A short-term precipitation prediction-oriented cycle parameter self-adaption and side boundary field driving method, system and storage medium

Abstract

Embodiments of the present application provide a short-term precipitation forecast-oriented cycle parameter self-adaptive and side boundary field driving method, system and storage medium. The method comprises: obtaining side boundary field data, historical model field data and real-time observation data, and inputting them into a model; calculating a gradient influence component based on the gradient trajectory of the model parameters, and calculating a feature influence component based on the activation correlation of the current input features; fusing to obtain a basic score, and performing boundary enhancement to obtain a dynamic influence score; matching with a pre-constructed joint index memory library; if successful, loading a historical parameter group to perform local optimization; if failed, screening sensitive parameters based on the dynamic influence score and performing elastic update; generating an original precipitation forecast field after updating, and performing energy conservation correction and water substance redistribution to obtain a correction result. In this way, the problems of insufficient response to the side boundary field and insufficient constraint on the physical generation and consumption process in the short-term precipitation forecast can be solved.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of regional meteorological forecasting and artificial intelligence, specifically to a cyclic parameter adaptive and lateral boundary field driven technology for short-term precipitation forecasting. Background Technology

[0002] Short-term (0-6 hour) precipitation forecasting is a core component of severe convective weather early warning, urban flood control, and support for major events. Currently, this field mainly employs two technical approaches: numerical weather prediction and artificial intelligence (AI) forecasting. Traditional numerical weather prediction methods, based on time integration of fluid dynamics equations, while physically interpretable, consume enormous computational resources, making it difficult to meet the operational requirements of high-frequency rolling updates for short-term forecasts. Furthermore, the lack of a real-time coupling mechanism with the model's internal state leads to a lag in response to boundary disturbances. Deep learning-based forecasting methods, on the other hand, are widely used due to their fast inference speed and strong fitting ability. However, existing deep learning models often employ offline training and online fixation, with model parameters remaining unchanged after deployment, failing to adapt to the rapid evolution of weather systems. Simultaneously, existing methods often focus on minimizing prediction errors, easily generating non-physical false precipitation in severe convective scenarios. Therefore, there is an urgent need for a method capable of efficient and accurate weather forecasting in regional hazardous weather scenarios such as short-term forecasts of severe convective weather and typhoon precipitation forecasts. Summary of the Invention

[0003] This application provides a method, system, and storage medium for adaptive cyclic parameters and lateral boundary field driving for short-term precipitation forecasting.

[0004] According to a first aspect of this application, a cyclic parameter adaptive and lateral boundary field driven method for short-term precipitation forecasting is provided.

[0005] Acquire lateral boundary field data, historical model field data, and real-time observation data of the target area, and input them into the neural network prediction model to be updated;

[0006] Based on the gradient trajectory of each parameter in the neural network prediction model, the gradient influence component is calculated; and based on the activation correlation between each parameter and the current input feature, the feature influence component is calculated; the gradient influence component and the feature influence component are fused to obtain the basic score, and the basic score is enhanced by the spatial gradient of the side boundary field to obtain the dynamic influence score.

[0007] The historical pattern field data is encoded to obtain the pattern features of the current time period; the side boundary field data is encoded to obtain the side boundary field features.

[0008] The pattern features and the side boundary field features are matched with a pre-built joint index memory. If the match is successful, the corresponding historical parameter group is loaded and local optimization is performed. If the match fails, sensitive parameters are selected based on the dynamic influence score and elastic updates are performed, and attenuation constraints are applied to the boundary sensitive parameters during the update process.

[0009] The original precipitation forecast field was generated using the updated neural network forecast model, and energy conservation correction and water redistribution were performed sequentially to obtain a short-term precipitation forecast field that satisfies the mass and energy conservation constraints.

[0010] The short-term precipitation forecast field and the real-time observation data of the next time period are used together as new inputs, and the aforementioned process is repeated to achieve rolling forecast updates.

[0011] Furthermore, the basic score is enhanced by boundary enhancement based on the spatial gradient of the lateral boundary field to obtain a dynamic influence score, including:

[0012] For the i-th parameter W in the neural network prediction model i Maintain a gradient queue ΔW of length T. i (1:T) Calculate the gradient mean μ i σ i 2 and autocorrelation coefficient ρ i (τ), the influence component of the gradient output from the two fully connected layers S grad (i) ;

[0013] For the current input feature map F∈R C×H×W Perform global average pooling to obtain the channel attention weights A. C Calculate the correlation between the parameter and the k-th channel I(W) i The feature influence components are obtained by aggregation (k). ;

[0014] Calculate the fusion coefficient using the gating unit. Receive a basic score ;

[0015] According to the spatial gradient of the side boundary field The boundary influence factor is output by the multilayer perceptron. The base score is modulated to obtain a dynamic influence score.

[0016] Furthermore, after obtaining the dynamic influence score, the method also includes a score stability control step:

[0017] Apply an exponential moving average to the dynamic influence score:

[0018] ;

[0019] Calculate the score mutation rate:

[0020] ;

[0021] When τ(t) is greater than the preset threshold, reset the smoothing score and re-execute the parameter influence assessment.

[0022] Furthermore, the composite index memory is constructed using the following algorithm:

[0023] Calculate the Pearson correlation coefficient between parameter updates:

[0024] ;

[0025] Group parameters with Corr>0.7 into the same parameter group;

[0026] The historical pattern field data was processed using an autoencoder to extract 128-dimensional pattern features. The convolutional neural network was used to extract 64-dimensional boundary features ψ from the boundary field data on the opposite side;

[0027] Use the locality-sensitive hash function family:

[0028] ;

[0029] Generate a composite index key of length L And bind the parameter group to the index key for storage.

[0030] Furthermore, the elastic update employs a joint algorithm of momentum acceleration and boundary decay:

[0031] Selected in descending order of dynamic influence score One sensitive parameter;

[0032] Perform the update:

[0033] ;

[0034] ;

[0035] ;

[0036] Where β = 0.9, ;

[0037] Apply attenuation to boundary-sensitive parameters:

[0038] .

[0039] Furthermore, the energy conservation correction includes:

[0040] Calculate the total energy:

[0041] ;

[0042] Calculate the energy imbalance:

[0043] ;

[0044] Correcting potential temperature using projection operator:

[0045] .

