Data enhancement method based on time sequence generative adversarial network

Abstract

The invention provides a data enhancement method based on a time sequence generative adversarial network. The method comprises the steps of obtaining a real rainfall sequence based on historical observation data; determining climate scene parameters and extreme rainfall identification conditions; when the Time GAN model is trained, a real rainfall sequence and a false rainfall sequence output by a generator are input into a discriminator, and parameters of the discriminator are optimized; inputting random noise, climate scene parameters and extreme rainfall identification conditions into a generator, and performing back propagation optimization on generator parameters by using a generator loss function; a physical equation regularization item is introduced into the loss function of the generator, the physical equation regularization item encodes physical rules of water vapor mass conservation and energy conservation to the loss function, and data is forcibly generated to meet physical constraints; and finally, adopting the trained Time GAN model to generate a required number of new rainfall samples. The method can solve the problems of misalignment of physical rules, insufficient generation of extreme rainfall and disjunction of climate response.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and urban meteorological warning technology, and particularly relates to a data enhancement method based on a time sequence generative adversarial network. BACKGROUND

[0002] In the task of urban extreme weather precipitation prediction, the data imbalance problem is one of the key factors affecting the performance of the model. Taking monthly observation data as an example, if the precipitation day accounts for about 5 days in a month, the remaining 25 days are non-precipitation days. Because the precipitation event accounts for a low proportion in the overall sample, the model is prone to overfitting to the "non-precipitation" label during the training process, thereby weakening its recognition ability for "precipitation" samples, especially the recognition ability for short-time heavy precipitation events. This sample imbalance problem not only reduces the sensitivity of the model, but also to some extent causes prediction bias, so that the model tends to predict the "non-precipitation" state, which seriously affects the accuracy of the warning system. Therefore, it is of great significance to enhance the data and adjust the proportion of precipitation day data and non-precipitation day data.

[0003] Traditional data enhancement methods, such as simple data duplication, translation, rotation, etc., cannot effectively generate high-quality data that conforms to the actual situation for precipitation data with complex spatio-temporal characteristics and physical laws. Although Time GAN, as a generative adversarial network for time series data generation, can generate data with time correlation to some extent, it has obvious technical defects when applied to precipitation data.

[0004] Time GAN is essentially a statistical learning model, and the generated data may violate atmospheric physical constraints (such as energy conservation, humidity continuity), resulting in abnormal phenomena such as "no cloud precipitation" and "negative humidity". At the same time, the long-tail distribution of precipitation data (such as a hundred-year rainstorm) is difficult to be captured by GAN, and the generated results often deviate to common precipitation intensity, weakening the characteristics of extreme events, and such events are very important. In addition, Time GAN also fails to fully consider the influence of climate scenarios on precipitation patterns. Therefore, there is an urgent need to develop a precipitation data enhancement technology that can effectively solve the above problems. SUMMARY

[0005] Therefore, the present application provides a data enhancement method based on an improved time sequence generative adversarial network (Time GAN), which solves the three major problems of physical law inaccuracy, insufficient extreme precipitation generation, and climate response disconnection in traditional data enhancement methods by embedding atmospheric physical equation constraints, constructing an extreme event triggering mechanism, and climate scenario parameters, and is suitable for the fields of meteorological warning, hydrological simulation, and climate risk assessment.

[0006] In order to solve the above technical problems, the present application is implemented as follows.

[0007] A data enhancement method based on a time sequence generative adversarial network, comprising:

[0008] Step 1: Based on historical observation data, obtain a real precipitation sequence, and determine corresponding climate scenario parameters and extreme precipitation identification conditions;

[0009] Step 2: Train a Time GAN model:

[0010] Input the real precipitation sequence and the false precipitation sequence output by the generator into the discriminator, and use the discriminator loss function to perform back propagation optimization on the discriminator parameters; the discriminator loss function is the weighted sum of the real precipitation sequence term and the false precipitation sequence term;

[0011] Input random noise, climate scenario parameters and extreme precipitation identification conditions into the generator, and use the generator loss function to perform back propagation optimization on the generator parameters; the generator loss function introduces a physical equation regularization term, which encodes the physical laws of water vapor mass conservation and energy conservation into the loss function, forcing the generated data to satisfy physical constraints;

[0012] Step 3: Generate a required number of new precipitation samples using the trained Time GAN model.

