first stage by the hybrid computation engine (130) with the actual values of IoT weather sensors (520) distributed throughout the city in real time, and calculate the residual between the predicted value and the actual value.
[0086] The analysis result of the error analysis unit (530) is input into an ensemble Kalman filter (EnKF)-based assimilation / correction processor (540) and is used to fine-tune the weights of the neural network to minimize the residual or to remove bias of the output value.
[0087] The final output value output by the above-mentioned assimilation / correction processor (540) may be microscopic weather information (550) in which CFD-based physical laws, learning data (AI), and real-time field observation values (IoT) are triple-combined / fused.
[0088] As such, the key feature of the feedback loop disclosed in this embodiment is that the correction effect at the observation point is physically and consistently propagated to the surrounding grid, which is fundamentally different from post-processing methods that simply correct only the value at the observation point.
[0090] FIG. 6 is a flowchart illustrating a method for generating ultra-high resolution local weather data according to an embodiment of the present invention.
[0091] Although FIG. 6 described below is described as being executed sequentially, this is merely an illustrative explanation of the technical concept of one embodiment of the present invention. In other words, a person skilled in the art to which one embodiment of the present invention belongs can modify and adapt it in various ways, such as changing the order described in FIG. 6 or executing one or more of the steps in parallel, without departing from the essential characteristics of one embodiment of the present invention; therefore, FIG. 6 is not necessarily limited to the chronological order depicted.
[0093] Referring to FIG. 6, it can be seen that the weather data generation method is performed by including a data collection and input step (S600); a CFD-based learning data generation step (S610); a physical information-based neural network learning step (S620); a real-time inference step (S630); a real-time correction and data assimilation step (S640); and an output and service provision step (S650).
[0095] Data collection and processing step (S600) ;
[0096] The data collection unit (110) of the device of the present invention begins by receiving input simultaneously from three major data sources.
[0097] First, when satellite and aerial images are received, they undergo a preprocessing process by the 3D terrain preprocessing unit (120) and are converted into 3D DSM (Digital Surface Model) terrain / building information. Specifically, the 3D terrain preprocessing unit (120) extracts building height H(x,y), terrain elevation Z(x,y), surface roughness R(x,y), and slope information from the DSM data and converts them into a 3D grid structure G₃D(x,y,z). At this time, the entire computational domain is divided into micro-grids of 5m units (5m × 5m × Δz), and boundary conditions are set such that non-penetration conditions and no-slip conditions are automatically applied to the building exterior wall and ground surface grids. Accordingly, the velocity vector at the corresponding boundary grid is set to converge to 0, so that physical constraints are applied to prevent the fluid from penetrating the building or passing through the ground surface. That is, as a non-penetration condition, “u = v = w = 0” (at the boundary grid).
[0098] Secondly, the Korea Meteorological Administration's LDAPS / GFS wide-area numerical weather forecast data is received at 1.5km grid units and prepared as wide-area weather data.
[0099] Third, real-time actual measurements are continuously collected from IoT / AWS sensors installed in the city center.
[0100] The three data streams mentioned above each play a different role in subsequent stages and are utilized in the corresponding stages.
[0102] CFD-based training data generation step (S610) ;
[0103] The above step (S610) is performed by the CFD simulator (132) of the hybrid computation engine (130). The CFD simulator (132) operates based on the Navier-Stokes equations and generates physically accurate high-resolution flow field data by iteratively calculating various wind direction and wind speed scenarios. At this time, a solution satisfying the conservation of mass and momentum of the fluid is calculated based on the Navier-Stokes equations, and the result is the correct answer dataset (W) for AI learning. highGT It is used as ).
[0104] Such CFD-based data may include flow patterns in complex urban environments, such as acceleration phenomena in building canyons or the formation of vortices behind buildings.
[0106] Physical information-based neural network learning step (S620);
[0107] In the above step (S620), the neural network model (210) is trained to receive wide-area weather data (low-resolution weather field) and a 3D terrain tensor as input and output a high-resolution weather vector field.
[0108] At this time, instead of a simple data error minimization method, a physics-based learning structure that includes residuals of the fluid continuity equation (∇·u = 0) and the Navier-Stokes equation in the loss function can be applied. That is, the total loss function is composed of a weighted sum of a data loss term (220) and a physics loss term (230), and the physics loss term evaluates whether the prediction vector field satisfies the laws of mass conservation and momentum conservation.