[0046] Furthermore, the water material redistribution includes employing the Wasserstein optimal transport algorithm:

[0047] Detection of water substance imbalance:

[0048] ;

[0049] Solving the optimal transmission problem:

[0050] ;

[0051] Redistribution of the water vapor field:

[0052] .

[0053] Furthermore, following the water substance redistribution, the precipitation phase change process is controlled using a threshold-triggered differential equation:

[0054] Cloud-water to rainwater conversion:

[0055] ;

[0056] Rainwater evaporation:

[0057] ;

[0058] in , and This is an empirical coefficient. This is an indicator function.

[0059] According to a second aspect of this application, a cyclic parameter adaptive and lateral boundary field driven system for short-term precipitation forecasting is provided, the system comprising:

[0060] The data input module is used to acquire the lateral boundary field data, historical model field data and real-time observation data of the target area, and input them into the neural network prediction model to be updated.

[0061] The dynamic influence score determination module is used to calculate the gradient influence component based on the gradient trajectory of each parameter in the neural network prediction model; and to calculate the feature influence component based on the activation correlation between each parameter and the current input feature; to fuse the gradient influence component and the feature influence component to obtain the basic score, and to perform boundary enhancement on the basic score according to the spatial gradient of the side boundary field to obtain the dynamic influence score.

[0062] The side boundary field feature determination module is used to encode the historical pattern field data to obtain the pattern features of the current time; and to encode the side boundary field data to obtain the side boundary field features.

[0063] The model parameter optimization module is used to match the mode features and the side boundary field features with a pre-built joint index memory. If the match is successful, the corresponding historical parameter group is loaded and local optimization is performed. If the match fails, sensitive parameters are selected based on the dynamic influence score and elastic updates are performed. During the update process, attenuation constraints are applied to the boundary sensitive parameters.

[0064] The model output and result correction module is used to generate the original precipitation forecast field using the updated neural network forecast model, and then perform energy conservation correction and water material redistribution in sequence to obtain a short-term precipitation forecast field that satisfies the mass and energy conservation constraints.

[0065] The forecast update module is used to take the short-term precipitation forecast field and the real-time observation data of the next time period as new inputs, and return to execute the aforementioned process to realize rolling forecast updates.

[0066] According to a third aspect of this application, a readable storage medium is provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps of the cyclic parameter adaptation and lateral boundary field driven method for short-term precipitation forecasting as described above.

[0067] According to a fourth aspect of this application, an electronic device is provided, the electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the cyclic parameter adaptation and lateral boundary field driven method for short-term precipitation forecasting as described above.

[0068] The technical solution provided in this embodiment, by integrating online adaptive adjustment of parameters, lateral boundary field constraints, and physical conservation mechanisms, is used to solve the problems of insufficient response of deep learning models to lateral boundary fields and insufficient constraints on physical generation and dissipation processes in short-term precipitation forecasts. It is applicable to regional disaster weather scenarios such as short-term forecasts of severe convective weather and typhoon precipitation forecasts.

[0069] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0070] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the application. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0071] Figure 1 This is a flowchart illustrating the cyclic parameter adaptive and lateral boundary field driven method for short-term precipitation forecasting provided in this embodiment.

[0072] Figure 2 This is a schematic diagram comparing the principles of the present invention with those of the prior art, provided in this embodiment;

[0073] Figure 3 This is a schematic diagram of the overall system execution flow provided in the embodiments of this application;

[0074] Figure 4 This is a schematic diagram illustrating the specific process of boundary-aware parameter influence assessment provided in the embodiments of this application;

[0075] Figure 5 This is a schematic diagram illustrating the specific process of enhanced fine-tuning and memory optimization provided in the embodiments of this application;

[0076] Figure 6 This is a schematic diagram illustrating the specific process of real-time correction of conservation constraints provided in the embodiments of this application;

[0077] Figure 7 This is a schematic diagram of the cyclic parameter adaptive and lateral boundary field driven system for short-term precipitation forecasting provided in this embodiment. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0079] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0080] Figure 1 This is a flowchart illustrating the cyclic parameter adaptive and lateral boundary field driven method for short-term precipitation forecasting provided in this embodiment. Figure 1 As shown, the method includes:

[0081] S11: Acquire the lateral boundary field data, historical model field data and real-time observation data of the target area, and input them into the neural network prediction model to be updated;

[0082] The target area can be a specific spatial range, such as a city, district, or village. It can be defined as the geographic area used for short-term precipitation forecasting, such as urban flood control areas, typhoon outer impact areas, or areas prone to severe convective weather.

[0083] Lateral boundary field data refers to boundary meteorological field data, such as wind field, temperature field, humidity field, or geopotential height field, provided by global numerical models or large-scale regional models and used to drive regional forecast calculations.

[0084] Historical model field data refers to time-series data composed of historical analysis fields and historical forecast fields output by numerical models over a period of time, used to characterize the early evolution of meteorological systems.

[0085] Real-time observation data refers to observation data acquired in real time by weather radar, meteorological satellites, and automatic ground observation stations and subjected to quality control, such as radar echo intensity, satellite brightness temperature, and ground precipitation intensity observations.

[0086] Neural network forecasting models refer to artificial intelligence models built on deep learning architectures and specifically designed for spatiotemporal prediction of short-term precipitation fields.

[0087] This solution enables unified access to multi-source meteorological data through standard data interfaces for reading, aggregation, format conversion, and quality checks. After spatiotemporal alignment, normalization, and outlier removal, the multi-source data is fed into the input layer of a neural network forecasting model, providing a data foundation for subsequent parameter evaluation and forecast calculations.

[0088] S12, calculate the gradient influence component based on the gradient trajectory of each parameter in the neural network prediction model; and calculate the feature influence component based on the activation correlation between each parameter and the current input feature; fuse the gradient influence component and the feature influence component to obtain the basic score, and perform boundary enhancement on the basic score according to the spatial gradient of the side boundary field to obtain the dynamic influence score.

[0089] Gradient trajectory refers to the time-series record of the update magnitude, update direction and change trend of each parameter in a neural network prediction model during multiple rounds of iterative updates, used to reflect the degree of contribution of the parameters to the model optimization.

[0090] The gradient influence component is a numerical scoring index used to quantify the degree to which changes in parameter gradients affect the model output. It can reflect the sensitivity of parameters in model optimization.