[0013] Preferably, the method further comprises: in the discriminator, when extreme precipitation is identified according to the extreme precipitation identification conditions, increase the weight of the false precipitation sequence term to make the model focus more on generating high-quality extreme precipitation data;

[0014] In the generator, specify part of the parameters as extreme precipitation parameters; when extreme precipitation is identified according to the extreme precipitation identification conditions, enable the extreme precipitation parameters to enhance the modeling capability of strong convective processes; if it is not extreme precipitation at present, the extreme precipitation parameters are not enabled.

[0015] Preferably, the extreme precipitation identification conditions include convective available potential energy (CAPE); if the convective available potential energy (CAPE) is greater than a set convective available potential energy threshold, it is identified as extreme precipitation.

[0016] Preferably, the extreme precipitation identification conditions are:

[0017] C ext = [CAPE, |Δv 850-200 |, RH BL ] T

[0018] Wherein, C ext is a vector composed of extreme precipitation identification conditions; CAPE is the convective available potential energy, |Δv 850-200| is the vertical wind shear modulus, RH BL is the relative humidity of the boundary layer.

[0019] Preferably, the discriminator loss function is:

[0020]

[0021] wherein, is the discriminator loss; E[] is the expectation operation; D() is the discriminator function, outputting a probability value between 0 and 1; X real is the real precipitation sequence; G(Z) is the precipitation sequence generated by the generator, Z is random noise, G is the generator network; γ is the weighting coefficient; when it is determined to be extreme precipitation, γ is selected as γ1, otherwise γ is selected as γ2, γ1> γ2.

[0022] Preferably, the determination of the climate scenario parameter is:

[0023] The selection of large-scale climate driving signals includes global temperature rise ΔT glob , regional sea surface temperature anomaly ΔSST reg , and Arctic sea ice change ΔIce arc .

[0024] The three large-scale climate driving signals are mapped into a vector of fixed dimension climate scenario parameters by the function f to respond to different climate scenarios.

[0025] Preferably, the generator loss function introduces a physical equation regularization term L phy :

[0026]

[0027] The first term in the formula corresponds to the conservation of water vapor mass, ensuring that precipitation generation matches water vapor transport; wherein, P gen is the generated precipitation, ε is the condensation efficiency coefficient, q is the specific humidity, v is the wind speed vector, ΔS is the loss term, denotes vertical integration first and then time integration; t0-Δt, t0 are the start and end of the time integration interval; z s , z c are the lower and upper boundaries of the vertical integration interval, which are determined according to the spatial range involved in the precipitation, denotes the divergence calculation.

[0028] The second term in the formula corresponds to the conservation of energy, relating precipitation to atmospheric energy changes; wherein, θ e is the equivalent potential temperature, Q latent is the latent heat flux; k1 is a conditional approximate proportionality coefficient obtained from the basic equations of thermodynamics; α and β are weight coefficients, respectively.

[0029] Preferably, the weight coefficients a and β adopt dynamic weight coefficients:

[0030] a = 0.2 + 0.6 * min(1, P obs / RES obs )

[0031] β = 0.3 + 0.4 * min(1, P obs / RES obs )

[0032] wherein P obs represents the observed precipitation, RES obs is the rainstorm daily precipitation threshold; min(,) represents taking the minimum value; when P obs ≥ RES obs , a rainstorm occurs, and the physical constraint is strengthened by increasing a and β; when it is not a rainstorm or no rain, the physical constraint is weakened by reducing a and β; when it rains but does not reach the rainstorm level, the greater the rainfall, the stronger the physical constraint.

[0033] Preferably, when a rainstorm occurs, a = 0.8 and β = 0.7.

[0034] Preferably, in step 3, the generator is used to obtain generated samples, the generated samples are merged with the original samples, and the balanced data set of the required proportion is constructed.

[0035] Beneficial effects:

[0036] (1) The present application performs physical constraint on the generator of the Time GAN model, encodes the atmospheric mass conservation, energy conservation and other physical laws into the loss function, forces the generated data to meet the physical constraint, ensures that the generated data conforms to the atmospheric physical law, and solves the problem of inaccurate physical law in the traditional data enhancement method.

