Extreme weather prediction method based on typical weather year

Abstract

The invention relates to the technical field of meteorological prediction, in particular to an extreme meteorological prediction method based on a typical meteorological year, which comprises the following steps: firstly, constructing a data set of the typical meteorological year, and generating high-quality meteorological data through interpolation and fusion technologies in combination with ground, reanalysis and satellite data; the generative adversarial network is used for enhancing data, the problem that extreme event samples are insufficient is solved, and key meteorological variables are extracted through super-resolution and feature dimension reduction; based on a convolutional neural network and a long-short-term memory network, constructing a deep learning model to perform multi-task prediction; a physical constraint rule is introduced, such as energy and mass conservation, and it is ensured that prediction conforms to a meteorological rule; and through an uncertainty quantification technology, generating a confidence interval of a prediction result, and providing risk assessment and decision support information. According to the method, the precision and credibility of extreme meteorological event prediction are improved, and disaster prevention and reduction, resource scheduling, agricultural production and the like are supported.

Description

Technical Field

[0001] The present invention relates to the field of meteorological forecasting technology, and in particular to an extreme meteorological forecasting method based on a typical meteorological year. Background Art

[0002] Extreme weather events (such as heavy rain, typhoons, high temperatures and droughts) pose a serious threat to human production and life, economic operations and the ecological environment due to their strong destructiveness and suddenness. With the intensification of global climate change, the frequency and intensity of extreme weather events have shown a clear upward trend. This trend has put forward higher weather forecasting requirements in the fields of disaster prevention and mitigation, energy scheduling and agricultural production. However, existing technologies have many deficiencies in the field of extreme weather event prediction and are difficult to meet the increasingly complex application requirements.

[0003] Deficiencies in data construction:

[0004] Meteorological data is the basis for the prediction of extreme meteorological events, and its quality and coverage directly determine the effectiveness of the prediction model. However, the current data construction method mainly relies on a single data source, such as ground observation data or satellite remote sensing data. These data have significant defects in temporal and spatial coverage, resolution and completeness:

[0005] Limitations of ground observation data: Although the temporal resolution is high, the limited distribution of observation sites cannot fully reflect regional or global meteorological changes, especially in remote areas and ocean regions, where data scarcity is a prominent problem.

[0006] Problems with satellite remote sensing data: Although the spatial coverage is wide, the temporal resolution is low, and the data quality is unstable due to cloud interference and atmospheric pollution.

[0007] Bias of reanalysis data: Although reanalysis data has global coverage, its resolution is relatively low and there are certain errors in the description of local features. In addition, meteorological data are often accompanied by missing values ​​and outliers. Existing interpolation and correction methods (such as simple interpolation and linear regression) are difficult to effectively supplement complex spatiotemporal dynamics, resulting in insufficient representativeness and completeness of the data. This deficiency directly affects the input quality of the prediction model, making subsequent predictions difficult to be accurate.

[0008] Limitations of extreme event modeling capabilities:

[0009] The formation of extreme meteorological events is highly complex, involving nonlinear dynamic processes and multi-scale interactions. However, existing technologies still have the following problems in event modeling:

[0010] The model’s ability to handle nonlinear characteristics is insufficient: Traditional statistical methods (such as generalized extreme value distribution and linear regression) usually assume that the data distribution conforms to a specific probability model, but the characteristics of extreme events deviate significantly from the normal distribution and have a long-tail distribution characteristic, which makes traditional models perform poorly in predicting event intensity.

[0011] The impact of sample scarcity on modeling: Extreme events occur at a low frequency and training samples are scarce, which makes traditional models prone to overfitting. This problem is particularly serious in event modeling at the regional scale.

[0012] Separation of spatiotemporal dynamic characteristics: Existing models usually treat meteorological time series and spatial distribution characteristics separately, lack the ability to jointly model the two, and it is difficult to accurately describe the dynamic evolution process of extreme meteorological events.

[0013] The meteorological system is a complex dynamic and thermodynamic system, and its evolution process strictly follows the laws of physics (such as conservation of energy, conservation of mass, and conservation of momentum). Existing extreme weather forecasting methods mostly rely on data-driven models (such as traditional machine learning and deep learning). Although they can capture the statistical patterns of data, they have significant deficiencies in the following aspects:

[0014] Ignoring the laws of meteorological physics: Existing methods fail to effectively embed the basic physical laws of the meteorological system, which may result in the prediction results being consistent with the data trend but violating basic meteorological constraints such as conservation of energy or mass, reducing the scientific nature of the prediction results.

[0015] Dynamic mechanisms are not fully utilized: The formation of extreme events is usually affected by multiple factors such as atmospheric circulation and the interaction between the surface and the atmosphere. Existing methods fail to combine these dynamic mechanisms for modeling, making it difficult to achieve a scientific description of the complex process of the event.

[0016] The lack of physical consistency not only limits the explanatory power of the model, but also makes the prediction results have great uncertainty in scientific research and practical applications.

[0017] Therefore, the present invention proposes an extreme weather forecasting method based on a typical meteorological year to address the deficiencies of the prior art. Summary of the invention

[0018] In view of the shortcomings of the existing technologies in data construction, nonlinear modeling capabilities, physical consistency and uncertainty assessment, the present invention provides an extreme weather forecasting method based on a typical meteorological year, aiming to comprehensively improve the accuracy, reliability and applicability of extreme weather event forecasts.