[0109] Through this, the neural network model goes beyond simple pattern learning and internalizes fluid dynamic characteristics.
[0111] Real-time inference step (S630) ;
[0112] In the actual operation phase, the neural network model (210) that has completed training is converted into an inference engine.
[0113] Therefore, in the real-time inference stage, CFD calculations are not performed, and only wide-area weather data and DSM topographic information are received in real time. A first prediction is performed within a predetermined time (e.g., 1 second) through a learned neural network model (210) to generate a high-resolution (5m) weather vector field. This result can be produced in the form of a three-dimensional vector field that includes micro-flow structures such as local acceleration zones or vortex sections between buildings in the city center.
[0114] Here, the above first prediction result may be a pure AI inference result before being corrected with actual data.
[0116] Real-time correction and data assimilation step (S640) ;
[0117] The real-time data correction unit (140) receives the first predicted weather vector field inferred in the real-time inference step described above and the actual data from the IoT weather sensor (520), and calculates a residual error by comparing the predicted value and the actual value. Then, it applies an algorithm such as an ensemble Kalman filter (EnKF) to fine-tune the weights of the neural network model or to remove output bias in order to minimize the residual error.
[0118] The corrected results are passed to the final output stage and packaged into an ultra-high-resolution weather information API.
[0120] Output and service provision step (S650) ;
[0121] The final 5m-class ultra-high resolution weather information is transmitted in real time to external systems in the form of an API.
[0122] As such, it can be seen that the present invention is performed in a closed-loop structure that collects satellite / meteorological agency / IoT data, pre-trains AI based on CFD, generates a 5m-level weather vector field through real-time AI inference, corrects errors using actual IoT measurements, and provides application services via API. Furthermore, through this closed-loop structure, prediction accuracy can be continuously improved over time.
[0124] Meanwhile, the aforementioned ultra-high-resolution weather information can be utilized in various application fields, such as urban air mobility (UAM) operation decision-making, construction site safety management, smart city wind path design, and digital twin-based environmental analysis.
[0125] Specifically, the scenarios in which the above-mentioned ultra-high-resolution weather information is applied may be as follows. Of course, it is not limited to this and can be applied to a wide variety of other scenarios.
[0126] First, vertiports located on the rooftops of high-rise buildings pose a high risk of accidents during the takeoff and landing of UAMs (Urban Air Mobility) due to unexpected building winds, vortices, and downwash caused by surrounding buildings. On the other hand, this embodiment can support a Go / No-go decision regarding takeoff and landing by calculating microscopic weather vectors at 5m intervals in real time for a 500m radius around the vertiport, and can provide optimal approach path guide information based on weather turbulence to the UAM control system in real time.
[0127] Second, by reflecting real-time 3D shape changes of buildings under construction at a construction site in the DSM, instantaneous gusts occurring at the top of tower cranes can be predicted. If the prediction results indicate that a specific wind speed threshold is exceeded, an immediate automatic alarm is issued and guidelines for stopping work are provided, thereby preventing safety accidents for workers at high-rise sites.
[0128] Third, it can be utilized for analyzing the urban heat island phenomenon and calculating natural ventilation efficiency for each building. Accordingly, it supports the establishment of optimized smart city policies that reduce the spread of fine dust and maximize heating and cooling energy efficiency through the design of urban wind corridors. Furthermore, by linking with a digital twin platform, it enables the pre-simulation of weather impacts during the urban development planning phase.
[0130] As described above, the present invention integrates satellite-based 3D terrain information, a CFD-based physical model, a neural network structure incorporating physical constraints, and real-time sensor assimilation technology, thereby having the effect of providing ultra-high-resolution urban weather information in real time while maintaining physical validity. In particular, it has technical differentiation in that it can significantly reduce computation time compared to existing CFDs, while deriving prediction results that do not violate physical laws, unlike simple statistical AI.
