Weather forecast model optimization method based on multi-scale modeling and PINNs
By introducing multi-scale modeling and PINNs technology into the meteorological forecast model, combining multi-source data and physical constraints, the existing meteorological forecast model's insufficient accuracy and physical irrationality in extreme weather forecasts are solved, and high accuracy and physical rationality of meteorological prediction are achieved.
Patent Information
- Application Number
- CN202411952095.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-27
- Publication Date
- 2025-04-25
AI Technical Summary
The existing meteorological forecasting models have problems of insufficient accuracy and physical irrationality when simulating extreme weather phenomena, especially when cumulative errors in long-term sequence predictions lead to deviations from actual climatic conditions.
The meteorological forecast model optimization method based on multi-scale modeling and PINNs is adopted, and the model parameters are optimized to improve forecast accuracy and physical rationality through multi-source data fusion, graph neural network and joint model construction, combining atmospheric motion equations and physical constraint terms.
High-precision meteorological predictions are achieved, the accuracy of forecasting for extreme weather phenomena is improved, the physical rationality and reliability of forecast results are ensured, and the development and application of atmospheric occultation technology is supported.
Smart Images
Figure CN119885871A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of weather forecasting, and in particular to a weather forecasting model optimization method based on multi-scale modeling and PINNs. Background Art
[0002] As a core component of modern meteorological science, weather forecasting is of great importance. Accurate and timely weather forecasts have a profound impact on social and economic activities and all aspects of people's lives. Whether it is agricultural planning, aviation scheduling or daily travel, reliable weather information provides a solid foundation for decision-making. With the advancement of science and technology, we have witnessed a shift from traditional observation and empirical judgment to more complex and technology-driven methods.
[0003] Currently, Numerical Weather Prediction (NWP) technology and data-driven large model technology constitute the main pillars of meteorological forecasting. NWP predicts future weather conditions by solving a set of basic physical equations that describe the behavior of the atmosphere. This approach relies on high-resolution global or regional models, such as the High Resolution Ensemble Prediction System (HRES) of the European Centre for Medium-Range Weather Forecasts (ECMWF), which can provide detailed and accurate short-term weather forecasts. However, in some cases, such as pressure layers of 50hPa and below, these emerging data-driven large models, such as GraphCast, may not perform as well as traditional high-resolution numerical weather prediction models. GraphCast uses advanced technologies such as graph neural networks to try to capture spatial relationships and speed up the calculation process, but the accuracy at specific vertical levels needs to be improved.
[0004] The performance of these models is particularly important when it comes to extreme events, such as tropical cyclones, atmospheric rivers, and extreme temperature phenomena. Although existing NWP models are already quite reliable under conventional conditions, they still have room for improvement for extreme situations with high uncertainty and nonlinear characteristics. For example, in the process of simulating the development of severe convective storms, due to the complexity and scale differences of cloud microphysical processes, even the most advanced models may not be able to fully accurately reproduce the actual situation. In addition, the physical unreasonableness problems in long-term series predictions are also worthy of attention. As time goes on, the cumulative error of the model increases, which may cause the forecast results to deviate from the actual climate conditions, thus affecting the effectiveness of long-term planning and risk assessment.
[0005] In recent years, deep learning technology has gradually become one of the research hotspots in the field of weather forecasting. This type of method is good at processing large amounts of unstructured data and can automatically extract features, so it is very suitable for application in meteorological big data analysis. However, how to effectively integrate traditional knowledge based on physical laws into a purely data-driven framework remains an urgent problem to be solved. On the one hand, deep learning models lack a clear physical basis, which makes it difficult to explain the working mechanism within the model; on the other hand, models trained solely on historical data are difficult to cope with new scenarios that have never been seen before. Summary of the invention
[0006] The present invention aims to provide a weather forecast model optimization method based on multi-scale modeling and PINNs, which can achieve high-precision weather forecasts and help promote the development and application of atmospheric occultation technology.