[0019] To achieve the above objectives, the present invention is implemented by the following technical solutions: An extreme weather forecasting method based on a typical weather year comprises the following steps:

[0020] Typical meteorological year data construction: Acquire meteorological data, including ground observation data, reanalysis data and satellite remote sensing data, clean and fuse the data, and construct a typical meteorological year;

[0021] Data enhancement and feature extraction: Expand extreme event data samples through data generation and enhancement methods, perform feature engineering and dimensionality reduction;

[0022] Deep learning modeling: build a spatiotemporal feature extraction model to perform multi-task learning on meteorological spatiotemporal data;

[0023] Integration of physical constraints: Introducing physical conservation rules to correct model prediction results;

[0024] Uncertainty quantification: quantify the uncertainty of prediction results through sampling methods or multi-model ensemble methods;

[0025] Result deployment and application: Provide data support based on prediction results to guide optimization and decision-making in specific scenarios.

[0026] Preferably, the typical meteorological year data construction includes the following contents:

[0027] The missing values ​​in the meteorological time series are interpolated using the local weighted regression method, and the formula is:

[0028]

[0029] in, represents the interpolated value at time point t, y i Indicates time point t i The existing observations, n represents the number of time points used for interpolation, w i represents the time weight, and λ represents the smoothing parameter that controls the weight decay rate.

[0030] Preferably, the data enhancement includes using a generative adversarial network to expand extreme weather event data samples, wherein the generator generation formula is:

[0031] G(z)=σ(W g ·z+b g );

[0032] Where z represents the input noise loudness, which follows a standard normal distribution. W g represents the weight matrix of the generator network, b g Represents the bias vector of the generator network, and σ represents the activation function, including ReLU and Sigmoid.

[0033] Preferably, the feature dimension reduction adopts the principal component analysis method, and the dimension reduction formula is:

[0034] Z = XW, W = argmaxTr(W T SW);

[0035] in, represents the original feature matrix, n is the number of samples, and d is the feature dimension; It represents the feature matrix after dimensionality reduction, k is the number of principal components; represents the projection matrix, which maps high-dimensional features to low-dimensional space; S represents the covariance matrix, which is defined as Tr(·) represents the matrix trace operation.

[0036] Preferably, the deep learning modeling includes constructing a spatiotemporal feature extraction model combining a convolutional neural network and a long short-term memory network, wherein the output formula of the convolutional layer is:

[0037]

[0038] in, K represents the value of position (i, j) in the output feature graph of the ,th layer of convolution; mn Represents the weight of the convolution kernel at position (m, n); Indicates the value of the position (i+m, j+n) of the input symbol graph; b (l) Represents the bias of the ,th layer of convolution; M and N represent the height and width of the convolution kernel respectively; σ represents the activation function.

[0039] Preferably, the time series processing of the long short-term memory network includes the following formula:

[0040] Cell status update:

[0041] c t =f t ⊙c t-1 +i t ⊙tanh(W c ·[h t-1 , x t ]+b c );

[0042] Hide status updates:

[0043] h t =o t ⊙tanh(c t );

[0044] Among them, c t Indicates cell status; h t represents the hidden state; f t 、i t , o t Respectively represent the forget gate, input gate, and output gate; Wc represents the cell state weight matrix; x t represents the input feature at time point t; ⊙ represents the element-by-element multiplication operation;

[0045] Preferably, the physical constraint integration includes:

[0046] Introducing the energy conservation constraint, the formula is:

[0047]

[0048] Among them, L energy represents the energy conservation loss, represents the heat flux divergence at the ith grid point, Q i represents the heat source at the ith grid point, Represents the time rate of change of the heat source.

[0049] Preferably, the uncertainty quantification adopts the MC-Dropout sampling method, and its formula is:

[0050]

[0051] in, Represents sample x i The expected prediction result of f(x i θ (m) ) represents the prediction result of the model under the mth random sampling, θ (m) represents the network parameters of the mth sampling, and M represents the number of sampling times.

[0052] Preferably, the result deployment and application includes:

[0053] Generate a risk level assessment of the meteorological event based on the forecast results, the assessment including:

[0054] According to the probability and intensity of extreme meteorological events, the events are divided into low risk, medium risk and high risk levels;

[0055] Combined with the prediction confidence interval, the reliability evaluation of the risk level is adjusted;

[0056] Provide risk-based decision support information, including:

[0057] Provide risk warning tips to clarify the time, location and impact range of possible events;

[0058] Provide optimization solutions for specific scenarios to guide scenario optimization and resource scheduling;

[0059] The prediction results are output in a visual way, including risk level distribution map, time dynamic prediction map and uncertainty distribution map of prediction results.

[0060] Preferably, an extreme weather forecasting system based on a typical meteorological year comprises:

[0061] Data acquisition module, used to obtain meteorological observation data, reanalysis data and satellite remote sensing data;

[0062] Data processing module for data cleaning, enhancement and feature extraction;

[0063] Deep learning module for spatiotemporal feature extraction and multi-task learning;

[0064] Physical constraint module, used to correct model output;

[0065] Uncertainty quantification module, used to quantify the uncertainty of prediction results;

[0066] Application module for generating forecast reports and supporting scenario optimization decisions.

