Continuous-time weather forecasting method, forecasting device, electronic device, and storage medium

Abstract

The application discloses a continuous-time weather forecast method, a weather forecast device, an electronic device and a storage medium. The weather forecast method comprises the following steps: acquiring a training sample pair composed of a historical time weather state field and a future time weather state field, constructing a continuous-time dynamic path; determining a target change rate field of the weather state in a unit normalized time based on the continuous-time dynamic path, taking the target change rate field as a supervision signal, and training a preset velocity field parameter model; acquiring an initial weather state field to be forecasted, taking the initial weather state field as an evolution starting point, performing segmented state deduction on a continuous time scale based on a dynamic evolution law depicted by the trained velocity field model, adapting a business required time resolution through a time scale mapping, and generating a continuous weather forecast result with an arbitrary time resolution. The application breaks through the limitation of discrete fixed time steps, and realizes continuous, stable and high-precision weather forecast with an arbitrary time resolution.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a continuous-time meteorological forecasting method, forecasting device, electronic equipment, and storage medium. Background Technology

[0002] Highly timely forecasts of meteorological elements are a crucial technological foundation for disaster prevention and mitigation, energy dispatching, aviation and maritime transport, and public services. With the continuous accumulation of global reanalysis and operational forecast data, deep learning-based meteorological forecasting methods have rapidly developed. Compared to traditional numerical weather prediction methods, deep learning-based methods offer significant advantages in computational efficiency, system deployment, and real-time response capabilities. Hourly and even higher temporal resolution meteorological element forecasts are increasingly becoming important operational applications, placing higher demands on the model's continuous forecasting capabilities at arbitrary time scales.

[0003] Existing deep learning-based weather forecasting methods typically employ discrete-time modeling, representing the evolution of weather conditions between adjacent fixed time steps as a mapping relationship, and generating multi-term forecast results through a step-by-step recursive approach. These methods often require model retraining when the time step changes, making it difficult to support outputs at arbitrary time resolutions. Furthermore, due to the highly nonlinear and multi-scale characteristics of atmospheric systems, discrete recursive forecasts are highly sensitive to model errors. Small deviations in a single-step forecast can be amplified over multiple recursive iterations, leading to rapid error accumulation and consequently affecting the stability and reliability of medium- and long-term, high-term forecast results.

[0004] Furthermore, hourly forecast models typically rely on reanalysis data as a source of supervision during training. Reanalysis data is generated by periodically assimilating observational data, and there is an objective degree of temporal discontinuity between different assimilation time windows. When discrete-time modeling is performed directly based on hourly data, the model is prone to mislearning these non-physical jumps as real meteorological evolution characteristics, thereby introducing instability factors into high temporal resolution forecasting and further limiting the model's generalization ability at arbitrary time scales.

[0005] Meanwhile, some deep learning-based weather forecasting methods incorporate modeling concepts from image or video generation, using random noise as the initial state of the generation process and learning the mapping relationship from noise distribution to meteorological data distribution through conditional denoising models. For example, diffusion-based methods are widely used in weather forecasting. These methods use random noise as the starting point and generate the target meteorological field through progressive denoising. However, these methods typically do not use the actual meteorological state as the evolutionary starting point, but rather rely on random noise to gradually generate the target meteorological field. This lack of explicit physical initial state constraints is detrimental to characterizing the continuous evolution of meteorological elements over time, especially in high temporal resolution and continuous-time forecasting scenarios, easily compromising the temporal consistency and stability of forecast results. Therefore, in the field of deep learning-based weather forecasting, there is an urgent need for a weather element forecasting scheme that can use real weather conditions as the starting point for evolution, suppress error accumulation from the modeling mechanism, avoid non-physical jump interference in reanalysis data, and support continuous output at any time resolution. This would improve the stability, reliability, and operational applicability of high-time-sensitivity data-driven forecasting systems. Summary of the Invention

[0006] The purpose of this invention is to provide a continuous time weather forecasting method, forecasting device, electronic device and storage medium, which at least partially solves one of the aforementioned technical problems.

[0007] To achieve the above objectives, a first aspect of the present invention provides a continuous-time weather forecasting method, the weather forecasting method comprising: acquiring training sample pairs consisting of historical weather state fields and future weather state fields; constructing a continuous-time dynamic path with the historical weather state field as the evolution starting point and the future weather state field as the evolution ending point, wherein the continuous-time dynamic path is used to characterize the evolution process of the physical field from the historical weather state field to the future weather state field; determining a target rate of change field of the weather state within a unit normalized time based on the dynamic path, and using the target rate of change field as a supervision signal to train a preset velocity field parameter model for fitting the atmospheric evolution law; acquiring an initial weather state field to be forecasted, using the initial weather state field as the evolution starting point, and performing segmented state deduction on a continuous time scale based on the dynamic evolution law characterized by the trained velocity field model, and generating a continuous weather forecast result with arbitrary time resolution by adapting the time scale mapping to the time resolution required for the operation.

[0008] According to one embodiment of the present invention, constructing a continuous time dynamic path with the historical meteorological state field as the starting point and the future meteorological state field as the ending point includes: within a preset standard normalized time interval, linearly weighting the historical meteorological state field and the future meteorological state field as they change with normalized time, and superimposing a random disturbance field of a preset magnitude to form a continuous dynamic path connecting the real historical state and the real future state along the time axis; calculating the instantaneous change of the continuous dynamic path with respect to normalized time to obtain the target rate of change field, wherein the target rate of change field is equivalent to the grid-by-grid difference between the future meteorological state field and the historical meteorological state field.

[0009] According to one embodiment of the present invention, the training of the preset velocity field parameter model adopts a two-stage progressive training strategy including a pre-training stage and a fine-tuning stage. In the pre-training stage, the training sample pairs are collected based on a first time interval to perform preliminary training on the velocity field parameter model in order to learn the dynamic evolution structure of the atmospheric system. In the fine-tuning stage, the training sample pairs are collected based on a second time interval to perform parameter fine-tuning on the pre-trained velocity field parameter model. The second time interval is shorter than the first time interval so that the velocity field parameter model can adapt to the high temporal resolution forecasting requirements.

[0010] According to an embodiment of the present invention, the fine-tuning of the pre-trained velocity field parameter model in the fine-tuning stage includes: using a learning rate smaller than that in the pre-training stage, freezing the low-level feature encoding layer in the velocity field parameter model, and updating only the high-level time modulation parameters and dynamic evolution mapping parameters of the model.

[0011] According to an embodiment of the present invention, the segmented state extrapolation on a continuous time scale includes: dividing the entire forecast duration into multiple standard time unit intervals corresponding to the model pre-training time scale, performing normalized time progression within each standard time unit interval; after completing the state extrapolation of a standard time unit interval, resetting the normalized time variable used to characterize the evolution progress to the initial value, and using the final state meteorological field obtained from the extrapolation of the current interval as the starting state field of the next interval, repeating the above process until the preset forecast lead time is reached.