[0007] The basic scheme provided by the present invention is: a weather forecast model optimization method based on multi-scale modeling and PINNs, comprising the following steps:
[0008] Step 1: Collect multi-source data and pre-process them;
[0009] Step 2, perform data fusion on the preprocessed multi-source data to obtain a basic data set;
[0010] Step 3, based on the basic data set, building a weather forecast model, including building a graph neural network and using meteorological parameters as node attributes, using the graph neural network for information dissemination and feature extraction, and then building a joint model as a weather forecast model based on the atmospheric motion equation; and using a loss function with specific constraints to optimize the model parameters; the specific constraints include atmospheric thermodynamic constraints and kinetic constraints;
[0011] Step 4: Use the weather forecast model to make weather forecasts and verify the forecast results.
[0012] Further, in step 1, the multi-source data includes satellite observation data, ground observation data, and radar data; the satellite observation data includes TM satellite data and NCEP data.
[0013] Furthermore, the data fusion includes: for the satellite observation data, converting the data formats of the TM satellite data and the NCEP data into data frames, and fusing the data frames.
[0014] Furthermore, when constructing the graph neural network, the ERA5 data is converted from pressure coordinates to the initial conditions in the σ coordinate system used by the model through an encoder; and based on the graph neural network, meteorological data based on graph representation at different levels of division are obtained.
[0015] Furthermore, when constructing a joint model, the following steps are included: setting an encoder, a dynamic core, a learning physics module, an ODE solver and a decoder; inputting meteorological data, noise data and mandatory item data represented by a graph into the encoder, the encoder encodes the data into a model state, and the dynamic core and the learning physics module process the data accordingly, the ODE solver promotes the evolution of the model state over time according to the processing results of the dynamic core and the learning physics module, thereby obtaining a new model state, and the decoder outputs the joint model based on the latest model state.
[0016] Further, the model state is input into a dynamic core, which is used to simulate large-scale fluid motion and thermodynamic processes under the influence of gravity and Coriolis forces, and output dynamic physical tendencies.
[0017] Furthermore, the noise data and the forced term data are input into a learning physics module, which is used to predict the impact of unresolved processes on the simulated field and output the rate of change or physical tendency of the unresolved processes; the unresolved processes include cloud formation, radiation transfer, precipitation and grid-scale dynamics.
[0018] Furthermore, the unanalyzed process is represented by a semi-empirical parameterized equation.
[0019] Furthermore, when the decoder outputs the joint model based on the latest model state, it includes direct output and conversion output; the direct output is: directly outputting the joint model as a weather forecast model; the conversion output is: outputting the joint model and converting it from σ coordinates to pressure coordinates for use in subsequent time steps.
[0020] Furthermore, after obtaining the new model state, the new model state is used as the input of the encoder and is then processed by the dynamic core, the learning physics module, and the ODE solver to continue advancing the time evolution of the model until the desired forecast end time is reached.
[0021] The working principle and advantages of the present invention are:
[0022] The present invention is based on a weather forecast model optimization method based on multi-scale modeling and PINNs, which combines multi-source data, multi-level grid structure, Graph network architecture and generative diffusion model, PINNs (physical information embedded network) and other technologies, and can effectively improve the accuracy and physical rationality of weather forecasts and achieve high-precision weather forecasts. The key points are:
[0023] First, this solution improves the quality and consistency of input data through multi-source data fusion technology, providing a reliable basis for model training. In particular, compared with existing weather forecasting solutions, the selected data source is more distinctive - the fusion of TM satellite data and NCEP data, the two can verify and calibrate each other, can improve the reliability of the overall data, and help improve the accuracy of subsequent weather forecasts.
[0024] Second, this solution sets up a multi-level grid structure and utilizes the Graph network architecture. Through the multi-level, multi-resolution grid structure, the model can better understand phenomena at different scales and improve the spatial consistency of forecasts. In addition, the Graph network architecture, i.e., the graph neural network, can effectively process data with spatial structure and capture local and global features at different scales, which helps to improve forecast accuracy.