[0067] The present invention provides an extreme weather forecasting method based on a typical meteorological year. It has the following beneficial effects:

[0068] 1. The present invention integrates ground observation data, reanalysis data and satellite remote sensing data, and comprehensively utilizes the advantages of different data sources to ensure the extensiveness and accuracy of meteorological data. At the same time, the local weighted regression (LOESS) method is used to interpolate missing data, and a high-quality typical meteorological year data set with complete spatiotemporal distribution is constructed, achieving the technical effect of improving data integrity, spatiotemporal resolution and reliability. Compared with the technical solutions in the prior art that rely on a single data source or use simple interpolation methods, the present invention significantly solves the problems of insufficient data coverage, data imbalance and low data quality, and lays a solid data foundation for subsequent extreme weather forecasts.

[0069] 2. The present invention uses a generative adversarial network (GAN) to generate virtual samples of extreme meteorological events, significantly increasing data diversity, and combines convolutional neural networks (CNN) and long short-term memory networks (LSTM) to achieve efficient extraction and modeling of spatiotemporal characteristics of meteorological data. At the same time, the deep learning framework supports dual-task learning of classification and regression, and can predict the probability of occurrence and intensity of extreme events. Through the above scheme, the present invention significantly improves the accuracy and stability of extreme meteorological event predictions, and achieves the technical effect of accurately modeling the nonlinear relationship between meteorological variables. Compared with the technical solutions that rely on statistical methods or traditional machine learning in the prior art, the present invention effectively solves the problems of scarce extreme event samples and insufficient complex spatiotemporal dynamic modeling, ensuring that the prediction results are closer to actual meteorological events.

[0070] 3. Based on deep learning modeling, the present invention corrects the model output by introducing physical constraint rules (such as conservation of energy and conservation of mass), and combines soft and hard constraint methods to ensure that the prediction results conform to the laws of meteorological dynamics and thermodynamics. Specifically, the energy conservation constraint is used to correct the heat exchange relationship between meteorological variables, and the mass conservation constraint ensures that the predicted wind speed field distribution conforms to the atmospheric dynamic balance. Through the above scheme, the present invention achieves the technical effect of ensuring the physical consistency of the prediction results while improving the prediction accuracy. Compared with the technical solutions in the prior art that rely on pure data-driven models, the present invention effectively solves the problem that the model prediction results may violate the laws of meteorological physics, and improves the scientificity and credibility of the prediction system.

[0071] 4. The present invention quantifies the uncertainty of the model prediction results through the MC-Dropout method, calculates the mean and confidence interval of the prediction value, and constructs an assessment framework based on the risk level in combination with the probability and intensity prediction results of extreme events. In addition, the present invention further adjusts the reliability evaluation of the risk level in combination with uncertainty information, and provides specific decision support information, including risk warning prompts, event occurrence time and impact range prediction, etc. In the result deployment, dynamic risk maps and prediction confidence intervals are presented through a visual interface, which improves the practical applicability of the system. The above scheme achieves the technical effect of improving the reliability, credibility and operability of the prediction results. Compared with the lack of technical solutions for prediction uncertainty assessment in the prior art, the present invention solves the problem that the existing system cannot provide quantitative credibility and risk level support in high-risk scenarios, and further enhances the practical value of the prediction results in complex application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0073] Figure 1 This is a flow chart of the extreme weather prediction method based on a typical meteorological year. DETAILED DESCRIPTION

[0074] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] Please see attached Figure 1 :

[0076] This embodiment provides an extreme weather forecasting method based on a typical weather year, comprising the following steps:

[0077] Typical meteorological year data construction: Acquire meteorological data, including ground observation data, reanalysis data and satellite remote sensing data, clean and fuse the data, and construct a typical meteorological year;

[0078] In this embodiment, meteorological data is first obtained from multiple data sources, specifically including the following:

[0079] Ground observation data:

[0080] Observational data from meteorological stations, including variables such as temperature, humidity, wind speed, precipitation, etc.

[0081] These data have high temporal resolution, but the observation sites are limited in distribution and the spatial coverage is insufficient.

[0082] Reanalysis data:

[0083] For example, ERA5 or MERRA2 reanalysis data provide temporal and spatial distribution information of the global meteorological field.

[0084] Reanalysis data are usually of lower resolution but have better global coverage.

[0085] Satellite remote sensing data:

[0086] For example, high-resolution meteorological data such as MODIS and TRMM are used to make up for the insufficient spatial coverage of ground observations.

[0087] This type of data is easily affected by clouds and requires subsequent cleaning.

[0088] In this embodiment, the acquired data is cleaned and preprocessed, mainly including the following steps:

[0089] Data outlier processing:

[0090] For outliers in the observed data, a Z-Score-based method is used to determine whether they are outliers:

[0091]

[0092] in:

[0093] x is the current data point;

[0094] μ and σ are the mean and standard deviation of the variable, respectively;

[0095] k is the threshold, usually 3.