[0091] Activation correlation refers to the response strength, contribution weight, and correlation degree between model parameters and the feature map channels of intermediate layers in the network.

[0092] Feature influence component refers to a numerical scoring index used to quantify the importance of parameters in feature extraction, feature mapping, and spatiotemporal feature expression.

[0093] The base score refers to the comprehensive score of the importance of the initial parameters obtained by fusing gradient dimension information and feature dimension information. Currently, the side boundary field constraint information has not been included.

[0094] The spatial gradient of the lateral boundary field refers to the rate of change of the physical quantities of the lateral boundary field in the horizontal spatial direction and the set of gradient vectors, which is used to characterize the intensity and propagation trend of boundary disturbances.

[0095] Boundary enhancement refers to the process of incorporating information about changes in the lateral boundary field into the parameter scoring system and weighting the base score to strengthen the weight of boundary-sensitive parameters.

[0096] Dynamic influence score refers to the final score of parameter importance that can adaptively adjust with changes in the lateral boundary field and real-time observation updates.

[0097] This scheme obtains the gradient influence component through gradient statistical analysis and the feature influence component through channel attention mechanism. The two influence components are then weighted and merged according to adaptive weights to obtain the parameter base score. The base score is then weighted and modulated using the spatial gradient of the side boundary field to obtain a dynamic influence score that can respond to boundary changes in real time.

[0098] S13, Encode the historical pattern field data to obtain the pattern features of the current time period; Encode the side boundary field data to obtain the side boundary field features;

[0099] Encoding refers to the feature extraction and dimensionality reduction operations that transform high-dimensional, high-resolution spatiotemporal meteorological field data into low-dimensional, fixed-length representation vectors through nonlinear mapping.

[0100] The current time-series pattern features refer to the low-dimensional embedding vectors that represent the recent state, structure, and evolution trend of the meteorological system, obtained by encoding the historical pattern field using an autoencoder (AE).

[0101] Lateral boundary field features refer to low-dimensional embedding vectors that characterize the boundary structure, perturbation intensity, and spatial gradient, obtained by extracting spatial features of the lateral boundary field using convolutional neural networks (CNNs).

[0102] Specifically, an autoencoder can be used to perform nonlinear dimensionality reduction encoding on historical model field data to obtain model features that characterize the state of the meteorological system, and a convolutional neural network can be used to extract spatial features from the lateral boundary field data to obtain lateral boundary field features that characterize boundary driving information.

[0103] S14: Match the pattern features and side boundary field features with the pre-built joint index memory; if the match is successful, load the corresponding historical parameter group and perform local optimization; if the match fails, filter sensitive parameters based on dynamic influence scores and perform elastic updates, and apply decay constraints to the boundary sensitive parameters during the update process.

[0104] Model features refer to low-dimensional embedding vectors that can characterize the current structure, intensity, and evolution trend of meteorological systems, obtained by extracting features from historical model field data through an autoencoder.

[0105] Side boundary field features refer to low-dimensional embedding vectors that can characterize the intensity of boundary perturbation, spatial gradient, and propagation trend, obtained by extracting spatial features from side boundary field data through a convolutional neural network.

[0106] A joint index memory is a pre-built structured database that associates and binds model features, lateral boundary field features, and corresponding optimal parameter sets, and supports fast retrieval. It is used for the reuse of parameter experience in similar weather scenarios.

[0107] Historical parameter set refers to a set of model parameters that have been fully trained, converged stably, and have excellent forecasting performance under similar historical weather conditions.

[0108] Local optimization refers to performing small-scale, high-precision constraint optimization on the loaded parameter set after matching similar historical scenarios, so that the parameters are adapted to the current weather conditions of the target area.

[0109] Sensitive parameters refer to key parameters that rank highly in the dynamic influence score and have the most significant impact on model output and precipitation forecast results.

[0110] Elastic updates refer to an optimization strategy that involves making lightweight, efficient, and minor online adjustments only to selected sensitive parameters, without updating all parameters.

[0111] Boundary-sensitive parameters are those that respond strongly to changes in the lateral boundary field and are easily affected by boundary disturbances, resulting in large update amplitudes. They require stability constraints.

[0112] Decay constraints are gradient magnitude limits and stability constraints imposed to prevent abnormal updates and model oscillations caused by excessive disturbances in boundary-sensitive parameters.

[0113] This scheme jointly encodes the pattern features extracted at the current moment with the side boundary field features, calculates the cosine similarity between the features and historical features in the joint index memory, completes the rapid retrieval and comparison of similar weather scenarios, and when the similarity reaches a preset threshold, reads and loads the historical optimal parameter set corresponding to the similar scenario from the memory, replaces or initializes the corresponding parameters of the current model, and uses a constraint optimization algorithm to perform local fine-tuning on the loaded historical parameter set, so that the parameters can quickly adapt to the current weather conditions and boundary conditions.

[0114] If a match fails, the parameters are sorted from highest to lowest according to their dynamic influence scores. The top K parameters with the greatest influence are selected as the sensitive parameters that need to be updated in this round. The momentum optimization algorithm is used to perform lightweight online updates on the selected sensitive parameters to maintain the model's fast response and prediction accuracy.

[0115] Furthermore, this scheme will weight and attenuate the update magnitude of boundary-sensitive parameters according to the gradient of the side boundary field during the parameter update process, so as to avoid drastic parameter fluctuations and model instability.

[0116] S15, the original precipitation forecast field is generated using the updated neural network forecast model, and energy conservation correction and water material redistribution are performed in sequence to obtain a short-term precipitation forecast field that satisfies the mass and energy conservation constraints.

[0117] The original precipitation forecast field refers to the spatial distribution field of precipitation intensity directly output by the model through forward inference driven solely by data, without any physical conservation constraints.

[0118] Energy conservation correction refers to the projection correction of physical quantities such as potential temperature, kinetic energy, and internal energy in the forecast field based on the atmospheric thermodynamic energy balance equation, so as to satisfy the physical constraint operation of the law of conservation of energy.

[0119] Water redistribution refers to the spatial redistribution of the water vapor field in the forecast field while keeping the total amount of water vapor, cloud water, and rainwater in the atmosphere constant, so as to satisfy the physical constraint of the law of conservation of mass.