[0025] Third, this scheme adopts a hybrid modeling strategy and creates a new paradigm of "physics-data" fusion. In this mode, it is possible to combine determinism and randomness, and through the constraints of physical information, ensure the correctness of the forecast results in the physical sense, and generate high-quality forecast effects that conform to the real data distribution. The forecast results are more in line with physical laws than the pure large model method. For example, for the forecast of extreme weather, it is difficult for the pure large model to give accurate short-term extreme forecast results because it does not conform to physical constraints. In addition, for physical processes that are not fully understood or difficult to parameterize, this scheme combines semi-empirical parameterization methods to represent them, which can effectively fill the observation gaps. The forecast results obtained are more accurate and reliable than ordinary meteorological forecasts. For the atmospheric occultation detection constellation, this hybrid paradigm can simulate atmospheric conditions more accurately, thereby optimizing satellite orbit design and observation strategies, so that the coverage and spatial distribution of occultation detection can be optimized, which helps to promote the development and application of atmospheric occultation technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 It is a method flow diagram of an embodiment of a method for optimizing a weather forecast model based on multi-scale modeling and PINNs of the present invention;
[0027] Figure 2 It is a first schematic diagram of the model building process of an embodiment of a weather forecast model optimization method based on multi-scale modeling and PINNs of the present invention;
[0028] Figure 3 This is the second schematic diagram of the model building process of an embodiment of the meteorological forecast model optimization method based on multi-scale modeling and PINNs of the present invention. DETAILED DESCRIPTION
[0029] The following is a further detailed description through specific implementation methods:
[0030] The embodiment is basically as shown in the attached Figure 1 As shown: The weather forecast model optimization method based on multi-scale modeling and PINNs includes the following steps:
[0031] Step 1: Collect multi-source data and preprocess them.
[0032] In this step, the multi-source data includes satellite observation data, ground observation data, and radar data; the satellite observation data includes TM satellite data, NCEP data, and ERA5 data. The ground observation data refers to atmospheric condition information collected from fixed or mobile stations on the ground. These stations may include but are not limited to airports, weather stations, ocean buoys, etc., and specifically include the following elements: temperature, humidity, air pressure, wind speed and direction, precipitation, cloud cover and cloud height, visibility. The radar data includes reflectivity factor, radial velocity, spectral width, etc.
[0033] Among them, TM satellite data includes multiple bands, which can be used to extract information about vegetation, water bodies, soil, etc. TM data usually has a high spatial resolution (30 meters), which is suitable for detailed classification of objects and change detection. NCEP data contains atmospheric variables (such as temperature, humidity, wind speed, etc.), surface parameters, and information on the state of the upper atmosphere. This type of data has a relatively low spatial resolution, but good temporal continuity, and is suitable for large-scale climate research and weather forecasting.
[0034] In this embodiment, the resolution of TM satellite data is 0.25°×0.25°, and TM satellite data is L2 product data inverted based on occultation payload. By combining such TM satellite data and NCEP data, the two can verify and calibrate each other, which can improve the reliability of the overall data and help improve the accuracy of subsequent weather forecasts. Combined with ERA5 data, ERA5 data can supplement the data that cannot be inverted from TM satellite data, ensuring the completeness of the data source.
[0035] The preprocessing includes data cleaning to filter out invalid data. In this embodiment, the preprocessing also includes radiation correction, geometric correction, atmospheric correction, cloud detection and classification for satellite observation data.
[0036] Step 2: perform data fusion on the preprocessed multi-source data to obtain a basic data set.
[0037] The data fusion includes: for the satellite observation data, converting the data formats of the TM satellite data and the NCEP data into data frames, and fusing the data frames. In this embodiment, this operation is completed based on the pandas tool of python.
[0038] Step 3: Based on the basic data set, a weather forecast model is constructed, including constructing a graph neural network and using meteorological parameters as node attributes, using the graph neural network for information propagation and feature extraction, and then constructing a joint model as a weather forecast model based on the atmospheric motion equation; and using a loss function with specific constraints to optimize the model parameters, such as Figure 2 As shown; the specific constraint items include atmospheric thermodynamic constraint items and kinetic constraint items. In this embodiment, the specific constraint items correspond to the learning physics module.
[0039] In this embodiment, the atmospheric motion equations used include the continuity equation, the momentum equation (ie, a simplified form of the Navier-Stokes equation), the thermodynamic energy equation, and the state equation.