[0096] Missing value interpolation:

[0097] The missing values ​​are interpolated using the locally weighted regression (LOESS) method:

[0098]

[0099] in, represents the interpolated value at time point t, y i Indicates time point t i The existing observations, n represents the number of time points used for interpolation, w i represents the time weight, and λ represents the smoothing parameter that controls the weight decay rate.

[0100] In this embodiment, multi-source data fusion is performed on the cleaned data, which specifically includes the following steps:

[0101] Multi-source data fusion method:

[0102] The weighted fusion method is used to integrate ground observation data, reanalysis data and satellite remote sensing data:

[0103] D(x, t) = αD obs (x,t)+βD rean (x,t)+γD sat (x, t);

[0104] in:

[0105] D obs (x, t): ground observation data;

[0106] D rean (x, t): reanalysis data;

[0107] D sat (x, t): satellite remote sensing data;

[0108] α, β, γ∈[0,1]: weight factors, satisfying α+β+γ=1. The weights can be adjusted according to the reliability of the data source.

[0109] Fusion strategy:

[0110] When there are multiple data sources in the region, sources with high data accuracy (such as ground observation data) are used first, and reanalysis data or satellite remote sensing data are used to supplement the missing areas.

[0111] In this embodiment, constructing a typical meteorological year based on the fused meteorological data specifically includes the following steps:

[0112] Selection of typical meteorological years:

[0113] According to the statistical characteristics of multi-year meteorological data, percentile distribution and long-term average methods are used to screen representative annual meteorological data.

[0114] Taking precipitation as an example, by calculating the percentile values ​​of precipitation distribution over many years (such as the 25th, 50th, and 75th percentiles), the year closest to these distribution characteristics is selected as the typical meteorological year.

[0115] Build results:

[0116] The constructed typical meteorological year data includes multivariate and multi-time resolution meteorological information, covering regional-scale meteorological characteristics, and is used to support subsequent extreme event predictions.

[0117] In this embodiment, the output of the typical meteorological year data set is:

[0118] Data format: time-space point data, covering meteorological variables such as temperature, humidity, precipitation, wind speed, etc.

[0119] Data resolution: The temporal resolution can reach the hourly level, and the spatial resolution supports the regional scale.

[0120] Data integrity: After cleaning and interpolation, the temporal and spatial continuity and representativeness of the data are improved.

[0121] Through the specific implementation of the above steps, a high-quality dataset of a typical meteorological year was constructed, which solved problems such as insufficient coverage of a single data source, inconsistency of multi-source data, and missing and anomaly data, providing a reliable data basis for the prediction of subsequent extreme meteorological events.

[0122] Data enhancement and feature extraction: Expand extreme event data samples through data generation and enhancement methods, perform feature engineering and dimensionality reduction;

[0123] In this embodiment, data enhancement is first performed, which specifically includes the following contents:

[0124] Generate data using Generative Adversarial Network (GAN):

[0125] In order to solve the problem of scarcity of extreme weather event samples, this embodiment uses a generative adversarial network (GAN) to generate virtual extreme weather;

[0126] Sample event:

[0127] G(z)=σ(W g ·z+b g );

[0128] Where z represents the input noise loudness, which follows a standard normal distribution. W g represents the weight matrix of the generator network, bg Represents the bias vector of the generator network, and σ represents the activation function, including ReLU and Sigmoid.

[0129] The discriminator (D) is used to evaluate the authenticity of the generated samples:

[0130] D(x)=σ(W d ·x+b d );

[0131] in:

[0132] x: input sample (real sample or generated sample); W d : The weight matrix of the discriminator; b d : The bias orientation of the discriminator.

[0133] The optimization goal of the GAN model is:

[0134]

[0135] in:

[0136] G is the generator r ):

[0137] The generator is a part of GAN, which is responsible for generating fake samples G(z) from random noise z. Its goal is to generate samples that are consistent with the real data distribution p data Samples that are close enough to deceive the discriminator D;

[0138] D is the discriminator:

[0139] The discriminator is another part of GAN, which is responsible for distinguishing whether the overflow sample comes from the real data distribution p data The sample x is still a fake sample G(z) generated by the generator.

[0140] V(G, D) is the objective function:

[0141] V(G, D) is the optimization goal of GAN. The generator G and the discriminator D are optimized through adversarial learning. The goal of D is to maximize V(G, D), that is, to distinguish real samples from forged samples as accurately as possible. The goal of G is to minimize V(G, D), that is, the generated samples can deceive D and make D think that G(z) is real data.

[0142] p data (x) is the real data distribution:

[0143] p data (x) is the distribution of real samples in the data space, describing the characteristics of real data.

[0144] Represents the expectation of the real data. The discriminator D needs to make the predicted value D(x) of the real data as close to 1 as possible.

[0145] p z (z) is the noise distribution:

[0146] p z (z) is the distribution of the input noise of the generator, usually a standard normal distribution

[0147] Represents the expectation of generating samples. The generator G hopes that the generated fake sample G(z) can make the predicted value D(G(z)) of the discriminator D close to the real sample.

[0148]

[0149] Indicates that in the data distribution p data Or the noise distribution p z The expected value of the function.

[0150] logD(x) is a logarithmic function:

[0151] Used to measure the prediction credibility of the star discriminator D for the real data x;

[0152] When the value of D(x) is close to 1, logD(x) tends to 0, indicating that D can distinguish real samples well.