[0012] According to one embodiment of the present invention, during the segmented state deduction process, when advancing the normalized time in a single standard time unit interval, the step scaling factor in the deduction process is adjusted according to the proportional relationship between the target output time resolution specified by the service and the standard time unit interval, and segmented iterative deduction is performed using a single-step advancement format or a multi-level advancement format to adapt to the output requirements of any time resolution.

[0013] According to an embodiment of the present invention, when constructing the velocity field parameter model, a deep network structure with time-conditional modulation function is adopted. The deep network structure includes: an encoding module for encoding the input meteorological state field into a latent feature representation; a time-modulated backbone network for performing deep feature extraction on the latent feature representation, wherein the backbone network introduces a time-adaptive control mechanism in its internal normalization stage to convert the normalized time variable into a feature control parameter, and performs scale scaling and offset correction on the normalized features of the intermediate layer of the network through the control parameter to characterize the dynamic evolution behavior under different propulsion positions; and a decoding module for reconstructing the feature representation output by the backbone network into a rate of change field, wherein the rate of change field has the same number of variable channels and spatial dimension as the input meteorological state field.

[0014] According to an embodiment of the present invention, training the velocity field parameter model using the target rate of change field as a supervisory signal includes: updating the model parameters by minimizing the grid-by-grid mean square difference between the rate of change field predicted by the velocity field parameter model and the target rate of change field; during the training process, the normalized time variable is randomly sampled within a preset standard interval or biased sampled according to a preset distribution strategy to enhance the model's adaptability to the dynamic characteristics of different evolutionary stages.

[0015] According to an embodiment of the present invention, the meteorological state field includes at least one of the following multi-source geophysical data: near-surface meteorological observation elements, upper-air meteorological elements of different isobaric surfaces, static underlying surface geographic features, and time-location coding features characterizing daily and annual cycle changes; the method further includes: performing spatial grid unified alignment and variable standardization processing on the acquired raw multi-source data, and organizing each variable into a multi-dimensional tensor format containing channel dimension and spatial dimension.

[0016] To achieve the above objectives, a second aspect of the present invention provides a continuous-time weather forecasting device, comprising: a construction unit, configured to acquire training sample pairs consisting of historical weather state fields and future weather state fields, and construct a continuous-time dynamic path with the historical weather state field as the evolution starting point and the future weather state field as the evolution ending point, wherein the continuous-time dynamic path is used to characterize the evolution process of the physical field from the historical weather state field to the future weather state field; a training unit, configured to determine the target rate of change field of the weather state within a unit normalized time based on the dynamic path, and use the target rate of change field as a supervision signal to train a preset velocity field parameter model for fitting the atmospheric evolution law; and a deduction unit, configured to acquire the initial weather state field to be forecasted, use the initial weather state field as the evolution starting point, and perform segmented state deduction on a continuous time scale based on the dynamic evolution law characterized by the trained velocity field model, and generate continuous weather forecast results with arbitrary time resolution by adapting the time scale mapping to the time resolution required for the operation.

[0017] To achieve the above objectives, a third aspect of the present invention provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform a continuous-time weather forecasting method as described in any one of the first aspects.

[0018] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the continuous-time weather forecasting method as described in any one of the first aspects.

[0019] It is evident that the present invention has the following advantages compared to the prior art: (1) This invention models the meteorological state evolution process as a continuous-time dynamic system, generates prediction results at any time point through an ordinary differential equation (ODE) solver, enabling the model to naturally support output at any time resolution, and alleviates the rapid accumulation of errors in the discrete recursive process from the modeling mechanism.

[0020] (2) The present invention proposes a data-to-data dynamic flow matching method, which directly takes the real meteorological state as the starting point of evolution and the real future meteorological state as the target endpoint, constructs a continuous probability path and learns the velocity field, so that the model can directly learn the real dynamic evolution law of the meteorological system, avoids the introduction of random noise, and enhances the time continuity and physical consistency from the modeling mechanism.

[0021] (3) The present invention adopts a two-stage training mechanism, which first pre-trains on data with a longer time interval (such as 6 hours or 12 hours) and then fine-tunes on data hourly. This reduces the interference of the time discontinuity of the reanalysis data on the model learning from the training strategy level, and improves the stability and reliability of the model in high time resolution prediction scenarios.

[0022] (4) This invention uses continuous-time dynamic system modeling and ODE solving to output prediction results at any time resolution under the same model framework without retraining the model, which significantly improves the model’s adaptability and deployment flexibility in multiple business scenarios.

[0023] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a continuous-time weather forecasting method according to an exemplary embodiment of the present invention; Figure 2 This is a logical architecture diagram of continuous-time weather forecasting based on dynamic flow matching according to an exemplary embodiment of the present invention; Figure 3 This is a schematic diagram illustrating a two-stage training process according to an exemplary embodiment of the present invention; Figure 4 This is a schematic diagram of a reasoning process according to an exemplary embodiment of the present invention; and Figure 5 This is a schematic block diagram of a continuous time weather forecasting device according to an exemplary embodiment of the present invention. Detailed Implementation

[0025] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0026] To address the three major challenges in current deep learning-based weather forecasting—limited discrete-time resolution, non-physical abrupt changes in reanalysis data, and lack of physical constraints on noise generation paradigms—this application proposes a continuous-time weather forecasting method based on dynamic flow matching.

[0027] Existing data-driven methods mostly employ a discrete modeling paradigm with a fixed time step, generating forecast results through recursive mapping. This makes it difficult for the model to support outputs at arbitrary time resolutions, and single-step errors are easily amplified nonlinearly in long-term recursion. At the same time, when training based on hourly reanalysis data, non-physical time jumps caused by assimilation window switching are easily mislearned by the model as spurious dynamic features. Furthermore, generative modeling based on noise-to-data lacks real physical initial state constraints, making it difficult to maintain the temporal consistency of meteorological field evolution.

[0028] To address the aforementioned issues, the core concept of this application lies in constructing a data-to-data dynamic flow matching mechanism. This scheme abandons the traditional random noise starting point and directly uses real meteorological states as the evolutionary starting point, modeling the meteorological element prediction process as a continuous-time dynamic system. By learning the instantaneous evolution laws between states, it suppresses error accumulation from a mechanistic perspective. Furthermore, this application introduces an ordinary differential equation solution framework, utilizing a time-series progressive solver to perform state derivation at any time point. This allows a single model to flexibly support hourly or even higher time-density forecast outputs without retraining, breaking through the dependence of discrete modeling on a fixed time step. Simultaneously, this application designs a two-stage anti-interference training strategy: first, pre-training is performed on data with longer time intervals to reduce the interference of non-physical time discontinuities caused by assimilation window switching in reanalysis data; then, supervised fine-tuning is performed on data with shorter time intervals to further improve the model's continuity, consistency, and stability at high temporal resolution.

[0029] Through the above technical solution, this application can achieve continuous, stable, and high-precision prediction of meteorological elements at any time scale within a single model framework, possessing both excellent engineering feasibility and operational value. The following will elaborate on this solution with specific embodiments.