[0037] (2) In a preferred embodiment, a dynamic weight mechanism is designed for the loss term corresponding to the atmospheric mass conservation and energy conservation in the generator: when a rainstorm occurs, the physical constraint is strengthened; when it is not a rainstorm or no rain, the constraint is weakened, and flexibility is retained.

[0038] (3) The present application designs an extreme event enhancement strategy: when extreme rainfall occurs, on the one hand, the extreme rainfall parameters in the generator are enabled to strengthen the modeling ability of the enhancement model for strong convective processes; on the other hand, the weight of the discriminator loss function is adjusted, the weight of the extreme sample is increased, the model focuses more on generating high-quality extreme precipitation data, and the problem of insufficient generation of traditional GAN extreme precipitation is solved.

[0039] (4)The application considers climate response, and a climate scenario parameter is added to the input features of the generator to associate climate scenarios, so that the generated data can respond to different climate scenarios, and the problem of disconnection of climate response in traditional data enhancement algorithms is solved. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 A schematic diagram of the data enhancement method based on the time sequence generative adversarial network.

[0041] Figure 2 A schematic diagram of the extreme precipitation enhancement strategy. DETAILED DESCRIPTION

[0042] The application provides a data enhancement method based on a time sequence generative adversarial network, and the design points are as follows:

[0043] (1) Physical constraint design for the generator of the Time GAN model: the scheme introduces a physical equation regularization term into the loss function of the generator, encodes atmospheric mass conservation, energy conservation and other physical laws into the loss function, forces the generated data to meet the physical constraints, ensures that the generated data conforms to the atmospheric physical laws, and solves the problem of inaccurate physical laws in traditional data enhancement methods. Further, a dynamic weight mechanism is designed: when a rainstorm occurs, the physical constraint is strengthened; when there is no rainstorm or rain, the constraint is weakened to retain flexibility.

[0044] (2) Design of the extreme event enhancement strategy: in view of the “long-tail distribution” characteristics of extreme precipitation, an extreme precipitation identification and threshold activation strategy is designed to strengthen the learning of the model on extreme samples. Further, the weight of the discriminator loss function is designed, the weight of the extreme samples is increased, the model focuses more on generating high-quality extreme precipitation data, and the problem of insufficient generation of extreme precipitation in the traditional GAN is solved.

[0045] (3) Consideration of climate response, the climate scenario parameter is added to the input features of the generator to associate climate scenarios, so that the generated data can respond to different climate scenarios (such as global warming and sea surface temperature anomaly), and the problem of disconnection of climate response in traditional data enhancement algorithms is solved.

[0046] The application will be described in detail below with reference to the drawings and examples.

[0047] Referring to Figure 1 The data enhancement method based on the time sequence generative adversarial network provided by the embodiment of the application includes the following steps:

[0048] Step 1: Acquisition and preprocessing of model input data.

[0049] Referring to Figure 1 The data input module and the preprocessing module.

[0050] The data input module obtains two types of data, observation data and CMIP6 climate scenario data.

[0051] Observation data: historical precipitation series, wind speed, humidity, PWV, etc. observed values recorded by actual weather stations and GNSS receiving stations, are the basis for the model to learn the "true precipitation rule". In particular, the data contains extreme precipitation precursor signals (CAPE (convective available potential energy), vertical wind shear, boundary layer humidity).

[0052] CMIP6 climate scenario: climate simulation data from the 6th phase of the international Coupled Model Intercomparison Project (CMIP6), containing large-scale climate signals such as global temperature rise, sea surface temperature anomaly under different emission scenarios (such as RCP), allowing the model to respond to large-scale climate scenarios.

[0053] The preprocessing module processes the observation data and the CMIP6 climate scenario to process the true precipitation series and the corresponding climate scenario parameters and extreme precipitation identification conditions.

[0054] The preprocessing module includes an event slicing module (observation data side) and a climate-weather cross-scale coupling module (CMIP6 side).