[0153] log(1-D(G(z))) is the logarithmic loss:

[0154] The credibility of the judgment of the star discriminator on the forged samples.

[0155] The closer D(G(z)) is to 0, the larger log(1-D(G(z))) is, which means that the generator has successfully deceived the discriminator.

[0156] Super-resolution reconstruction: To improve the quality of low-resolution data, this embodiment uses SRCNN (Super-Resolution Convolutional Neural Network) to perform super-resolution reconstruction on the data. The specific steps include:

[0157] Initial feature extraction:

[0158] f1(X) = ReLU(W1·X+b1);

[0159] Where W1 and b1 are the convolution kernel and bias respectively:

[0160] Non-linear mapping:

[0161] f2(f1(X))=ReLU(W2·f1(X)+b2);

[0162] Where W2 and b2 are weight and bias respectively:

[0163] Reconstruct high-resolution data:

[0164] Y=W3·f2(f1(X))+b3;

[0165] Where Y is the output high-resolution data, W3 and b3 are weight and bias respectively.

[0166] In this embodiment, feature extraction and dimensionality reduction are performed on the enhanced data, specifically including the following:

[0167] Feature Selection:

[0168] Extract key meteorological variables that are highly correlated with extreme meteorological events, including temperature, humidity, precipitation, wind speed, etc.;

[0169] Filter irrelevant or redundant variables to ensure the simplicity and effectiveness of input features.

[0170] Feature Dimensionality Reduction:

[0171] The principal component analysis (PCA) method is used to reduce the dimension of high-dimensional meteorological data:

[0172] Z = XW, W = argmaxTr(W T SW);

[0173] in:

[0174] in, represents the original feature matrix, n is the number of samples, and d is the feature dimension; It represents the feature matrix after dimensionality reduction, k is the number of principal components; represents the projection matrix, which maps high-dimensional features to low-dimensional space; S represents the covariance matrix, which is defined as Tr(·) represents the matrix trace operation.

[0175] The results of principal component analysis ensure that the features after dimensionality reduction significantly reduce the data dimension while maintaining the main information of the data, thereby alleviating the computational burden of model training.

[0176] In this embodiment, the enhanced and extracted data has the following characteristics:

[0177] Sample expansion:

[0178] The virtual samples generated by GAN increase the diversity of extreme event data and improve the generalization ability of the model in small sample scenarios.

[0179] Improved data quality:

[0180] Super-resolution reconstruction significantly improves the spatial and temporal resolution of the data, enabling the model to capture more detailed spatiotemporal features.

[0181] Feature Optimization:

[0182] Through feature dimensionality reduction, data redundancy is removed, model efficiency is improved, and key feature input is provided for subsequent modeling.

[0183] Through the implementation of this step, data enhancement alleviates the problem of scarcity of extreme event samples, feature extraction and dimensionality reduction optimize the quality of data input, and provide a high-quality data foundation for subsequent deep learning modeling.

[0184] Deep learning modeling: build a spatiotemporal feature extraction model to perform multi-task learning on meteorological spatiotemporal data;

[0185] In this embodiment, the basic architecture of the deep learning model is first constructed, which specifically includes the following contents:

[0186] Spatial feature extraction (CNN part):

[0187] Convolutional Neural Network (CNN) is used to extract the spatial features of meteorological data. Input data Where H is the spatial height, W is the width, and C is the number of channels (such as different meteorological variables).

[0188] The convolution operation formula is:

[0189]

[0190] in, K represents the value of position (i, j) in the output feature graph of the ,th layer of convolution; mn Represents the weight of the convolution kernel at position (m, n); Indicates the value of the position (i+m, j+n) of the input symbol graph; b (l) Represents the bias of the th layer of convolution; M and N represent the height and width of the convolution kernel respectively; σ represents the activation function (such as ReLU).

[0191] After each convolution layer, the data dimension is reduced by the maximum pooling operation while retaining the main spatial features. The formula is:

[0192]

[0193] Temporal feature extraction (LSTM part):

[0194] Based on the spatial features extracted by CNN, LSTM is further used to capture the temporal dynamic characteristics of meteorological data.

[0195] The LSTM unit update formula is as follows:

[0196] Input Gate:

[0197] i t =σ(W i ·[h t-1 , x t ]+b i );

[0198] Forget Gate:

[0199] f t =σ(W f ·[h t-1 , x t ]+b f );

[0200] Cell status update:

[0201] c t =f t ⊙c t-1 +i t ⊙tanh(W c ·[h t-1 , x t ]+b c );

[0202] Output Gate:

[0203] o t =σ(W o ·[h t-1 , x t ]+b o );

[0204] Hidden state:

[0205] h t =o t ⊙tanh(c t );

[0206] in:

[0207] h t , c t : The hidden state and cell state at the current moment; W i , W f , W c ,W o : weight matrix; b i , b f , b c ,b o : bias term; σ: activation function; ⊙: element-wise multiplication operation.