[0030] Figure 1 A continuous-time weather forecasting method is shown, such as Figure 1 As shown, this weather forecasting method includes: Step S110: Obtain training sample pairs consisting of historical meteorological state fields and future meteorological state fields, and construct a continuous time dynamic path with the historical meteorological state field as the evolution starting point and the future meteorological state field as the evolution ending point, wherein the continuous time dynamic path is used to characterize the evolution process of the physical field from the historical meteorological state field to the future meteorological state field.

[0031] Specifically, embodiments of the present invention can obtain a large number of paired meteorological state field samples from reanalysis data or numerical weather prediction products. Each pair of samples consists of a meteorological state field at a historical moment and a meteorological state field at a future moment. For example, the global meteorological field at 00:00 on January 1, 2020, is extracted from ERA5 reanalysis data as the historical state, and the global meteorological field at 06:00 on the same day is extracted as the future state; the two constitute a training sample pair.

[0032] After acquiring sample pairs, this embodiment of the invention does not directly learn the black-box generation process from historical states to future states as in traditional methods. Instead, it constructs a continuous evolution path between the two. The starting point of this path is the actual historical weather state, and the ending point is the actual future weather state. The path itself describes the complete evolution trajectory of the physical field from the starting point to the ending point on the time axis. In this way, the system transforms the discrete two-point mapping problem into a continuous path learning problem, enabling the model to obtain the evolution information of the weather state at any intermediate time.

[0033] To facilitate unified data processing, this embodiment of the invention normalizes the path evolution interval to the [0,1] interval, where t=0 corresponds to a historical moment and t=1 corresponds to a future moment. Under this normalization framework, the path state at any intermediate moment t∈[0,1] can be represented as a linear combination of the historical state and the future state, that is, the state smoothly transitions from the starting point to the ending point over time.

[0034] Therefore, through this step, the embodiments of the present invention establish a data-to-data evolution path starting from real meteorological conditions, which is fundamentally different from the traditional generative modeling paradigm that starts from random noise.

[0035] Step S120: Based on the dynamic path, determine the target rate of change field of meteorological state within a unit normalized time, and use the target rate of change field as a supervision signal to train a preset velocity field parameter model for fitting the atmospheric evolution law.

[0036] Specifically, for the continuous dynamic path constructed in step S110, this embodiment of the invention calculates its instantaneous change with respect to normalized time to obtain the target rate of change field corresponding to the path. In actual processing, this rate of change field has a clear physical meaning; it is equal to the grid-by-grid difference between the meteorological state field at future times and the meteorological state field at historical times. For example, if the temperature at a certain grid point rises from 15°C to 18°C ​​in 6 hours, then the target rate of change field value for that grid point is +3°C (representing the total change in 6 hours). When a normalized time frame is used, this value is the average rate of change per unit normalized time.

[0037] Subsequently, this embodiment of the invention constructs a velocity field parameter model (e.g., implemented using a deep neural network). This model takes the meteorological state field at any given time and the corresponding normalized time position as input, and outputs a predicted value of the meteorological state change rate field at that time. The training objective of the model is to make its predicted change rate field as close as possible to the aforementioned target change rate field.

[0038] Unlike traditional methods that directly learn the mapping from state to state, the core of this step lies in learning the mapping from state to rate of change. This shift allows the model to learn the evolutionary laws of the physical field (i.e., how the field changes over time) rather than a simple correspondence, giving the trained model stronger generalization ability and physical consistency.

[0039] Therefore, through this step, the embodiments of the present invention obtain a velocity field parameter model that can characterize the instantaneous evolution of meteorological conditions, thereby laying the foundation for subsequent continuous time forecast inference.

[0040] Step S130: Obtain the initial meteorological state field to be forecasted. Using the initial meteorological state field as the starting point of evolution, based on the dynamic evolution law characterized by the trained velocity field model, perform segmented state deduction on a continuous time scale. Adapt the time resolution required by the operation through time scale mapping to generate continuous meteorological forecast results with arbitrary time resolution.

[0041] Specifically, this embodiment of the invention first obtains the analysis field at the current moment as the initial meteorological state field, which serves as the starting state for the evolution of the dynamic system. Then, this embodiment of the invention inputs the initial state into a pre-trained velocity field model, which describes the continuous evolution of the meteorological state along the time axis. This process is expressed as an initial value problem of an ordinary differential equation, that is, given the initial state and the law of state change over time (characterized by the velocity field model), the solution is to find the state at any future moment.

[0042] In the actual solution process, this embodiment of the invention does not integrate from the initial time to the target time all at once, but adopts a segmented state extrapolation strategy. Specifically, the system divides the entire forecast duration into multiple standard time unit intervals corresponding to the model pre-training time scale (e.g., each interval is 6 hours), and advances the normalized time from 0 to 1 within each interval. After the extrapolation of each interval is completed, this embodiment of the invention resets the normalized time to the initial value, and uses the final state of the current interval as the starting state of the next interval, repeating this process until the preset forecast time is reached. This segmented restart strategy effectively suppresses the cumulative amplification of errors during long-term extrapolation.

[0043] Meanwhile, embodiments of the present invention adapt to different operational needs through a time-scale mapping mechanism. Since the model learns the evolutionary patterns within a standard time interval (e.g., 6 hours) during the pre-training phase, when the operational needs require forecast results with a finer time resolution (e.g., 1 hour), the step scaling factor in the extrapolation process is adjusted accordingly based on the ratio between the required time step and the standard time interval. This mechanism allows a single model to support forecast output at any time resolution without retraining.

[0044] Through this step, the embodiments of the present invention generate continuous weather forecast results with arbitrary time resolution under a single model framework, combining the stability of long-term forecasts with the flexibility of fine-grained forecasts.

[0045] As can be seen, the embodiments of the present invention, on the one hand, transform weather forecasting from discrete time step modeling to continuous time dynamic system modeling, fundamentally breaking through the limitation of fixed time step size; on the other hand, by directly learning the instantaneous evolution law between states rather than recursive mapping, the rapid accumulation of errors in the time progression process is suppressed from the modeling mechanism; furthermore, a single model can support forecast output at any time resolution, without the need to retrain the model for different time resolutions, significantly improving the applicability and flexibility of the model.

[0046] In a preferred embodiment, constructing a continuous time dynamic path with the historical meteorological state field as the starting point and the future meteorological state field as the ending point includes: within a preset standard normalized time interval, linearly weighting the historical meteorological state field and the future meteorological state field as they change with normalized time, and superimposing a random disturbance field of a preset magnitude to form a continuous dynamic path connecting the real historical state and the real future state along the time axis; calculating the instantaneous change of the continuous dynamic path with respect to normalized time to obtain the target rate of change field, wherein the target rate of change field is equivalent to the grid-by-grid difference between the future meteorological state field and the historical meteorological state field.

[0047] Specifically, within the normalized time interval, this embodiment of the invention represents the meteorological state field at any intermediate moment as a linear weighted combination of the historical state field and the future state field: when the normalized time approaches the starting point, the path state approaches the historical state; when the normalized time approaches the ending point, the path state approaches the future state; at intermediate moments, the path state is the weighted average of the two.