[0055] Among them, the event slicing module: taking "precipitation event" as the core, the data is intercepted according to the window [T0-Ta, T0+Tb] (Ta days before T0 and Tb days after T0 of precipitation day), the time evolution process of the precipitation event (such as water vapor accumulation before precipitation, energy release after precipitation) is retained, and the model learns the complete precipitation life cycle. The event slicing module obtains the true precipitation series as a sample.

[0056] The climate-weather cross-scale coupling module: used to determine the climate scenario parameters. This module encodes large-scale climate factors (such as global temperature rise, sea surface temperature anomaly) into a low-dimensional vector to extract the key features of the climate scenario, so that the generated data can respond to different climate scenarios (such as global warming, sea surface temperature anomaly).

[0057] In a preferred embodiment, the determination method of the climate scenario parameter is:

[0058] Selecting large-scale climate driving signals including global temperature rise ΔT glob (℃), regional sea surface temperature anomaly ΔSST reg (℃) and Arctic sea ice change ΔIce arc (%) as input; map the three large-scale climate driving signals into a 10-dimensional climate scenario parameter vector through a function f to encode the large-scale climate driving signals to respond to different climate scenarios.

[0059] The climate scenario vector C rcp is generated as:

[0060] C rcp = f (ΔT glob , ΔSST reg , ΔIce arc ) ∈ R 10

[0061] The climate scenario vector C rcp is input to the Time GAN as a condition, and the "downscaling" conversion from large-scale climate signals to mesoscale precipitation processes is realized through the model structure, ensuring that the generated precipitation data can reflect climate scenario changes.

[0062] In practice, other scenario vectors can also be selected for climate-weather cross-scale coupling to adapt to different climate model prediction needs.

[0063] Step 2: Train the Time GAN model.

[0064] The Time GAN model is used in the present application as a generation model for new precipitation samples. The model includes a discriminator (D) and a generator (G). The two are trained in opposition, with the generator (G) generating precipitation sequences and the discriminator (D) distinguishing between "real / generated data", improving the authenticity of the generated data through adversarial games. In the present embodiment, the generator (G) uses BiLSTM+TCN to capture long / short-term temporal dependencies of precipitation (such as daily variation and seasonal evolution of precipitation), achieving temporal modeling.

[0065] The training steps of the Time GAN model include:

[0066] Step a1: Initialize the model and set the initial parameters of the generator and discriminator.

[0067] Step a2: Train the discriminator: input real precipitation sequences and false precipitation sequences generated by the generator, and let the discriminator learn to distinguish between the two, using the discriminator loss function to optimize the discriminator parameters through backpropagation. The discriminator loss function is a weighted sum of the real precipitation sequence term and the false precipitation sequence term; in particular, in order to enhance extreme precipitation, an extreme precipitation identification condition also needs to be determined. When extreme precipitation is identified according to the extreme precipitation identification condition, a higher weight (e.g. γ = 3.0) is given to the extreme precipitation sample in the discriminator loss function, strengthening the modeling ability of extreme events and allowing the model to focus more on generating high-quality extreme precipitation data.

[0068] Step a3: Train the generator: input random noise, extreme precipitation identification conditions, and climate scenario parameters to generate false precipitation sequences, and jointly optimize the generator parameters through adversarial loss, supervised error loss, and physical constraint loss. In particular, when the CAPE in the input extreme precipitation identification condition is greater than a threshold, the extreme precipitation parameters in the generator are activated (normally not activated), thereby strengthening the learning of extreme samples.

[0069] Step a4: Loop iteration: repeat training discriminators and generators until the model converges, i.e. the generator can generate precipitation sequences that are difficult for the discriminator to distinguish between true and false and conform to physical laws.

[0070] In the above step a3 of the generator training process, the core is to increase the physical constraints and extreme precipitation enhancement strategy.

[0071] For the physical constraint problem: the invention embeds atmospheric physical equations (such as mass conservation, energy conservation) to make the generated precipitation sequence conform to meteorological laws (avoiding physical contradictions such as "no cloud precipitation").

[0072] Embedding atmospheric physical equations is achieved by introducing a physical equation regularization term in the generator loss function.