[0208] In this embodiment, a multi-task learning framework is designed to achieve joint prediction, which specifically includes the following contents

[0209] Classification task (event probability prediction): output the probability of an extreme event occurring P(y=1|x);

[0210] P(y=1|x)=σ(W cls ·h+b cls );

[0211] in:

[0212] W cls , b cls : weights and biases for the classification task; h: final hidden state of the LSTM.

[0213] Regression task (event intensity prediction):

[0214] Output the intensity forecast of extreme events

[0215]

[0216] in:

[0217] W reg b reg : Weights and biases for regression tasks.

[0218] Loss function design:

[0219] Combining classification and regression tasks, define the total loss function:

[0220] L=αL cls +βL reg ;

[0221] in:

[0222] Cross entropy loss for classification tasks;

[0223] Mean squared error for regression tasks;

[0224] α, β: weight coefficients used to adjust the relative importance of the two tasks.

[0225] In this embodiment, the specific process of training and prediction is as follows:

[0226] Model training:

[0227] Use typical meteorological year data for model training;

[0228] The optimization goal is to minimize the total loss L, and the optimizer uses the Adam algorithm.

[0229] Model predictions:

[0230] Given future meteorological input data, the model outputs both the probability of an event occurring and its intensity forecast.

[0231] Through the implementation of the above steps, the deep learning model can effectively capture the complex spatiotemporal characteristics of meteorological data, and achieve accurate prediction of extreme meteorological events, providing reliable prediction results for subsequent steps.

[0232] Integration of physical constraints: Introducing physical conservation rules to correct model prediction results;

[0233] In this embodiment, physical consistency is achieved by introducing energy conservation constraints, which specifically include the following:

[0234] Energy conservation constraint definition:

[0235] The energy conservation relationship of the meteorological system can be expressed as:

[0236]

[0237] in:

[0238] q i : heat flux at the i-th grid point; Q i : heat source of the ith grid point; Divergence of heat flux; The time rate of change of the heat source.

[0239] Energy conservation constraint loss function:

[0240] Embed the above equation into the model training process and define the energy conservation constraint loss:

[0241]

[0242] in:

[0243] N: total number of grid points; |·|: absolute value symbol, used to calculate constraint deviation.

[0244] Constraint modification process:

[0245] Based on the meteorological variables (such as temperature, humidity, etc.) output by the model, the energy conservation loss is calculated.

[0246] The prediction results are corrected by optimizing the physical constraints obtained by the extraction and matching them with the output of the deep learning model.

[0247] In this embodiment, the wind speed field prediction is further corrected by introducing mass conservation constraints, which specifically include the following contents:

[0248] Mass conservation constraint definition:

[0249] In the meteorological system, the mass conservation relation of the wind speed field can be expressed as:

[0250]

[0251] in:

[0252] v i =(u i , v i , w i ): 3D wind speed vector of the ith grid point; Divergence of the wind speed field.

[0253] Mass conservation constraint loss function:

[0254] Add the mass conservation equation as a loss term to the model optimization objective:

[0255]

[0256] in:

[0257] L mass : mass conservation loss; N: total number of grid points.

[0258] Post-processing corrections:

[0259] If the wind speed field output by the model does not satisfy the mass conservation, it can be corrected by post-processing methods:

[0260]

[0261] in:

[0262] v: wind speed field originally predicted by the model; φ: correction term; v*: corrected wind speed field.

[0263] In this embodiment, physical constraints are integrated into the deep learning model optimization objectives, specifically including the following:

[0264] The total loss function is defined as:

[0265] Combining model prediction loss (such as classification and regression task loss) with physical constraint loss, the total loss function is defined as:

[0266] L total =αL cls +βL reg +γ1L energy +γ2L mass ;

[0267] in:

[0268] L cls : classification task loss; Lreg : regression task loss; L energy : Energy conservation loss; L mass : mass conservation loss; a, β, γ1, γ2: weight coefficients used to adjust the relative importance of different loss terms.

[0269] Optimization and training:

[0270] During model training, by minimizing the total loss L total Simultaneously optimize prediction accuracy and physical consistency.

[0271] In this embodiment, the final effect of physical constraint integration is:

[0272] By introducing energy conservation and mass conservation constraints, the model prediction results are more consistent with the basic laws of meteorology;

[0273] The corrected wind speed field, temperature field and other variables are more scientific and reasonable, providing reliable support for the uncertainty quantification and application of subsequent predictions.

[0274] Through the implementation of the above steps, the present invention enhances physical consistency based on deep learning modeling, and significantly improves the scientificity and credibility of the model prediction results.

[0275] Uncertainty quantification: quantify the uncertainty of prediction results through sampling methods or multi-model ensemble methods;

[0276] In this embodiment, the MC-Dropout method is first used to quantify the uncertainty of the prediction results, which specifically includes the following contents:

[0277] Sampling multiple times generates a predictive distribution:

[0278] In a trained deep learning model, activating the Dropout layer keeps the property of randomly dropping neurons even during the inference phase:

[0279] For the same input x i , perform M forward propagations and generate M prediction values:

[0280]

[0281] in:

[0282] The predicted value of the mth sampling; f(x i θ (m) ): Network model for input x i The mth sampling output; θ (m) : Model parameters for the mth sampling.