[0120] The mass and energy conservation constraint refers to the dual physical constraints that require the total energy of the system to remain in balance and the total water content to remain unchanged in order to ensure that the forecast results conform to the actual atmospheric physical laws.

[0121] Short-term precipitation forecast field refers to the final operational forecast result that is physically reasonable, free of false precipitation, and spatiotemporally continuous after undergoing energy conservation correction and water redistribution dual physical constraints.

[0122] This scheme performs forward inference calculations on the input data based on the updated neural network forecast model, outputting the original spatial distribution of precipitation without physical constraints. Then, the original forecast field is projected and corrected according to the atmospheric energy balance equation to eliminate energy imbalance terms and ensure the rationality of energy physics. Furthermore, the optimal transport algorithm is used to spatially redistribute water vapor and water substances to ensure that the total water volume remains unchanged and achieve mass conservation. Finally, the short-term precipitation forecast results are obtained after double physical constraint correction.

[0123] S16 takes the short-term precipitation forecast field and the real-time observation data of the next time period as new inputs, returns to execute the aforementioned process, and realizes rolling forecast updates.

[0124] The next real-time observation data refers to the latest real-time observation data collected after the current time forecast is completed, such as radar, satellite, and ground automatic station data, which is used for cyclic assimilation and model correction.

[0125] The technical solution provided in this embodiment achieves real-time adaptive adjustment of model parameters with the boundary field and observations through parameter dynamic influence assessment, memory matching and retrieval, parameter elastic updating and physical conservation correction. This significantly improves the accuracy, physical rationality and boundary response capability of short-term precipitation forecasts, effectively suppresses false precipitation, and is suitable for short-term forecasting scenarios of severe convection, typhoons and other disastrous weather.

[0126] In one embodiment, optionally, the base score is enhanced based on the spatial gradient of the lateral boundary field to obtain a dynamic influence score, including:

[0127] For the i-th parameter W in the neural network prediction model i Maintain a gradient queue ΔW of length T. i (1:T) Calculate the gradient mean μ i σ i 2 and autocorrelation coefficient ρ i(τ), the influence component of the gradient output from the two fully connected layers S grad (i) ;

[0128] For the current input feature map F∈R C×H×W Perform global average pooling to obtain the channel attention weights A. C Calculate the correlation between the parameter and the k-th channel I(W) i The feature influence components are obtained by aggregation (k). ;

[0129] Calculate the fusion coefficient using the gating unit. Receive a basic score ;

[0130] According to the spatial gradient of the side boundary field The boundary influence factor is output by the multilayer perceptron. The base score is modulated to obtain a dynamic influence score.

[0131] Among them, channel attention weight refers to the weight coefficient that characterizes the importance of each channel and is used to measure the correlation strength between parameters and feature channels.

[0132] A gating unit is an adaptive weight calculation unit based on the Sigmoid function, used to dynamically adjust the fusion ratio of gradient and feature components.

[0133] A multi-layer perceptron (MLP) is a network structure consisting of multiple fully connected layers used for nonlinear mapping of boundary gradients to influence factors.

[0134] The boundary influence factor refers to the modulation coefficient calculated from the boundary field gradient, which is used to dynamically enhance the parameter scoring based on boundary information.

[0135] This scheme can continuously record parameter gradient updates and maintain a fixed-length sequence to ensure the gradient trajectory is traceable. Statistical operations are performed on the gradient sequence to obtain the gradient influence component. The feature influence component is obtained through an attention mechanism. The boundary influence factor is used to weight and adjust the basic score to obtain a boundary-aware dynamic influence score.

[0136] This technical solution achieves precise quantification of parameter influence through dual-modal perception of gradient trajectory and feature activation, combined with gated fusion and boundary field gradient modulation. This enables the model to respond sensitively to changes in the side boundary field, improves the pertinence and effectiveness of parameter updates, reduces invalid computation and parameter perturbation, and enhances the model's responsiveness to boundary-driven events.

[0137] In one embodiment, optionally, after obtaining the dynamic influence score, the method further includes a score stability control step:

[0138] Apply an exponential moving average to the dynamic influence score:

[0139] ;

[0140] Calculate the score mutation rate:

[0141] ;

[0142] When τ(t) is greater than the preset threshold, reset the smoothing score and re-execute the parameter influence assessment.

[0143] Among them, the exponential moving average refers to a filtering method that uses weighted smoothing to enhance historical information and reduce instantaneous noise in a scoring sequence.

[0144] The rating mutation rate refers to the relative change between the current rating and the smoothed rating, and is used to detect abnormal jumps in ratings.

[0145] This technical solution eliminates instantaneous noise and abnormal fluctuations through exponential moving average smoothing and abrupt change detection, maintains long-term stability of parameter influence scores, avoids abnormal parameter updates due to instantaneous disturbances in the boundary field, improves the robustness and continuous operational reliability of the forecast system, and ensures stable and controllable parameter updates.

[0146] In one embodiment, optionally, the composite index memory is constructed using the following algorithm:

[0147] Calculate the Pearson correlation coefficient between parameter updates:

[0148] ;

[0149] Group parameters with Corr>0.7 into the same parameter group;

[0150] The historical pattern field data was processed using an autoencoder to extract 128-dimensional pattern features. The convolutional neural network was used to extract 64-dimensional boundary features ψ from the boundary field data on the opposite side;

[0151] Use the locality-sensitive hash function family:

[0152] ;

[0153] Generate a composite index key of length L And bind the parameter group to the index key for storage.

[0154] The Pearson correlation coefficient is a statistical indicator that measures the degree of linear correlation between parameter updates and is used for parameter correlation grouping.

[0155] An autoencoder is a neural network structure used for unsupervised feature extraction and dimensionality reduction. It can output a 128-dimensional pattern representation vector, while a convolutional neural network is used to output a 64-dimensional boundary feature vector.

[0156] Locality Sensitive Hashing (LSH) is a hash algorithm that supports fast approximate retrieval of high-dimensional vectors.

[0157] A composite index key is a hash index generated by fusing pattern features and boundary features, used for fast retrieval in the memory.

[0158] This technical solution enables rapid retrieval of similar meteorological scenarios and reuse of parameter experience through parameter grouping, dual-feature encoding, and local sensitive hash indexing. This significantly improves online fine-tuning efficiency, reduces computational overhead, and allows the model to quickly reuse historically optimal parameters to improve forecast accuracy, thus adapting to high-frequency rolling forecast services.