[0040] Among them, the continuity equation expresses the principle of conservation of mass. For incompressible fluids, it ensures that the mass flowing into a volume is equal to the mass flowing out of the volume plus or minus the rate of change of mass in the volume; in atmospheric science, air can be considered as an approximately compressible fluid. The momentum equation is used to describe the change of velocity of air parcels or fluid elements with time and space. The thermodynamic energy equation describes the process of conversion between the internal energy, kinetic energy and other forms of energy of a system; in atmospheric science, it is used to track the change of temperature with time and space. The state equation connects the state variables of a substance, such as temperature, pressure and density.
[0041] These equations interact with each other to simulate various phenomena in the atmosphere (from daily weather patterns to long-term climate trends), and based on the above atmospheric motion equations, they can help predict future atmospheric conditions.
[0042] Specifically, when constructing the graph neural network, the ERA5 data is converted from pressure coordinates to the initial conditions of the σ coordinate system used by the model through an encoder; and based on the graph neural network, the meteorological data based on graph representation at different levels of division are obtained. The division basis of the different levels of division includes the height basis, so as to facilitate the encoding of different scale features.
[0043] like Figure 3As shown, when constructing a joint model, the following steps are included: setting an encoder, a dynamic core, a learning physics module, an ODE solver (i.e., an implicit-explicit ordinary differential equation solver) and a decoder; inputting meteorological data, noise data and mandatory data represented by a graph into the encoder, the encoder encodes the data into a model state, and the dynamic core and the learning physics module process the data accordingly, the ODE solver promotes the evolution of the model state over time according to the processing results of the dynamic core and the learning physics module, thereby obtaining a new model state, and the decoder outputs the joint model based on the latest model state. The noise data is used to simulate uncertainty or complexity in the real world to improve the generalization ability of the model. The mandatory data includes boundary conditions, initial conditions, source and sink terms, etc. The mandatory data helps the model to accurately reproduce the physical phenomena of the real world, thereby improving the accuracy of the forecast.
[0044] Specifically, the model state is input into a dynamic core, which is used to simulate large-scale fluid motion and thermodynamic processes under the influence of gravity and Coriolis forces, and output dynamic physical tendencies. In this embodiment, the atmospheric motion equations constitute the dynamic core to promote time evolution and calculate changes in atmospheric variables.
[0045] The noise data and the forced term data are input into a learning physics module, which is used to predict the impact of unresolved processes on the simulated field and output the rate of change or physical tendency of the unresolved processes; the unresolved processes include cloud formation, radiation transfer, precipitation, and grid-scale dynamics.
[0046] The simulated fields refer to the spatial distribution maps obtained after numerical simulation of atmospheric state variables (such as temperature, humidity, wind speed, etc.). These simulated fields represent the atmospheric conditions calculated by physical equations and parameterization schemes within a specific time and space range according to initial conditions and boundary conditions.
[0047] The unresolved process is represented by a semi-empirical parameterized equation. Taking radiation transfer as an example, since the propagation paths of solar shortwave radiation and earth longwave radiation in the atmosphere are affected by factors such as aerosols and clouds, appropriate parameterization processing is required. The parameterized equations corresponding to radiation transfer include estimation equations based on the two-stream approximation method and equations based on the discrete coordinate method (DISORT). Among them, the former simplifies the radiation flux in the vertical direction into two components, upward and downward, and assumes a uniform distribution in the horizontal direction; it is suitable for roughly estimating the net radiation flux. The latter is based on the discrete coordinate method for precise calculations, which can solve the radiation problem of arbitrary angles of incidence in multi-layer media, and is particularly suitable for situations involving multiple scattering sources.
[0048] Taking cloud formation as an example, the formation, development and dissipation of clouds involve complex phase change processes, such as condensation, freezing, melting, etc. These processes affect the generation and development of precipitation. The parameterized equations used include dual-parameter equations and multi-category particle equations. Among them, the former simultaneously tracks the number concentration and mass mixing ratio of water vapor and liquid / solid water (ie, dual parameters), which can more efficiently obtain the cloud formation state. The latter divides different types of cloud particles (such as cloud droplets, snow crystals, and graupel particles) more finely, and establishes independent continuity equations for each particle, which can better reproduce the dynamic and thermodynamic characteristics inside the cloud with higher accuracy.