[0283] Compute the forecast mean and variance:

[0284] By taking the mean of all sampled results, we can get the expected value of the predicted result:

[0285]

[0286] in For sample x i The final predicted value of

[0287] By calculating the variance of the sampling results, we can get the quantitative index of uncertainty of the prediction results:

[0288]

[0289] Generate confidence intervals:

[0290] Assuming that the sampling results follow a normal distribution, a confidence interval is constructed based on the mean and variance:

[0291]

[0292] in:

[0293] z a / 2 Quantile of the standard normal distribution (e.g., 95% confidence interval corresponds to z = 1.96); CI α : Confidence interval range.

[0294] In this embodiment, the risk level assessment is performed in combination with the uncertainty quantification result, which specifically includes the following contents:

[0295] Risk level classification criteria:

[0296] According to the probability of occurrence and forecast intensity of extreme events, combined with confidence interval information, they are divided into low risk, medium risk and high risk:

[0297] Low risk: P(y=1|x)<0.3 and

[0298] Medium risk: 0.3≤P(y=1|x)<0.7 and

[0299] High risk: P(y=1|x)≥0.7 and

[0300] in,

[0301] Predict upper bounds;

[0302] Predict lower bounds;

[0303] T1, T2: Risk intensity thresholds.

[0304] Adjust the reliability of risk level:

[0305] If the confidence interval spans Larger, reducing the reliability rating of risk assessment results;

[0306] If the confidence interval is small, the credibility of the risk assessment is improved.

[0307] In this embodiment, decision support information is generated based on the uncertainty quantification result, which specifically includes the following contents:

[0308] Risk distribution visualization:

[0309] Draw a regional risk level distribution map to show the risk classification of different regions;

[0310] Output a time dynamic prediction chart to show the evolution trend of risk level over time.

[0311] Decision support suggestions:

[0312] In high-risk areas, provide specific recommendations for preventive measures;

[0313] For medium-risk areas, the possible time of occurrence of events and the potential scope of impact are indicated.

[0314] Through the implementation of this step, not only the confidence interval of the prediction results is generated, but also a comprehensive risk level assessment is provided in combination with the occurrence probability and intensity information, providing reliable technical support for decision-making scenarios such as disaster prevention and mitigation, energy scheduling and agricultural management in practical applications.

[0315] Result deployment and application: Provide data support based on prediction results to guide optimization and decision-making in specific scenarios.

[0316] In this embodiment, the prediction results are first visualized, which specifically includes the following contents:

[0317] Risk level distribution map:

[0318] Based on the risk level assessment results, a spatial distribution map is generated to visualize the risk classification of different areas;

[0319] The data input includes the predicted risk level, confidence interval and time distribution, and the output is a two-dimensional or three-dimensional risk distribution map:

[0320]

[0321] in:

[0322] R(x, y): regional risk level, divided into three levels: low (1), medium (2), and high (3); P(y=1|x): probability of event occurrence; Upper and lower limits of predicted values; T1, T2: risk level thresholds.

[0323] Dynamic time evolution diagram:

[0324] Using time as the dimension, chart the dynamic changes in the possible time range and intensity of extreme events.

[0325]

[0326] in:

[0327] D(t): dynamic prediction result of time; Prediction strength at time t; CI α (t): confidence interval.

[0328] Uncertainty distribution map:

[0329] Demonstrate the uncertainty range of the results by showing upper and lower bounds of the predicted values ​​and the distribution of the variance.

[0330] In this embodiment, decision support information is provided in combination with the prediction results, specifically including the following:

[0331] Risk warning report:

[0332] Automatically generate early warning reports based on high-risk areas and high-risk time periods, including the following:

[0333] the likely timing of extreme events;

[0334] The predicted intensity of the event and the scope of its impact;

[0335] Uncertainty assessment information (such as confidence intervals and risk levels).

[0336] Suggestions for optimizing resource allocation:

[0337] Provide specific optimization solutions for different risk levels:

[0338] High-risk areas: It is recommended to deploy disaster prevention resources in advance and strengthen protective measures in key areas;

[0339] Medium-risk areas: It is recommended to maintain monitoring and make emergency preparations;

[0340] Low-risk areas: Only routine monitoring recommendations are provided.

[0341] In this embodiment, the prediction results are deployed to the application system, which specifically includes the following contents:

[0342] Online service platform:

[0343] The prediction model is deployed as an online service, and users can obtain prediction results by uploading input data (such as real-time meteorological data).

[0344] The service output includes the probability of event occurrence, intensity prediction value and its confidence interval.

[0345] API interface:

[0346] Provide standardized interfaces to embed prediction results into other application systems (such as disaster prevention systems and agricultural decision-making platforms).

[0347] Local deployment:

[0348] For environments without Internet access or offline, the models and application systems are packaged as locally running tools, and users can directly input data and obtain results.

[0349] In this embodiment, the result deployment and application achieve the following goals:

[0350] Through visualization and reporting, the prediction results can be presented to users more intuitively;

[0351] Combined with risk assessment, generate practical decision support recommendations;

[0352] Provides a variety of system deployment methods to meet the needs of different scenarios.

[0353] Through the implementation of this step, the prediction results of extreme meteorological events are converted into effective supporting information in actual application scenarios, which improves the practicality and usability of the prediction technology.