[0159] In one embodiment, optionally, the elastic update employs a joint algorithm of momentum acceleration and boundary decay:

[0160] Selected in descending order of dynamic influence score One sensitive parameter;

[0161] Perform the update:

[0162] ;

[0163] ;

[0164] ;

[0165] Where β = 0.9, ;

[0166] Apply attenuation to boundary-sensitive parameters:

[0167] .

[0168] Among them, momentum acceleration is an optimization strategy that incorporates historical update gradients to improve the convergence speed and stability of parameter updates.

[0169] This technical solution achieves lightweight and efficient online fine-tuning by using Top-K sensitive parameter screening, momentum acceleration, and boundary decay joint updates. It ensures that the parameters are updated sufficiently and quickly adapt to weather changes, while avoiding excessive updates of boundary sensitive parameters, thereby improving model stability and forecast consistency.

[0170] In one embodiment, optionally, performing the energy conservation correction includes:

[0171] Calculate the total energy:

[0172] ;

[0173] Calculate the energy imbalance:

[0174] ;

[0175] Correcting potential temperature using projection operator:

[0176] .

[0177] Energy imbalance refers to the residual between the change of energy over time and the advection transport, characterizing the non-physical degree of energy.

[0178] The projection operator is a mathematical correction operator that removes non-physical energy components, and the projection coefficients are weighting coefficients used for energy correction to ensure energy conservation after correction. This scheme can project the potential temperature into the energy conservation space using the projection operator, thereby eliminating non-physical energy terms.

[0179] This technical solution, through total energy calculation and projection correction, forces the forecast field to meet the energy conservation constraint, eliminates non-physical energy anomalies, improves the physical rationality of the forecast results, reduces forecast drift caused by false heavy precipitation and energy imbalance, and enhances forecast credibility.

[0180] In one embodiment, optionally, the water material redistribution includes employing the Wasserstein optimal transport algorithm:

[0181] Detection of water substance imbalance:

[0182] ;

[0183] Solving the optimal transmission problem:

[0184] ;

[0185] Redistribution of the water vapor field:

[0186] .

[0187] Wasserstein optimal transport refers to a mathematical optimization algorithm that maintains a constant total amount while achieving optimal distribution adjustment.

[0188] Optimal transport mapping refers to a spatial mapping function that adjusts a non-equilibrium distribution to a conserved distribution. Water vapor field redistribution refers to the optimization and adjustment of the spatial distribution of water vapor while keeping the total water volume constant.

[0189] This technical solution achieves strict conservation and redistribution of water substances based on Wasserstein optimal transport, ensuring that the total water vapor content remains unchanged, avoiding false precipitation and water imbalance, improving the physical accuracy of precipitation location, intensity and evolution process, and improving the spatiotemporal continuity of forecasts.

[0190] In one embodiment, optionally, after the water material redistribution, the precipitation phase change process is further controlled using a threshold-triggered differential equation:

[0191] Cloud-water to rainwater conversion:

[0192] ;

[0193] Rainwater evaporation:

[0194] ;

[0195] in , and This is an empirical coefficient. This is an indicator function.

[0196] The precipitation phase transition process refers to the atmospheric physical process in which cloud water transforms into rainwater and rainwater evaporates back into water vapor. Critical water content refers to the critical threshold value that triggers the transformation of cloud water into rainwater.

[0197] Empirical coefficients are phase transition rate coefficients determined based on observations and physical laws.

[0198] This technical solution accurately simulates the physical processes of cloud-water-rainwater conversion and rainwater evaporation through threshold-triggered differential equations, ensuring that precipitation generation and dissipation conform to real atmospheric laws, further suppressing false precipitation and unreasonable intensity abrupt changes, and improving the reliability of short-term forecasts in severe convective scenarios.

[0199] The following is a more intuitive explanation of the technical solution provided to help those skilled in the art to better understand this solution.

[0200] This invention provides a cyclic parameter adaptive and lateral boundary field driven method for short-term precipitation forecasting. The method introduces a boundary-aware parameter influence assessment system to construct boundary influence factors, achieving a quantitative mapping from lateral boundary field gradients to parameter updates. It also constructs a joint index memory of model features and lateral boundary field features to enable rapid retrieval of similar meteorological scenarios and reuse of parameter experience. Furthermore, it establishes a real-time correction system with conservation constraints to achieve energy conservation correction and water redistribution.

[0201] To achieve the above objectives, the present invention provides the following technical solution, comprising the following steps:

[0202] Step 1: Construct a boundary-aware parameter influence assessment system;

[0203] Dual-modal sensing mechanism:

[0204] Gradient trajectory analysis: Records the historical gradient changes of parameters during training and calculates the mean. ,variance and autocorrelation coefficient ;

[0205] Feature activation analysis: Calculates the correlation between parameters and features through channel attention. ;

[0206] Influence scores are generated through gating fusion. .

[0207] Enhanced boundary constraints:

[0208] Calculate the boundary influence factor: ;

[0209] Adjust rating: .

[0210] Sliding window optimization:

[0211] Using exponential moving average scoring: .

[0212] Step 2: Establish a system for enhanced fine-tuning and memory optimization;

[0213] Composite index memory construction:

[0214] Calculate the correlation coefficient. ,

[0215] Those with a correlation coefficient greater than 0.7 were grouped together.

[0216] Extracting pattern features (128-dimensional autoencoder vector) and boundary features (64-dimensional CNN features)

[0217] Using a family of hash functions for composite indexes:

[0218] .

[0219] Real-time retrieval and optimization:

[0220] Calculate the dynamic weighted similarity:

[0221] ;

[0222] Perform dual-path optimization: if a match is successful, load the parameter module and perform L-BFGS optimization; if a match fails, activate Top-K parameter momentum update.

[0223] Boundary constraint update:

[0224] Apply attenuation to sensitive parameters .

[0225] Step 3: Implement real-time correction of conservation constraints;

[0226] Energy conservation correction:

[0227] Calculate the total energy:

[0228] ;

[0229] Perform projection correction:

[0230] .

[0231] Water substance redistribution

[0232] Wasserstein's optimal transmission:

[0233] .

[0234] Precipitation phase change monitoring:

[0235] Achieving cloud water-rainwater conversion:

[0236] .