[0132] Although the present invention has been described with reference to illustrated embodiments as above, these are merely exemplary, and it will be obvious to those skilled in the art that various modifications, changes, and equivalent alternative embodiments are possible without departing from the gist and scope of the invention. Accordingly, the true technical scope of protection of the present invention should be determined by the technical spirit of the appended claims. Explanation of the symbols
[0133] 100: Device of the present invention 110: Data Collection Unit 120: 3D terrain preprocessing unit 130: Hybrid computing engine 132: CFD Simulator 140: Real-time data correction unit 150: Service Interface 210: Neural Network Model 220: Data Loss 230: Physical loss term 231: Constraint term of the law of conservation of mass 232: Constraint term of the law of conservation of momentum
Claims
Claim 1 A device for generating ultra-high-resolution weather data comprises: a data acquisition unit that receives three-dimensional digital surface model (DSM) data acquired from a satellite and low-resolution weather data from a wide-area numerical weather prediction model; a preprocessing unit that matches urban building and topographic information into a three-dimensional topographic grid based on the three-dimensional digital surface model; a hybrid computation engine that includes a Physics-Informed Neural Network trained on high-resolution flow field data for various wind direction and wind speed scenarios generated in advance through Computational Fluid Dynamics (CFD) simulation, and receives the low-resolution weather data and the three-dimensional grid information as input to calculate a high-resolution weather vector with a resolution of several meters or less in real time; and an output unit that provides the calculated high-resolution weather vector field to an external system. An ultra-high-resolution weather data generation device comprising, wherein the preprocessing unit sets fluid non-transparency conditions and no-slip boundary conditions for the 3D terrain grid, and the physics-informed neural network is trained to include residuals of the mass conservation equation and the momentum conservation equation in the physics-informed loss function of the neural network. Claim 2 An ultra-high-resolution weather data generation device according to claim 1, wherein the three-dimensional numerical surface model data includes data generated from satellite imagery, aerial LiDAR, or aerial orthophotos, and the preprocessing unit generates a three-dimensional geospatial tensor including building height, surface elevation, and surface roughness information using the data. Claim 3 delete Claim 4 In claim 1, the hybrid computation engine is an ultra-high-resolution weather data generation device that is learned by applying boundary conditions such that the velocity vector converges to 0 for building surfaces and ground surface grids. Claim 5 An ultra-high resolution weather data generation device according to claim 1, further comprising a real-time data correction unit that receives real-time wind speed and wind direction data from a weather sensor installed in an urban area, calculates the difference between a predicted value and a measured value calculated by the hybrid computation engine, and fine-tunes the weights of the physical information-based neural network or biases corrects the output value. Claim 6 In claim 1, the loss function of the physical information-based neural network is a data loss term (L data ) and physical loss term (L physics An ultra-high resolution weather data generating device comprising a weighted sum of ), wherein the physical loss term includes a mass conservation constraint term based on the continuity equation (∇·u = 0); and a momentum conservation constraint term based on the residual of the Navier-Stokes equation. Claim 7 A method for generating ultra-high-resolution weather data, comprising: a collection step of collecting three-dimensional digital surface model data and low-resolution weather data acquired from a satellite; a preprocessing step of generating a three-dimensional terrain grid tensor including building height and terrain information using the three-dimensional digital surface model, interpolating the low-resolution weather data into micro-grid units, and expanding the altitude-specific weather profile in the vertical direction to generate an initial weather field; a generation step of inputting the initial weather field and the three-dimensional terrain grid tensor into a physical information-based neural network to generate a high-resolution weather vector field with a resolution of several meters or less that satisfies mass conservation and momentum conservation constraints; and a providing step of providing the generated high-resolution weather vector field to an external system. Claim 8 A method for generating ultra-high resolution weather data according to claim 7, further comprising the step of receiving a measured value from a weather sensor, calculating an error between a predicted value generated in the generation step and the measured value, and correcting the predicted value. Claim 9 In claim 7, the above-mentioned physical information-based neural network is a method for generating ultra-high-resolution meteorological data, which is pre-trained using the results of computational fluid dynamics simulations performed based on the Navier-Stokes equations as training data. Claim 10 A method for generating super high-resolution weather data according to claim 7, wherein the learning of the above-mentioned physical information-based neural network is performed using a physical-based loss function comprising a data loss term and a physical loss term that satisfies the conservation of mass or momentum of the fluid. Claim 11 In claim 7, the method for generating ultra-high-resolution weather data, wherein the ultra-high-resolution weather data is generated in the form of a weather vector field with a spatial resolution of 5 meters or less for urban areas or complex terrain areas.
Citation Information
Patent Citations
Method and Apparatus for Evaluating Risk Impacts in Pedestrian Environment due to Hazardous Substance Dispersion in Urban Area
KR101943758B1
Urban Climate Impact Assessment method of Reflecting Urban Planning Scenarios and Analysis System using the same
KR1020170033711A
Dynamically Forecasting High Resolution Air Temperature in Real-Time Using Multiple Sources
US20240393498A1