[0048] Based on this, the embodiments of the present invention superimpose a random perturbation field of extremely small magnitude (preset magnitude 1e-3). The magnitude of this perturbation is negligible relative to the value of the meteorological variable itself, does not change the physical evolution direction of the path, and only plays a regularization role during the training process to prevent the model from overfitting.

[0049] Furthermore, in this embodiment of the invention, the instantaneous change of the above-mentioned continuous dynamic path with respect to normalized time is obtained to obtain the target rate of change field. For the path constructed by linear weighted superposition, its instantaneous change is mathematically equal to the grid-by-grid difference between the meteorological state field at future time and the meteorological state field at historical time, and this difference remains constant throughout the entire normalized time interval.

[0050] The target rate of change field represents the total change of meteorological elements within a given time interval, equivalent to a finite-time difference form. Taking a grid point temperature as an example, if the historical temperature is 15℃ and the future temperature is 18℃, then the target rate of change field at that grid point is +3℃, indicating that the temperature at that grid point has risen by a total of 3℃ within a preset time interval. The model is required to predict a rate of change close to +3℃ for any intermediate time of the sample pair, that is, to learn the evolutionary pattern of temperature rising uniformly at a rate of 0.5℃ per hour.

[0051] For example, embodiments of the present invention are based on training sample pairs ( , Within a predefined standard normalized time interval t ∈ [0, 1], a continuous dynamic path of the following form is constructed: The corresponding sample mapping form is: (x) = in, For historical meteorological conditions, For the meteorological state field at a future moment; For normalized time variables; Noise scale that is zero or a very small constant It is a random noise field.

[0052] As can be seen from the formula, the extremely small order of magnitude mentioned above is used... This indicates that its value can be zero or a very small constant, where the very small constant can be 1e-3. The present invention further addresses the aforementioned continuous dynamic path Find the normalized time. The partial derivatives are used to obtain the target rate of change field. : Ignoring the extremely small order of random noise field In the extreme case, the target change rate field, in finite-time difference form, directly reflects the average evolution trend of the meteorological system within a given time interval.

[0053] As can be seen, the embodiments of the present invention achieve the continuity and stability of the forecast over time by superimposing a very small-scale random perturbation field in the linear weighted path.

[0054] In a preferred embodiment, the training of the preset velocity field parameter model employs a two-stage progressive training strategy including a pre-training stage and a fine-tuning stage. In the pre-training stage, training sample pairs are collected based on a first time interval to perform preliminary training on the velocity field parameter model in order to learn the dynamic evolution structure of the atmospheric system. In the fine-tuning stage, training sample pairs are collected based on a second time interval to perform parameter fine-tuning on the pre-trained velocity field parameter model. The second time interval is shorter than the first time interval to adapt the velocity field parameter model to the requirements of high temporal resolution forecasting.

[0055] Specifically, the pre-training phase uses longer time intervals (such as 6 or 12 hours) to collect training sample pairs. Because the meteorological state changes more significantly over longer time intervals, the evolutionary trends contained in the sample pairs are more pronounced. This is beneficial for the model to learn the overall dynamic evolution structure of the meteorological system from a global perspective, while effectively reducing the interference of non-physical time discontinuities and small-scale noise in the reanalysis data on model training.

[0056] In this stage, the model, based on a large number of sample pairs with long time intervals, prioritizes learning the relatively stable large-scale evolution patterns in atmospheric motion. These large-scale structures constitute the basic framework of meteorological evolution and are less affected by assimilation window switching. Mastering these fundamental physical laws through pre-training provides a stable parameter foundation for subsequent fine-tuning at high temporal resolution.

[0057] For example, embodiments of the present invention extract sample pairs arranged at 6-hour intervals from ERA5 reanalysis data, i.e. ( , ), ( , ), ( , ), ( , These constitute training sample pairs. It is worth noting that the above-mentioned 6-hour sampling nodes happen to avoid the switching jump points of the 12-hour assimilation period of ERA5 reanalysis data (09:00-10:00 UTC and 21:00-22:00 UTC), thereby further reducing the misleading effect of non-physical jumps on model learning.

[0058] After pre-training, this embodiment of the invention uses sample pairs with shorter time intervals (e.g., 1 hour) for fine-tuning. The fine-tuning stage employs a continuous multi-step rolling supervision mechanism: using the meteorological state at a certain moment (e.g., Starting with a timescale of 6 hours, the model continuously extrapolates hourly, with the rolling endpoint gradually adjusted as the training progresses (e.g., initially set to 6 hours, then switched to a longer duration, such as 12 hours or 18 hours, after a preset number of training steps (e.g., 1000 steps)). At each step, the model's predictions are supervised and aligned with the corresponding ERA5 reanalysis data to calculate the loss. The system continues training based on the pre-trained model parameters, focusing on learning the fine-grained changes in meteorological elements at the hourly scale, thus enabling the model to possess both the accuracy of large-scale structures and the precision of hourly evolution.

[0059] As can be seen, the embodiments of the present invention, on the one hand, conduct training based on samples with longer time intervals during the pre-training stage. This avoids the non-physical time discontinuities caused by the switching of the ERA5 reanalysis data assimilation window and allows the model to prioritize learning the overall dynamic evolution structure of the meteorological system. This effectively reduces the interference of non-physical jumps on model learning from the training strategy perspective. On the other hand, the fine-tuning stage conducts supervised training based on samples with shorter time intervals. While retaining the large-scale dynamic evolution structure learned in the pre-training stage, it further adapts to the short-term evolution characteristics at finer time scales, enabling the model to simultaneously meet the requirements of accuracy of large-scale dynamic structure and continuity and consistency of high temporal resolution forecasts. Therefore, the two-stage progressive training strategy of the present invention takes into account both the stability of large-scale physical structure and the fine-grained requirements of high temporal resolution, avoids the overfitting problem that is prone to occur when training directly on hourly data, and reduces the model's sensitivity to non-physical perturbations, significantly improving the stability and reliability in high temporal resolution prediction scenarios.

[0060] More preferably, the fine-tuning of the pre-trained velocity field parameter model in the fine-tuning stage includes: using a learning rate smaller than that in the pre-training stage, freezing the low-level feature encoding layer in the velocity field parameter model, and updating only the high-level time modulation parameters and dynamic evolution mapping parameters of the model.

[0061] Specifically, the pre-training phase has already learned the overall dynamic evolution structure of the meteorological system. Fine-tuning only needs to adapt to the short-term evolution requirements at the hourly level, so a learning rate smaller than that used in the pre-training phase is adopted. Small step sizes can avoid drastic parameter fluctuations, prevent damage to the learned basic physical structure, and suppress the interference of non-physical time discontinuities in hourly data on the model.

[0062] The bottom coding layer contains 3D convolutions for processing upper-air meteorological variables and 2D convolutions for processing ground variables and static features. It is responsible for extracting general spatial features and has been fully learned during the pre-training phase. During fine-tuning, all parameters of this layer are frozen, which can avoid overfitting the model to non-physical temporal discontinuities in the data analysis and significantly reduce the number of trainable parameters, thereby improving training efficiency.