[0237] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation scenarios.

[0238] Figure 2 This is a schematic diagram comparing the principles of the present invention with those of the prior art, provided in this embodiment. For example... Figure 2 As shown, this invention provides a real-time data assimilation and short-term precipitation forecasting system based on adaptive cyclic parameters, lateral boundary field driving, and physical conservation constraints. Figure 2 Figure 101 illustrates a traditional unconstrained forecasting method. This method directly feeds multiple inputs, such as historical meteorological fields, real-time observation fields, and cloud images, into a black-box neural network model. Although it can generate precipitation predictions, the lack of physical constraints and conservation constraint correction mechanisms within the model makes it prone to problems such as mass / energy imbalances and false local heavy precipitation, making it difficult to meet the requirements of physical consistency and stability for regional short-term precipitation warnings.

[0239] Figure 2Figure 102 illustrates the physical constraint forecasting method of the present invention. The present invention inputs historical model fields, real-time observation data, and regional lateral boundary field data into a prediction model with physical conservation constraints. At the model output, energy projection correction and a water material redistribution mechanism based on Wasserstein transmission are introduced to ensure that the prediction results meet the requirements of mass conservation and energy balance during the generation process, thereby improving the physical rationality and spatiotemporal continuity of precipitation forecast results.

[0240] Figure 2 Figure 103 illustrates the time-progression method of existing static parameter forecast models. In this method, the model parameters are basically fixed after pre-training, and the same prediction model is driven to make hourly predictions only by the observation input and lateral boundary field input at different times. The parameter blocks inside the model remain static and cannot respond in real time to the evolution of upstream weather systems, lateral boundary field disturbances, and the development of local severe convection. Therefore, forecast lag and error accumulation are likely to occur under rapidly evolving weather conditions.

[0241] Figure 2 Figure 104 illustrates the cyclic assimilation and online parameter update mechanism of the present invention. Between adjacent time steps, the system continuously receives observation inputs and, in conjunction with cyclic assimilation, online fine-tuning, and dynamic parameter update strategies, adaptively adjusts some sensitive parameters of the forecast model. Specifically, the system detects the intensity of changes in the lateral boundary field through boundary gradient triggers. When the boundary disturbance reaches a preset threshold, it retrieves a parameter set similar to the current weather scenario from the joint index memory and performs parameter set retrieval, loading, and update operations. This allows the model parameters to be adjusted in real time according to boundary evolution and observational changes, improving the ability to track sudden heavy precipitation events.

[0242] Figure 2 Figure 105 illustrates the overall architecture of the real-time data assimilation and short-term forecasting system. This system centers on a regional forecast domain and, driven by the background field of a global or large-scale model, receives real-time meteorological observation data, radar echo data, satellite cloud imagery data, and lateral boundary field information. Through a cyclic assimilation node, external observations are continuously introduced into the system, and the regional forecast status is updated under the coupling lateral boundary forcing. Simultaneously, the system utilizes an adaptive parameter update module to dynamically correct key parameters within the model, ultimately generating real-time meteorological data assimilation and short-term precipitation forecast results for the target region.

[0243] Furthermore, Figure 3This is a schematic diagram of the overall execution flow of the system provided in the embodiments of this application, which can be summarized as follows: First, acquire multi-source observation data, historical model data, and lateral boundary field data; then, perform a boundary-aware parameter influence assessment to identify model parameters sensitive to changes in the current weather situation; next, perform rapid matching and enhanced fine-tuning of parameter groups based on the joint index memory; then, implement energy conservation projection correction and optimal transport redistribution of water and matter on the prediction field; finally, output the prediction results of short-term precipitation intensity and spatial distribution within the target time period, feed back new observation information to the next cycle to achieve real-time rolling updates, and continue to feed back new observation information to the next cycle to achieve real-time rolling updates.

[0244] Compared with traditional black-box static parameter forecasting methods, this invention introduces lateral boundary driving, cyclic assimilation, adaptive parameter updating, and mass and energy conservation constraints, enabling the model to maintain the efficient inference capabilities of deep learning while possessing stronger boundary response capabilities, physical consistency, and online correction capabilities. It is particularly suitable for short-term warning scenarios such as severe convection, typhoon periphery precipitation, and localized sudden rainstorms.

[0245] Figure 4 This is a schematic diagram illustrating the specific process of boundary-aware parameter influence assessment provided in an embodiment of this application. For example... Figure 4 As shown, the specific implementation steps for assessing the influence of boundary-aware parameters are as follows:

[0246] In parameter influence assessment, an innovative dual-modal perception mechanism was used, which can evaluate the dynamic importance of parameters in each layer of the neural network in real time and quantify their contribution to the evolution of meteorological models. The dual-modal perception mechanism incorporates gradient trajectory modal analysis and feature activation modal analysis. During the training phase, for each parameter W... i Maintain a gradient queue ΔW of length T. i (1:T) Record the direction and magnitude of its most recent T updates. Perform time-frequency analysis on the gradient sequence of each parameter, and calculate the mean μ. i σ i 2 and extracting the autocorrelation coefficient ρ i (τ), the above features are input into two fully connected layers, and the gradient influence score is output. During model inference, for each layer's output feature map F∈R... C×H×W Perform global average pooling to compute channel attention weights. For each parameter W of this layer... i Calculate its correlation with the feature channels. (W i (connected to the k-th channel), where Here, K is the indicator function, and K is the number of channels. The gated fusion unit can adjust the fusion coefficients based on the signal-to-noise ratio of the current input data. The final output is an influence score. .

[0247] Extracting the spatial gradient of the boundary field Input three fully connected layers to calculate the boundary influence factor. Calculate the parameter update magnitude Where W5 and b5 are the weights and biases of the first fully connected layer of the MLP, and W6 and b6 are the weights and biases of the second fully connected layer of the MLP, thus dynamically adjusting the score. A sliding window optimizer is designed to eliminate instantaneous noise in parameter importance scores and extract stable long-term trends. In the sliding window optimizer, an exponential moving average is applied to the score sequence of each parameter. ,in Control the intensity of historical memory; calculate the rate of change of scores between adjacent time steps. To perform mutation detection and reset.