[0073] ① Physical equation regularization term L phy Specifically:

[0074]

[0075] The first term in the above formula corresponds to the water vapor mass conservation, ensuring that the precipitation generation matches the water vapor transport; where P gen is the generated precipitation, ε is the condensation efficiency coefficient, q is the specific humidity, v is the wind speed vector, ΔS is the loss term, indicates vertical integration first and then time integration; t0-Δt, t0 are the start and end of the time integration interval, which can be determined according to the research object and the prediction frequency (such as the start and end of a precipitation event), z s , z c are the lower and upper boundaries of the vertical integration interval, which can be determined according to the spatial range mainly involved in precipitation (such as the ground and the upper layer of the troposphere), indicates the divergence calculation.

[0076] The second term in the above formula corresponds to the energy conservation, which relates precipitation to atmospheric energy changes; where θ e is the equivalent potential temperature, Q latent is the latent heat flux, k1 is a conditional approximate proportionality coefficient, which can be obtained from the basic equations of thermodynamics. α and β are weight coefficients.

[0077] The above physical equation regularization term can be simplified as long as it can express the constraints of the physical model.

[0078] Based on the physical equation regularization term, a total loss function can be constructed including an adversarial loss term, a physical equation regularization term, and a supervised error loss term. A strength adjustment factor can be added before the physical equation regularization term.

[0079] ② Dynamic weight mechanism

[0080] For the weight coefficients a and β in the regularization term of the above physical equation, a dynamic weight mechanism is designed, which can dynamically adjust the proportion of the physical constraint term according to the precipitation intensity. The formula is expressed as:

[0081] a = 0.2 + 0.6 x min (1, P obs / RES obs )

[0082] β = 0.3 + 0.4 x min (1, P obs / RES obs )

[0083] Where P obs represents the observed precipitation, RES obs is the threshold of daily heavy rain precipitation; min(,) represents the minimum value; when P obs ≥ RES obs , heavy rain occurs, and the physical constraint is strengthened by increasing a and β, preferably a = 0.8, β = 0.7; when it is not raining, the physical constraint is weakened by reducing a and β, and the flexibility is retained, preferably a = 0.2, β = 0.3; when it is raining but not reaching the heavy rain level, the greater the rainfall, the stronger the physical constraint.

[0084] The dynamic weight mechanism can also have other expressions, but the core idea is to strengthen the physical constraint in heavy rain and weaken the physical constraint in weak rain or no rain.

[0085] For the extreme precipitation enhancement problem:

[0086] For the "long tail distribution" problem of extreme precipitation (such as ≥ 50 mm / day), an extreme precipitation identification condition judgment and an extreme precipitation special network parameter are added.

[0087] When extreme precipitation is identified according to the extreme precipitation identification condition, two treatments are performed; as shown in Figure 2 , one is the adjustment of the generator: part of the parameters in the generator is specified as an extreme precipitation parameter, for example, part of the dimensions of the generator hidden layer is specified as an extreme precipitation parameter; when extreme precipitation is identified according to the extreme precipitation identification condition, the extreme precipitation parameter is enabled to enhance the modeling ability of the model for strong convective processes; if it is not extreme precipitation, the extreme precipitation parameter is not enabled. Two is the adjustment of the discriminator: increase the weight of extreme samples, so that the model focuses more on generating high-quality extreme precipitation data (the specific scheme is described in the specific introduction of the discriminator).

[0088] In this embodiment, the extreme precipitation identification condition includes: CAPE (convective available potential energy), vertical wind shear, and boundary layer relative humidity. The extreme precipitation identification condition vector C ext is constructed as:

[0089] C ext =[CAPE,|Δv 850-200 |,RH BL ] T

[0090] Among them, CAPE (Convective Available Potential Energy) can intuitively and quantitatively reflect the atmospheric convection potential. High CAPE easily triggers heavy precipitation; therefore, this component is used to determine whether the extreme precipitation parameter in the generator is activated. |Δv 850-200 |(850-200hPa vertical wind shear modulus) can affect storm organization; strong shear tends to sustain extreme precipitation. RH BL (Boundary layer relative humidity) reflects the supply of water vapor; high humidity provides the water vapor basis for extreme precipitation. |Δv 850-200 |and RH BL This is used to better simulate extreme precipitation events by interacting with other physical quantities in the generator after activating the extreme precipitation parameters in the generator.