[0063] The high-level parameters include the temporal embedding generation layer of the Adaptive Normalization (adaLN) module (output scale parameter γ(t), bias parameter β(t)) and the high-level dynamic modeling module. These parameters directly control the dynamic evolution behavior at different time scales and are strongly correlated with the time scale. The 6-hour scale evolution rules learned in the pre-training stage cannot be directly adapted to the hourly forecasting requirements. Therefore, these parameters are updated in a targeted manner so that the model can grasp the short-term refined evolution characteristics at finer time scales without destroying the basic dynamic structure obtained in the pre-training stage.

[0064] As can be seen, the embodiments of the present invention, on the one hand, effectively avoid catastrophic forgetting during the fine-tuning stage by using a smaller learning rate and freezing the bottom coding layer, fully preserving the overall dynamic evolution structure of the meteorological system learned in the pre-training stage, ensuring the physical rationality of long-term forecasts, and avoiding the destruction of learned large-scale dynamic laws by hourly training; on the other hand, the targeted updating of high-level time modulation and dynamic evolution mapping parameters that are strongly correlated with the time scale enables the model to quickly adapt to hourly short-term evolution laws without destroying the basic physical structure, significantly improving the continuity and consistency of high-temporal-resolution forecasts, and solving the non-physical oscillation problem that is prone to occur in short-interval data training; furthermore, freezing the general feature layer while targeting and updating the task-related layers can significantly reduce the number of trainable parameters, reduce the computational cost of fine-tuning and the risk of overfitting, and effectively suppress the misleading effect of non-physical time discontinuities in hourly data on the model, further improving the stability and reliability of high-term forecasts.

[0065] In a preferred embodiment, the step of performing segmented state extrapolation on a continuous time scale includes: dividing the entire forecast duration into multiple standard time unit intervals corresponding to the model pre-training time scale, performing normalized time progression within each standard time unit interval; after completing the state extrapolation of each standard time unit interval, resetting the normalized time variable used to characterize the evolution progress to its initial value, and using the final state meteorological field obtained from the extrapolation of the current interval as the starting state field of the next interval, repeating the above process until the preset forecast lead time is reached.

[0066] Specifically, the embodiments of the present invention employ a multi-segment restart rolling simulation mechanism, which decomposes long-term forecasts into multiple short-term integrals based on pre-trained time scales, thereby suppressing error accumulation from a mechanism perspective.

[0067] First, the total forecast duration is divided into several standard time intervals that match the pre-training timescale (e.g., 6 hours). Since the model has fully learned the dynamic evolution patterns within the timeframe during the pre-training phase, this division ensures that the inference tasks within each interval fall within the model's existing knowledge domain, avoiding performance degradation caused by timescales exceeding the training range. For example, forecasting the next 72 hours is divided into 12 standard intervals, each corresponding to a 6-hour evolution period.

[0068] Secondly, within each standard interval, the ODE solver is driven to integrate in the normalized time domain. This corresponds to the initial state and the final state of the interval; during integration, intermediate evolution states corresponding to any intermediate value can be extracted at any time resolution.

[0069] Furthermore, after completing an interval extrapolation, a restart operation is performed: the normalized time variable is reset to 0, and the final state meteorological field output by the current interval is used as the starting state field for the next interval. This design ensures that the model always operates within a familiar range, while simultaneously cutting off the chain propagation path of errors between intervals by refreshing the initial conditions with the actual forecast field, significantly suppressing the accumulation of errors in long-term forecasts. For example, after the first interval outputs a 6-hour forecast field, it is reset to zero, and the 6-12 hour interval extrapolation is started with this field as the initial state, and this cycle continues until 72 hours.

[0070] As can be seen, the embodiments of the present invention, on the one hand, divide the long forecast period into multiple standard time unit intervals, decompose the long-term forecast task into multiple short-time integration tasks, and each interval integrates independently and the error is generated independently, effectively suppressing the cumulative amplification effect of error during long-term integration; on the other hand, by resetting the normalization time after each integration segment, the model always works within the normalization time range covered by its training, avoiding the degradation of prediction performance caused by the normalization time exceeding the training range, and ensuring the consistency of the extrapolation accuracy within each interval; furthermore, by using the final state of the previous segment as the initial state of the next segment for state restart, the system uses the real (i.e., the state field updated by the extrapolation of the previous interval) rather than the accumulated error field as the starting point at the beginning of each segment, so that the baseline state of subsequent extrapolations always closely resembles the real atmospheric state, effectively improving the stability and reliability of long-term forecasts.

[0071] More preferably, during the segmented state deduction process, when advancing the normalized time within a single standard time unit interval, the step scaling factor in the deduction process is adjusted according to the proportional relationship between the target output time resolution specified by the business and the standard time unit interval, and segmented iterative deduction is performed using a single-step advancement format or a multi-level advancement format to adapt to the output requirements of any time resolution.

[0072] Specifically, within a single standard time unit interval, it is necessary to extrapolate from the initial state to the final state of that interval. Unlike directly advancing with a fixed step size, this implementation dynamically adjusts the advancement step size according to the output time resolution required by the service.

[0073] Within a single standard time unit interval (the length of which is the pre-training time scale, such as 6 hours), calculate the step scaling factor based on the target output time resolution required by the business needs: This coefficient maps physical time requirements to a normalized time-domain advance step size. Each completed normalized time advance step corresponds to the dynamic evolution of physical world duration. This method eliminates the need to retrain the model; it can output prediction results for arbitrary time intervals simply by adjusting the settings. Once the step size scaling factor is determined, the ordinary differential equations can be discretized and solved using numerical integration. The choice of numerical solution format directly affects prediction accuracy and computational efficiency. This implementation method can use different ordinary differential equation solving algorithms to solve continuous-time dynamical systems, including but not limited to the Euler method, the Runge-Kutta method, or adaptive step-size integration algorithms. Both formats are compatible with multi-stage restart inference mechanisms, allowing users to flexibly switch according to their business scenarios without modifying the model structure.

[0074] For example, the velocity field function learned by the model is denoted as... ,in Represents the normalized time variable. ∈ [0,1]. Given any initial weather state The model characterizes its continuous dynamic evolution process within a standard time interval: By numerically solving this ordinary differential equation, the predicted meteorological conditions at any intermediate time point can be obtained.

[0075] (2) Piecewise ODE solver and time scale mapping In actual reasoning, the time step required by the business is used. (For example, 1 hour or 30 minutes can be smaller than the step size in the fine-tuning phase) compared to the model pre-training timescale. The proportional relationship between (e.g., 6 hours) is r= / The velocity field output is time-scaled and then piecewise integrated using either the explicit Euler method or a higher-order Runge-Kutta method. Taking the Euler method as an example: Until the points are obtained .