[0248] Then, the model is enhanced and fine-tuned, and a joint index memory is built. Figure 5 This is a schematic diagram illustrating the specific process of enhanced fine-tuning and memory optimization provided in the embodiments of this application. For example... Figure 5 As shown, the specific implementation steps for enhancing fine-tuning and parameter memory system optimization are as follows:

[0249] Update the correlation coefficient by calculating parameters. Grouping parameters with a correlation coefficient > 0.7 into the same parameter group. Extracting pattern features. and boundary features Construct a family of LSH hash functions Generate hash keys .

[0250] Weighted similarity calculation If a similar pattern is found in the memory (Cosine similarity > 0.85), then load the associated parameter group { Constrained optimization was performed using the L-BFGS algorithm. A flexible parameter update strategy was used to fine-tune the model, and all network parameters were sorted in descending order of their smoothed parameter influence scores, selecting those that met the criteria. The first K parameters of the condition, Given the total number of parameters, the Top-K parameters are obtained. Momentum acceleration is used for updating, and the update formula is as follows:

[0251] ;

[0252] , ;

[0253] Where β = 0.9, Adaptive adjustment based on signal-to-noise ratio:

[0254] ;

[0255] For ΔW i Apply boundary constraints to additionally attenuate boundary-sensitive parameters. .

[0256] If no matching pattern is found in the memory, activate the Top-K parameter update channel and perform elastic parameter updates to generate the corrected prediction field. .

[0257] Figure 6 This is a schematic diagram illustrating the specific process of real-time correction of conservation constraints provided in the embodiments of this application. For example... Figure 6 As shown, the implementation steps of the real-time correction system for conservation constraints are as follows:

[0258] The total energy is obtained through the projection method. and energy imbalance Using projection coefficients Correcting the potential temperature Detection of water substance imbalance Wasserstein optimal transmission is adopted. Achieving water substance redistribution. Water vapor quantity is also redistributed. The above results are then input into the precipitation physical balance module to monitor the phase change process and realize the conversion of cloud water to rainwater. Evaporation of rainwater Set a critical value .

[0259] This invention constructs a boundary-aware parameter influence assessment framework. By introducing boundary influence factors and dynamically adjusting parameter influence scores, it achieves quantitative coupling between the lateral boundary field and the model's internal parameter update process. It also constructs a joint index memory that integrates meteorological model features and lateral boundary field features to achieve similar scene matching and rapid parameter group loading. Furthermore, it designs a water and material redistribution algorithm based on Wasserstein optimal transmission and an energy projection correction operator to maintain energy and water and material balance during neural network prediction, thereby reducing non-physical spurious precipitation. This method, while maintaining the efficient inference characteristics of deep learning, overcomes the limitations of static model parameters, improving the response to lateral boundary field disturbances and the physical consistency of prediction results.

[0260] This application also provides a cyclic parameter adaptive and lateral boundary field driven system for short-term precipitation forecasting. Figure 7 This is a schematic diagram of the cyclic parameter adaptive and lateral boundary field driven system for short-term precipitation forecasting provided in this embodiment, as shown below. Figure 7 As shown, the system includes:

[0261] The data input module 701 is used to acquire the lateral boundary field data, historical model field data and real-time observation data of the target area, and input them into the neural network prediction model to be updated.

[0262] The dynamic influence score determination module 702 is used to calculate the gradient influence component based on the gradient trajectory of each parameter in the neural network prediction model; and to calculate the feature influence component based on the activation correlation between each parameter and the current input feature; to fuse the gradient influence component and the feature influence component to obtain the basic score, and to perform boundary enhancement on the basic score according to the spatial gradient of the side boundary field to obtain the dynamic influence score.

[0263] The side boundary field feature determination module 703 is used to encode the historical pattern field data to obtain the pattern features of the current time; and to encode the side boundary field data to obtain the side boundary field features.

[0264] The model parameter optimization module 704 is used to match the mode features and the side boundary field features with the pre-built joint index memory; if the match is successful, the corresponding historical parameter group is loaded and local optimization is performed; if the match fails, sensitive parameters are selected based on the dynamic influence score and elastic updates are performed, and attenuation constraints are applied to the boundary sensitive parameters during the update process.

[0265] The model output and result correction module 705 is used to generate the original precipitation forecast field using the updated neural network forecast model, and sequentially perform energy conservation correction and water material redistribution to obtain a short-term precipitation forecast field that satisfies the mass and energy conservation constraints.

[0266] The forecast update module 706 is used to take the short-term precipitation forecast field and the real-time observation data of the next time period as new inputs, and return to execute the aforementioned process to realize rolling forecast updates.

[0267] It is understood that the cyclic parameter adaptive and lateral boundary field driven system for short-term precipitation forecasting provided in this embodiment has functional modules and beneficial effects corresponding to the above-mentioned cyclic parameter adaptive and lateral boundary field driven method for short-term precipitation forecasting. To avoid repetition, these will not be described again here.

[0268] This application also provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiments of the cyclic parameter adaptation and lateral boundary field driving method for short-term precipitation forecasting, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0269] It should be noted that the electronic devices in the embodiments of this application include mobile electronic devices and non-mobile electronic devices as described above.

[0270] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiments of the cyclic parameter adaptation and lateral boundary field driving method for short-term precipitation forecasting, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0271] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0272] This application also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described embodiments of the cyclic parameter adaptive and lateral boundary field driven method for short-term precipitation forecasting, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0273] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in a different order, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0274] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A cyclic parameter adaptive and lateral boundary field driven method for short-term precipitation forecasting, characterized in that, The method includes: Acquire lateral boundary field data, historical model field data, and real-time observation data of the target area, and input them into the neural network prediction model to be updated; Based on the gradient trajectory of each parameter in the neural network prediction model, the gradient influence component is calculated; and based on the activation correlation between each parameter and the current input feature, the feature influence component is calculated; the gradient influence component and the feature influence component are fused to obtain the basic score, and the basic score is enhanced by the spatial gradient of the side boundary field to obtain the dynamic influence score. The historical pattern field data is encoded to obtain the pattern features of the current time period; the side boundary field data is encoded to obtain the side boundary field features. The pattern features and the side boundary field features are matched with a pre-built joint index memory. If the match is successful, the corresponding historical parameter group is loaded and local optimization is performed. If the match fails, sensitive parameters are selected based on the dynamic influence score and elastic updates are performed, and attenuation constraints are applied to the boundary sensitive parameters during the update process. The original precipitation forecast field was generated using the updated neural network forecast model, and energy conservation correction and water redistribution were performed sequentially to obtain a short-term precipitation forecast field that satisfies the mass and energy conservation constraints. The short-term precipitation forecast field and the real-time observation data of the next time period are used together as new inputs, and the aforementioned process is repeated to achieve rolling forecast updates.