[0354] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The extreme weather forecasting method based on a typical meteorological year is characterized by: The following steps are involved: Typical meteorological year data construction: Acquire meteorological data, including ground observation data, reanalysis data and satellite remote sensing data, clean and fuse the data, and construct a typical meteorological year; Data enhancement and feature extraction: Expand extreme event data samples through data generation and enhancement methods, perform feature engineering and dimensionality reduction; Deep learning modeling: build a spatiotemporal feature extraction model to perform multi-task learning on meteorological spatiotemporal data; Integration of physical constraints: Introducing physical conservation rules to correct model prediction results; Uncertainty quantification: quantify the uncertainty of prediction results through sampling methods or multi-model ensemble methods; Result deployment and application: Provide data support based on prediction results to guide optimization and decision-making in specific scenarios.

2. The extreme weather forecasting method based on a typical weather year according to claim 1, characterized in that: The typical meteorological year data construction includes the following contents: The missing values ​​in the meteorological time series are interpolated using the local weighted regression method, and the formula is: in, represents the interpolated value at time point t, y i Indicates time point t i The existing observations, n represents the number of time points used for interpolation, w i represents the time weight, and λ represents the smoothing parameter that controls the weight decay rate.

3. The extreme weather forecasting method based on a typical weather year according to claim 1, characterized in that: The data enhancement includes using a generative adversarial network to expand extreme weather event data samples, where the generator generation formula is: G(z)=σ(W g ·z+b g ); Where z represents the input noise loudness, which follows a standard normal distribution. W g represents the weight matrix of the generator network, b g Represents the bias vector of the generator network, and σ represents the activation function, including ReLU and Sigmoid.

4. The extreme weather forecasting method based on a typical weather year according to claim 1, characterized in that: The feature dimension reduction adopts the principal component analysis method, and the dimension reduction formula is: Z=ZW,W=argmaxTr(W T SW); in, represents the original feature matrix, n is the number of samples, and d is the feature dimension; It represents the feature matrix after dimensionality reduction, k is the number of principal components; represents the projection matrix, which maps high-dimensional features to low-dimensional space; S represents the covariance matrix, which is defined as Tr(·) represents the matrix trace operation.

5. The extreme weather forecasting method based on a typical weather year according to claim 1, characterized in that: The deep learning modeling includes constructing a spatiotemporal feature extraction model combining a convolutional neural network and a long short-term memory network, wherein the output formula of the convolutional layer is: in, K represents the value of position (i, j) in the l-th layer convolution output feature map; mn Represents the weight of the convolution kernel at position (m, n); Indicates the value of the position (i+m, j+n) of the input symbol graph; b (l) represents the bias of the lth layer of convolution; M and N represent the height and width of the convolution kernel respectively; σ represents the activation function.

6. The extreme weather forecasting method based on a typical weather year according to claim 1, characterized in that: The time series processing of the long short-term memory network includes the following formula: Cell status update: c t =f t ⊙c t-1 +i t ⊙tanh(W c ·[h t-1 ,x t ]+b c ); Hide status updates: h t =o t ⊙tanh(c t ); Among them, c t Indicates cell status; h t represents the hidden state; f t 、i t , o t Respectively represent the forget gate, input gate, and output gate; W c represents the cell state weight matrix; x t Represents the input features at time point t; ⊙ represents an element-by-element multiplication operation.

7. The extreme weather forecasting method based on a typical weather year according to claim 1, characterized in that: The physical constraint integration includes: Introducing the energy conservation constraint, the formula is: Among them, L energy represents the energy conservation loss, represents the heat flux divergence at the ith grid point, Q i represents the heat source at the ith grid point, Represents the time rate of change of the heat source.

8. The extreme weather forecasting method based on a typical weather year according to claim 1, characterized in that: The uncertainty quantification adopts the MC-Dropout sampling method, and its formula is: in, Represents sample x i The expected prediction result of f(x i θ (m) ) represents the prediction result of the model under the mth random sampling, θ (m) represents the network parameters of the mth sampling, and M represents the number of sampling times.

9. The extreme weather forecasting method based on a typical weather year according to claim 1, characterized in that: The deployment and application of the results include: Generate a risk level assessment of the meteorological event based on the forecast results, the assessment including: According to the probability and intensity of extreme meteorological events, the events are divided into low risk, medium risk and high risk levels; Combined with the prediction confidence interval, the reliability evaluation of the risk level is adjusted; Provide decision support information based on risk level, including: Provide risk warning tips to clarify the time, location and impact range of possible events; Provide optimization solutions for specific scenarios to guide scenario optimization and resource scheduling; The prediction results are output in a visual way, including risk level distribution map, time dynamic prediction map and uncertainty distribution map of prediction results.

10. An extreme weather forecasting system based on a typical weather year, applied to the method according to any one of claims 1 to 9, characterized in that: The system comprises: Data acquisition module, used to obtain meteorological observation data, reanalysis data and satellite remote sensing data; Data processing module for data cleaning, enhancement and feature extraction; Deep learning module for spatiotemporal feature extraction and multi-task learning; Physical constraint module, used to correct model output; Uncertainty quantification module, used to quantify the uncertainty of prediction results; Application module for generating forecast reports and supporting scenario optimization decisions.

Images