[0076] As can be seen, the embodiments of the present invention can dynamically adjust the step size scaling factor according to the ratio between the output time resolution required by the business and the standard time unit interval, so that a single model can be directly adapted to any business time resolution requirement without retraining, significantly improving the deployment flexibility and versatility of the model. In addition, the embodiments of the present invention can support multiple solution strategies, enabling users to flexibly balance computational efficiency and prediction accuracy according to specific business scenarios, taking into account the different needs of high-timeliness and high-quality scenarios.

[0077] In a preferred embodiment, when constructing the velocity field parameter model, a deep network structure with time-conditional modulation is adopted. The deep network structure includes: an encoding module for encoding the input meteorological state field into a latent feature representation; a time-modulated backbone network for deep feature extraction of the latent feature representation, wherein the backbone network introduces a time-adaptive control mechanism in its internal normalization stage to convert the normalized time variable into a feature control parameter, and performs scaling and offset correction on the normalized features of the intermediate layer of the network through the control parameter to characterize the dynamic evolution behavior at different propulsion positions; and a decoding module for reconstructing the feature representation output by the backbone network into a rate of change field, wherein the rate of change field has the same number of variable channels and spatial dimension as the input meteorological state field.

[0078] Specifically, the velocity field model in this embodiment of the invention adopts a three-level structure of encoding-time modulation-decoding. The core is to introduce a time adaptive control mechanism in the backbone network, so that a single network can uniformly represent the dynamic evolution law under any normalized time, without the need to train independent models for different time steps.

[0079] The encoding module is used for multi-source heterogeneous meteorological feature fusion, and is responsible for converting the input multi-source meteorological state field into a unified-dimensional latent feature representation, including: High-altitude variable encoding: for dimensions of The upper-air meteorological variables are extracted using 3D convolution to extract vertical-horizon joint spatial features and output a high-dimensional latent tensor. Ground / Static Feature Encoding: Dimensional To extract planar spatial features from surface meteorological variables and static geographical characteristics, 2D convolution is used. Feature concatenation: The encoding results of high altitude and ground are concatenated along the channel dimension to form a unified joint feature representation, which is then input into the subsequent time-modulated backbone network.

[0080] The backbone network is used for deep extraction of joint features. Its core is the introduction of an adaptive layer normalization (adaLN) mechanism in the normalization layer. Embedded features are fed into the deep network for feature extraction. To account for the influence of the time condition t, an adaptive layer normalization module is introduced into the key layers of the network (such as Transformer Blocks or convolutional residual blocks): adaLN(h; t) = γ(t) ⊙ LN(h) + β(t) Where: h represents the input feature; LN(·) represents Layer Normalization; γ(t) and β(t) are the scale and bias parameters generated by the temporal embedding through a small feedforward network; ⊙ represents element-wise multiplication.

[0081] Through this mechanism, time information t is injected into the features of each layer of the network, enabling a single network to dynamically adjust its dynamic evolution prediction behavior according to different time span requests.

[0082] The decoding module is used to restore the latent features output by the backbone to a rate of change field with the same dimension as the input meteorological field. The latent feature tensors output by the backbone network are mapped back to the ground variable channel and the upper-air variable channel, respectively, and restored to the original spatial resolution and variable dimension through linear projection and inverse patch embedding operations, thereby outputting the velocity field prediction result with the same dimension as the input meteorological state.

[0083] As can be seen, this invention, on the one hand, encodes the high-altitude three-dimensional field and the ground two-dimensional field separately through 3D / 2D convolution, fusing them into a unified joint feature to support full-element dynamic evolution modeling. On the other hand, it injects time variables into each layer of the network through adaptive layer normalization (adaLN), enabling a single model to accurately characterize the dynamic evolution law at any normalized time position, breaking the limitation of fixed time step. Furthermore, the decoding module outputs a rate of change field with the same dimension as the input, providing a structural basis for grid-by-grid finite difference supervision, ensuring training stability and physical consistency.

[0084] In a preferred embodiment, training the velocity field parameter model using the target rate of change field as a supervisory signal includes: updating the model parameters by minimizing the grid-by-grid mean square difference between the rate of change field predicted by the velocity field parameter model and the target rate of change field; during training, the normalized time variable is randomly sampled within a preset standard interval or biased sampled according to a preset distribution strategy to enhance the model's adaptability to the dynamic characteristics of different evolutionary stages.

[0085] Specifically, during the training process, the system pairs each training sample with... In normalized time Sample one within [0,1] The value is used to calculate the intermediate state at that moment based on the continuous dynamic path. Then and Input the velocity field parameter model to obtain the predicted rate of change field. Simultaneously, the target rate of change field is calculated based on the dynamic path. .

[0086] The model updates its parameters by minimizing the grid-by-grid mean square difference between the predicted rate of change field and the target rate of change field. The loss function is: The expectation is that in the training sample pairs With normalized time Calculate on the joint distribution During training, the normalized time variable The sampling method for ∈[0,1] can be uniform random sampling or non-uniform sampling strategy, including but not limited to bias sampling based on time weight, adaptive sampling or segmented sampling, to enhance the model's ability to learn dynamic features at different time scales and improve training stability and convergence efficiency.

[0087] Through the above training method, the model can directly learn the dynamic evolution relationship between real meteorological states without introducing random noise, thereby obtaining a stable and more physically consistent continuous-time dynamic model.

[0088] In a preferred embodiment, the meteorological state field includes at least one of the following multi-source meteorological state variables: surface layer variables, upper-level layer variables, static geographic features, and temporal condition features; the method further includes: performing spatial grid unification alignment and variable standardization processing on the acquired raw multi-source data, and organizing each variable into a multi-dimensional tensor format containing channel dimensions and spatial dimensions.

[0089] Specifically, the data inputs in this embodiment of the invention include multi-source meteorological state variables, preferably including: Surface layer variables: such as near-surface temperature, air pressure, humidity, wind speed, wind direction, precipitation, etc. Variables in the upper atmosphere: such as temperature, specific humidity, wind field, and geopotential at different pressure levels; Static geographical features: such as topographic elevation, land-sea cover, and soil type; Time-related characteristics: such as sine and cosine coding of latitude and longitude, and local time daily and annual cycle coding.

[0090] Furthermore, to ensure the consistency, stability, and computational efficiency of the input data, the present invention preferably performs the following standardized operations: Spatial grid uniform alignment: at fixed time intervals (The pre-training phase can last for 6 hours) Construct training sample pairs : Indicates time The multivariable meteorological state tensor; Indicates time The corresponding meteorological state tensor.

[0091] This sample pair is used to construct a dynamic flow matching path, enabling data-driven learning from real state to real state.

[0092] Standardization of variables: Standardize all continuous variables according to their respective dimensions. in and The mean and standard deviation are obtained from the training set statistics. They are used to ensure that the numerical scale of each variable is consistent and to improve the stability of model training.

[0093] Multidimensional tensor organization: The processed variables are concatenated along the channel dimension to organize them into a multi-channel tensor with a uniform format, and then fed into the model.