2. The method according to claim 1, characterized in that, The basic score is enhanced by applying boundary enhancement based on the spatial gradient of the lateral boundary field to obtain a dynamic influence score, including: For the i-th parameter W in the neural network prediction model i Maintain a gradient queue ΔW of length T. i (1:T) Calculate the gradient mean μ i σ i 2 and autocorrelation coefficient ρ i (τ), the influence component of the gradient output from the two fully connected layers S grad (i) ; For the current input feature map F∈R C×H×W Perform global average pooling to obtain the channel attention weights A. C Calculate the correlation between the parameter and the k-th channel I(W) i The feature influence components are obtained by aggregation (k). ; Calculate the fusion coefficient using the gating unit. Receive a basic score ;in, This is the transpose of the learnable weight vector for the gating unit; For real-time observation of the data field; For gating unit bias terms; To observe the signal-to-noise ratio of the data, This is a function for calculating the standard deviation of the observation field. According to the spatial gradient of the side boundary field The boundary influence factor is output by the multilayer perceptron. The base score is modulated to obtain a dynamic influence score.

3. The method according to claim 2, characterized in that, After obtaining the dynamic influence score, the method further includes a score stability control step: Apply an exponential moving average to the dynamic influence score: ； in, The smoothing coefficient for the exponential moving average; This is the current forecast time. The adjusted dynamic influence score after being modulated by the lateral boundary field gradient; Calculate the score mutation rate: ； When τ(t) is greater than the preset threshold, reset the smoothing score and re-execute the parameter influence assessment.

4. The method according to claim 1, characterized in that, The composite index memory is constructed using the following algorithm: Calculate the Pearson correlation coefficient between parameter updates: ； in, Let i be the update amount for the i-th parameter. Let j be the update amount for the j-th parameter. These are the parameters of the i-th and j-th network models, respectively; Group parameters with Corr>0.7 into the same parameter group; The historical pattern field data was processed using an autoencoder to extract 128-dimensional pattern features. The convolutional neural network was used to extract 64-dimensional boundary features ψ from the boundary field data on the opposite side; Use the locality-sensitive hash function family: ;in, This is the transpose of the linear weight vector of the hash function. The hash bias constant. This is the hash scaling factor; Generate a composite index key of length L And bind the parameter group to the index key for storage.

5. The method according to claim 4, characterized in that, The elastic update employs a joint algorithm of momentum acceleration and boundary decay: Selected in descending order of dynamic influence score One sensitive parameter; Perform the update: ； ； ； Where β = 0.9, ; Apply attenuation to boundary-sensitive parameters: in, This represents the total number of parameters in the neural network model. denoted as the gradient of the loss function; O represents the observed precipitation field, and P represents the model output forecast precipitation field; This represents the update amount of the i-th parameter at time t. These are the original parameters at time t; The gradient of the side boundary field.

6. The method according to claim 1, characterized in that, The energy conservation correction includes: Calculate the total energy: ； Calculate the energy imbalance: ； Correcting potential temperature using projection operator: ； in, The specific heat of air at constant pressure; Atmospheric temperature; The latent heat of vaporization of water vapor; The atmospheric specific humidity; This is the horizontal wind speed vector; It is the acceleration due to gravity; Vertical height; This refers to atmospheric potential temperature.

7. The method according to claim 1, characterized in that, The water material redistribution includes the use of the Wasserstein optimal transport algorithm: Detection of water substance imbalance: ;in, The residual of the imbalance between water material budget; This refers to surface evaporation. This refers to the amount of water vapor condensation. Solving the optimal transmission problem: ;in, , These are the source water vapor grid coordinates and the target water vapor grid coordinates, respectively. For the divergence of the transport mapping function; To map to Water vapor transport function of coordinates; Redistribution of the water vapor field: ;in, The original water vapor field before redistribution; This represents the spatial transport mapping relationship for water vapor.

8. The method according to claim 7, characterized in that, Following the water redistribution, the process also includes controlling the precipitation phase change process using a threshold-triggered differential equation: Cloud-water to rainwater conversion: ； Rainwater evaporation: ； in , The cloud-water mixing ratio, The rainwater mixing ratio; This is an empirical coefficient for the conversion of cloud water into rainwater. This is the empirical coefficient for rainwater evaporation. It is the acceleration due to gravity; This is an indicator function.

9. A cyclic parameter adaptive and lateral boundary field driven system for short-term precipitation forecasting, characterized in that, The system includes: The data input module is used to acquire the lateral boundary field data, historical model field data and real-time observation data of the target area, and input them into the neural network prediction model to be updated. The dynamic influence score determination module is used to calculate the gradient influence component based on the gradient trajectory of each parameter in the neural network prediction model; and to calculate the feature influence component based on the activation correlation between each parameter and the current input feature; to fuse the gradient influence component and the feature influence component to obtain the basic score, and to perform boundary enhancement on the basic score according to the spatial gradient of the side boundary field to obtain the dynamic influence score. The side boundary field feature determination module is used to encode the historical pattern field data to obtain the pattern features of the current time; and to encode the side boundary field data to obtain the side boundary field features. The model parameter optimization module is used to match the mode features and the side boundary field features with a pre-built joint index memory. If the match is successful, the corresponding historical parameter group is loaded and local optimization is performed. If the match fails, sensitive parameters are selected based on the dynamic influence score and elastic updates are performed. During the update process, attenuation constraints are applied to the boundary sensitive parameters. The model output and result correction module is used to generate the original precipitation forecast field using the updated neural network forecast model, and then perform energy conservation correction and water material redistribution in sequence to obtain a short-term precipitation forecast field that satisfies the mass and energy conservation constraints. The forecast update module is used to take the short-term precipitation forecast field and the real-time observation data of the next time period as new inputs, and return to execute the aforementioned process to realize rolling forecast updates.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the cyclic parameter adaptive and lateral boundary field driven method for short-term precipitation forecasting as described in any one of claims 1-8.

Images