[0049] Since these unresolved processes often occur at a scale smaller than the model resolution or are too complex to be solved directly, a parameterization scheme is introduced here through learning the physics module to approximate the impact of these unresolved processes. This can simplify the sub-grid scale processes into empirical formulas that depend on large-scale variables, so as to accurately describe the physical processes.
[0050] The decoder outputs the joint model based on the latest model state, including direct output and conversion output; the direct output is: directly outputting the joint model as a weather forecast model; the conversion output is: outputting the joint model and converting it from σ coordinates to pressure coordinates for use in subsequent time steps.
[0051] After the new model state is obtained, it is used as the input of the encoder and is processed by the dynamic core, the learning physics module and the ODE solver to continue to advance the time evolution of the model until the desired forecast end time is reached.
[0052] Step 4: Use the weather forecast model to make weather forecasts and verify the forecast results.
[0053] Specifically, the forecast results of the weather forecast include precipitation, temperature, wind speed, wind direction, and humidity. In this embodiment, the weather forecast verification library meteva is used to perform verification and evaluation of the forecast results.
[0054] The present embodiment provides a weather forecast model optimization method based on multi-scale modeling and PINNs, which can achieve high-precision weather forecasting and help promote the development and application of atmospheric occultation technology.
[0055] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field are aware of all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement the scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for the ordinary technicians in the relevant field to implement this application. It should be pointed out that for the technicians in this field, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent.
Claims
1. A meteorological forecast model optimization method based on multi-scale modeling and PINNs, characterized in that: The following steps are involved: Step 1: Collect multi-source data and pre-process them; Step 2, perform data fusion on the preprocessed multi-source data to obtain a basic data set; Step 3, based on the basic data set, building a weather forecast model, including building a graph neural network and using meteorological parameters as node attributes, using the graph neural network for information dissemination and feature extraction, and then building a joint model as a weather forecast model based on the atmospheric motion equation; and using a loss function with specific constraints to optimize the model parameters; the specific constraints include atmospheric thermodynamic constraints and kinetic constraints; Step 4: Use the weather forecast model to make weather forecasts and verify the forecast results.
2. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 1, characterized in that: In step 1, the multi-source data includes satellite observation data, ground observation data, and radar data; the satellite observation data includes TM satellite data and NCEP data.
3. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 2, characterized in that: The data fusion includes: for satellite observation data, converting the data formats of TM satellite data and NCEP data into data frames, and fusing the data frames.
4. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 1, characterized in that: When constructing the graph neural network, the ERA5 data is converted from pressure coordinates to the initial conditions in the σ coordinate system used by the model through an encoder; and based on the graph neural network, meteorological data based on graph representation at different levels of division are obtained.
5. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 4, characterized in that: When constructing a joint model, the following steps are included: setting an encoder, a dynamic core, a learning physics module, an ODE solver and a decoder; inputting meteorological data, noise data and mandatory item data represented by a graph into the encoder, the encoder encodes the data into a model state, and the dynamic core and the learning physics module process the data accordingly, the ODE solver promotes the evolution of the model state over time according to the processing results of the dynamic core and the learning physics module, thereby obtaining a new model state, and the decoder outputs the joint model based on the latest model state.
6. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 5, characterized in that: The model states are input into a dynamic core, which is used to simulate large-scale fluid motion and thermodynamic processes under the influence of gravity and Coriolis forces, and outputs dynamic physical tendencies.
7. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 5, characterized in that: The noise data and the forced term data are input into a learning physics module, which is used to predict the impact of unresolved processes on the simulated field and output the rate of change or physical tendency of the unresolved processes; the unresolved processes include cloud formation, radiation transfer, precipitation, and grid-scale dynamics.
8. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 7, characterized in that: The unresolved process is represented by semi-empirical parameterized equations.
9. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 5, characterized in that: The decoder outputs the joint model based on the latest model state, including direct output and conversion output; the direct output is: directly outputting the joint model as a weather forecast model; the conversion output is: outputting the joint model and converting it from σ coordinates to pressure coordinates for use in subsequent time steps.
10. The method for optimizing a weather forecast model based on multi-scale modeling and PINNs according to claim 5, characterized in that: After the new model state is obtained, it is used as the input of the encoder and is processed by the dynamic core, the learning physics module and the ODE solver to continue to advance the time evolution of the model until the desired forecast end time is reached.