[0094] As can be seen, the embodiments of the present invention, by including multi-source data such as ground, upper air, static, and time data, enable the model to acquire as complete atmospheric system state information as possible, providing a data foundation for accurately depicting the laws of meteorological evolution; by uniformly aligning the spatial grid, the consistency of data from different sources is ensured in space, avoiding training errors caused by spatial misalignment; and by standardizing variables, the differences in dimensions and magnitudes between different variables are eliminated, improving the stability and convergence efficiency of model training.

[0095] To explain in more detail, Figure 2 The present invention illustrates a continuous-time weather forecasting logic architecture based on dynamic flow matching, which fully covers the entire lifecycle from data preprocessing to forecast output.

[0096] Among them, the meteorological data acquisition module serves as the data input end, responsible for collecting raw multi-source geophysical data, including but not limited to near-surface meteorological observation elements, upper-air meteorological elements of different isobaric surfaces, static underlying surface geographic features, and time and location coding features.

[0097] The data preprocessing module receives the raw data and performs standardization processing, executes spatial grid alignment and variable standardization operations, and organizes multi-source heterogeneous data into multi-channel multi-dimensional tensors in a unified format, providing a standardized data foundation for subsequent modeling.

[0098] The sample building module and two-stage training path are used to receive preprocessed data and are responsible for constructing training sample pairs.

[0099] Pre-training phase: Sample pairs are constructed based on the first time interval, and the target rate of change field is used with finite difference. The basic velocity field model is trained by minimizing the mean square error so that it can grasp the overall dynamic evolution structure of the atmospheric system.

[0100] Fine-tuning stage: Based on the second time interval, construct high temporal resolution sample pairs, load pre-trained weights to fine-tune parameters, and generate an optimized continuous-time velocity field model.

[0101] The initial meteorological state and ODE solver inference module receives the preprocessed initial meteorological state input. The ODE solver inference module calls the trained continuous-time velocity field model, performs numerical integration based on ordinary differential equations, and outputs future weather forecasts at arbitrary time resolution through a multi-segment restart rolling inference mechanism.

[0102] Figure 3 The diagram illustrates the two-stage training process of this invention. This process, by breaking down long-duration tasks into consecutive short-duration sub-tasks and combining them with step-by-step supervision based on physical constraints, significantly improves the model's ability to capture high temporal resolution evolutionary details without increasing additional computational costs. Specifically, it includes the following core steps: Phase 1: Pre-training This stage establishes the basic dynamic understanding of the model. The process begins with the initial state. Sampling Get the state at a specific moment Then it is input into the velocity field model. Forward inference is performed, and finally the network weights are updated by calculating the loss function.

[0103] Phase 2: Fine-tuning This stage, based on the pre-trained weights, focuses on improving the model's adaptability to fine-grained temporal changes. The system uses... Starting from this point, we first perform a standard single-step deduction to obtain... .

[0104] Subsequently, the core multi-step rolling monitoring mechanism is activated: the system does not directly perform long-term prediction, but instead uses the output of the previous step... As new input, it is fed into the velocity field model. At this point, the input time stamp of the model is... This allows for the calculation of the predicted rate for the next step. By introducing a supervised intermediate state module, the model is forced to predict the next state. It must be aligned with actual observation data. This process is continuously repeated. Figure 3 (represented by ellipses "..."), forming a continuous forward integral trajectory of ODE.

[0105] Figure 4 This more intuitively demonstrates how the present invention achieves long-term, high-resolution weather forecasts during the inference phase. Its core lies in employing a multi-stage restart strategy to solve the ordinary differential equation (ODE) piecewise, as detailed below: Based on initial meteorological conditions This serves as the starting point. In the first stage of the derivation, this state is input into the velocity field model. At this point, the model's input time stamp is 0 (i.e., This marks the beginning of a new standard time interval. Subsequently, the system performs forward integration of the ODE, generating it sequentially along the time axis until the end of the interval.

[0106] When the first segment of integration is completed (i.e., when the point is reached) At any given time, the system performs a rolling inference operation: it sets the endpoint state of the previous interval... This is directly used as the starting state for the next interval. The normalized time variable is reset to 0. In the next deduction, although physical time is... However, the time stamp input to the model is reset to 0 (i.e., ).

[0107] Starting from a new point and with a reset time stamp, perform ODE forward integration again, repeating the above process until the preset total forecast duration is covered (e.g., ...). Or further.

[0108] In summary, this invention establishes a modeling paradigm for continuous-time dynamic systems through a data-to-data dynamic flow matching training mechanism, achieving stable and continuous forecasting of meteorological elements at arbitrary time resolutions. It also demonstrates significant technical advantages in terms of forecast stability, error accumulation suppression, temporal consistency, model generalization ability, and flexibility in engineering deployment. Specific technical effects are as follows: (1) Physical consistency is significantly improved by using a data-to-data dynamic flow matching training mechanism.

[0109] By constructing a data-to-data probability path from real weather conditions to real weather conditions, and directly using finite time difference as the velocity field monitoring signal, the traditional generative modeling method of using random noise as the initial state is avoided. This ensures that the model always uses the real weather field as the starting point for evolution, thereby enhancing the physical rationality and temporal continuity of the prediction results from the modeling mechanism.

[0110] (2) By using continuous dynamic evolution modeling, the accumulation of multi-step prediction errors can be suppressed from a mechanistic perspective.

[0111] By learning the continuous-time velocity field and generating future weather conditions using an ODE solver, this invention avoids the chain propagation mechanism of error amplification in traditional discrete recursive prediction, making the model error exhibit a smoother and more controllable growth characteristic over time, and has higher stability and reliability in medium- and long-term high-time-efficiency prediction scenarios.

[0112] (3) By using a two-stage training mechanism, the adverse effects of the time discontinuity of the reanalysis data on model learning can be reduced.

[0113] By pre-training on data with a longer time interval and then fine-tuning on hourly data, this invention effectively reduces the interference of time discontinuity caused by assimilation window switching in reanalysis data on the model learning process, avoids the model from mislearning non-physical jumps as real dynamic features, and thus further improves the stability and consistency of hourly and high temporal resolution prediction results.

[0114] (4) Achieving multi-timescale prediction through a unified model framework reduces system deployment and maintenance costs.

[0115] By modeling continuous-time dynamic systems and introducing a time modulation mechanism into the velocity model, this invention can simultaneously support hourly prediction and higher time resolution prediction tasks within a single model framework. This avoids training independent models for different time scales, thereby significantly reducing system training, deployment, and maintenance costs and improving engineering application efficiency.

[0116] Accordingly, based on the same inventive concept as the above method, embodiments of the present invention provide a continuous-time weather forecasting device, such as... Figure 5 As shown, the weather forecasting device 500 includes: a construction unit 510, used to acquire training sample pairs consisting of historical weather state fields and future weather state fields, and construct a continuous-time dynamic path with the historical weather state field as the evolution starting point and the future weather state field as the evolution ending point, wherein the continuous-time dynamic path is used to characterize the evolution process of the physical field from the historical weather state field to the future weather state field; a training unit 520, used to determine the target rate of change field of the weather state within a unit normalized time based on the dynamic path, and use the target rate of change field as a supervision signal to train a preset velocity field parameter model for fitting the atmospheric evolution law; and a deduction unit 530, used to acquire the initial weather state field to be forecasted, use the initial weather state field as the evolution starting point, and perform segmented state deduction on a continuous time scale based on the dynamic evolution law characterized by the trained velocity field model, and generate continuous weather forecast results with arbitrary time resolution by adapting the time scale mapping to the time resolution required for the operation.

[0117] More details on the implementation and advantages of this device can be found in the above-described optimized method embodiments, and will not be elaborated further here.

[0118] Accordingly, the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the continuous time weather forecasting method described in the above embodiments.

[0119] Accordingly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the continuous time weather forecasting method described in the above embodiments.

[0120] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0121] The units described in the embodiments of the present invention can be implemented in software or in hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0122] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0123] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0124] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0125] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in sequential order. In certain circumstances, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0126] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A continuous-time weather forecasting method, characterized in that, The weather forecasting methods include: A training sample pair consisting of a historical meteorological state field and a future meteorological state field is obtained, and a continuous time dynamic path is constructed with the historical meteorological state field as the evolution starting point and the future meteorological state field as the evolution ending point. The continuous time dynamic path is used to characterize the evolution process of the physical field from the historical meteorological state field to the future meteorological state field. Based on the dynamic path, the target rate of change field of meteorological state within a unit normalized time is determined, and the target rate of change field is used as a supervision signal to train a preset velocity field parameter model for fitting the atmospheric evolution law. The initial meteorological state field to be forecasted is obtained. Using the initial meteorological state field as the starting point of evolution, based on the dynamic evolution law characterized by the trained velocity field model, segmented state deduction is performed on a continuous time scale. By adapting the time scale mapping to the time resolution required by the operation, continuous meteorological forecast results with arbitrary time resolution are generated.

2. The weather forecasting method according to claim 1, characterized in that, The constructed continuous-time dynamic path, starting from the historical meteorological state field and ending with the future meteorological state field, includes: Within a preset standard normalized time interval, the meteorological state fields at the historical time and the meteorological state fields at the future time are linearly weighted and superimposed as normalized time, and a random disturbance field of a preset magnitude is superimposed to form a continuous dynamic path connecting the real historical state and the real future state along the time axis. The instantaneous change of the continuous dynamic path with respect to normalized time is obtained to obtain the target rate of change field, which is equivalent to the grid-by-grid difference between the meteorological state field at the future time and the meteorological state field at the historical time.

3. The weather forecasting method according to claim 1, characterized in that, The preset velocity field parameter model for training employs a two-stage progressive training strategy, including a pre-training stage and a fine-tuning stage. During the pre-training phase, the training sample pairs are collected based on the first time interval to perform preliminary training on the velocity field parameter model in order to learn the dynamic evolution structure of the atmospheric system. During the fine-tuning phase, the training sample pairs are collected based on a second time interval, and the parameters of the pre-trained velocity field parameter model are fine-tuned. The second time interval is shorter than the first time interval, so that the velocity field parameter model can be adapted to the high temporal resolution forecasting requirements.

4. The weather forecasting method according to claim 3, characterized in that, The fine-tuning of the pre-trained velocity field parameter model during the fine-tuning phase includes: A learning rate smaller than that used in the pre-training phase is adopted, and the low-level feature encoding layer in the velocity field parameter model is frozen. Only the high-level time modulation parameters and dynamic evolution mapping parameters of the model are updated.

5. The weather forecasting method according to claim 1, characterized in that, The segmented state deduction over a continuous time scale includes: The entire forecast duration is divided into multiple standard time unit intervals corresponding to the model pre-training time scale, and normalized time progression is performed within each standard time unit interval. After each standard time unit interval is completed, the normalized time variable used to characterize the evolution progress is reset to the initial value, and the final state meteorological field obtained from the current interval is used as the starting state field for the next interval. The above process is repeated until the preset forecast lead time is reached.

6. The weather forecasting method according to claim 5, characterized in that, During the segmented state deduction process, when advancing the normalized time within a single standard time unit interval, the step scaling factor in the deduction process is adjusted according to the ratio between the target output time resolution specified by the business and the standard time unit interval. Segmented iterative deduction is then performed using a single-step advancement format or a multi-level advancement format to adapt to the output requirements of any time resolution.

7. The weather forecasting method according to claim 1, characterized in that, When constructing the velocity field parameter model, a deep network structure with time-conditional modulation function is adopted. The deep mesh structure includes: The encoding module is used to encode the input meteorological state field into a latent feature representation; A time-modulated backbone network is used to extract deep features from the latent feature representation. The backbone network introduces a time-adaptive control mechanism in the internal normalization stage to transform the normalized time variable into a feature control parameter. The control parameter is used to scale and offset the normalized features of the intermediate layer of the network to characterize the dynamic evolution behavior at different propulsion positions. The decoding module is used to reconstruct the feature representation output by the backbone network into a rate of change field, which has the same number of variable channels and spatial dimension as the input meteorological state field.

8. The weather forecasting method according to claim 1, characterized in that, The method of training the velocity field parameter model using the target rate of change field as a supervision signal includes: The model parameters are updated by minimizing the grid-by-grid mean square difference between the rate of change field predicted by the velocity field parameter model and the target rate of change field. During training, the normalized time variable is randomly sampled within a preset standard interval or biased sampled according to a preset distribution strategy to enhance the model's adaptability to the dynamic characteristics of different evolutionary stages.

9. The weather forecasting method according to claim 1, characterized in that, The meteorological state field includes at least one of the following multi-source geophysical data: Near-surface meteorological observation elements, upper-air meteorological elements of different isobaric layers, static underlying surface geographical features, and time-location coding features characterizing daily and annual cycle changes; The method further includes: performing spatial grid unification alignment and variable standardization on the acquired raw multi-source data, and organizing each variable into a multi-dimensional tensor format containing channel dimension and spatial dimension.

10. A continuous-time weather forecasting device, characterized in that, The weather forecasting device includes: A construction unit is used to acquire training sample pairs consisting of historical meteorological state fields and future meteorological state fields, and to construct a continuous time dynamic path with the historical meteorological state field as the evolution starting point and the future meteorological state field as the evolution ending point, wherein the continuous time dynamic path is used to characterize the evolution process of the physical field from the historical meteorological state field to the future meteorological state field. The training unit is used to determine the target rate of change field of meteorological state within a unit normalized time based on the dynamic path, and to train a preset velocity field parameter model for fitting the atmospheric evolution law using the target rate of change field as a supervision signal. The extrapolation unit is used to obtain the initial meteorological state field to be forecasted. Taking the initial meteorological state field as the starting point of evolution, based on the dynamic evolution law characterized by the trained velocity field model, it performs segmented state extrapolation on a continuous time scale. By mapping the time scale to adapt to the time resolution required by the operation, it generates continuous meteorological forecast results with arbitrary time resolution.

11. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the continuous time weather forecasting method as described in any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the continuous time weather forecasting method as described in any one of claims 1 to 9.

Images