[0091] The C ext As part of the generator's input features, C ext CAPE is also used as a criterion for identifying extreme precipitation. Threshold activation of some parameters in the generator is implemented based on CAPE. When CAPE ≥ RES1 (convective effective potential energy threshold), the extreme precipitation parameters in the generator are activated (these parameters are not activated during ordinary precipitation), which enhances the generation capability of extreme precipitation and makes the model more likely to produce extreme events that are consistent with reality.

[0092] In the training process of the discriminator in step a2 above, the core lies in adding an extreme precipitation enhancement strategy.

[0093] As mentioned earlier, when extreme precipitation is identified based on the extreme precipitation identification criteria, the weight of extreme samples is increased to enhance the discriminator's ability to judge extreme precipitation, allowing the model to focus more on generating high-quality extreme precipitation data.

[0094] In this embodiment, the loss function of the discriminator is:

[0095]

[0096] in, X is the discriminator loss, which measures the discriminator's ability to distinguish between "real samples" and "generated samples"; E[·] is the expectation operation; D(·) is the discriminator function, which outputs a probability value of 0-1, where a value closer to 1 indicates a stronger likelihood of being classified as a "real sample" and a value closer to 0 indicates a stronger likelihood of being classified as a "generated sample"; X realis the real precipitation sequence (from observation data); G(Z) is the precipitation sample generated by the generator (Z is the noise vector, and G is the generator network); and γ is the weighting coefficient.

[0097] When the sample corresponds to the extreme precipitation condition identified according to the extreme precipitation identification condition, the extreme sample γ = 3.0; otherwise, the ordinary sample γ = 1.0. That is, by increasing the weight of the extreme sample, the model focuses more on generating high-quality extreme precipitation data, solving the problem of insufficient generation of traditional GAN extreme precipitation.

[0098] In practice, the discriminator loss weighting can have other ways, but the core idea is to increase the weight of the extreme sample, so that the model focuses more on generating high-quality extreme precipitation data, solving the problem of insufficient generation of traditional GAN extreme precipitation.

[0099] Step 3: Generate the required number of new precipitation samples using the trained Time GAN model.

[0100] The data generation process in this embodiment includes:

[0101] Input conditions: Determine the extreme precipitation identification conditions (CAPE, vertical wind shear, boundary layer humidity) and climate scenario parameters (such as global temperature rise, regional sea surface temperature anomaly).

[0102] Construct a vector: Integrate the above two conditions into an input vector C.

[0103] Generate noise: Generate a random noise vector Z.

[0104] Generate precipitation: Input the input vector C and the random noise vector Z into the trained generator, and output the precipitation sequence.

[0105] Result output: Obtain a precipitation sequence that meets the physical laws and contains extreme event characteristics, which can be used for meteorological warning, hydrological simulation, etc.

[0106] In practice, the generated samples and the original samples can also be combined to construct a balanced data set with the required proportion.

[0107] Thus, the process ends.

[0108] The embodiment of the application solves the three major problems of physical law misalignment, insufficient generation of extreme precipitation, and climate response disconnection in traditional data enhancement methods by embedding atmospheric physical equation constraints, constructing extreme event triggering mechanisms, and climate scenario parameters, and is suitable for meteorological warning, hydrological simulation, and climate risk assessment fields.

[0109] The above specific embodiments only describe the design principles of the present application, and the shapes and names of the components in the description can be different and are not limited. Therefore, those skilled in the art of the present application can modify or equivalently replace the technical solutions described in the foregoing embodiments; and these modifications and replacements do not deviate from the purpose and technical solutions of the present application, and should all belong to the protection scope of the present application.

Claims

1. A data augmentation method based on a time sequence generative adversarial network, characterized in that, The method comprises the following steps: Step 1: Based on historical observation data, obtain a real precipitation sequence, and determine corresponding climate scenario parameters and extreme precipitation identification conditions; Step 2: Train a Time GAN model: Input the real precipitation sequence and the false precipitation sequence generated by the generator into the discriminator, and optimize the parameters of the discriminator by back propagation using the discriminator loss function; the discriminator loss function is the weighted sum of the real precipitation sequence term and the false precipitation sequence term; Input random noise, climate scenario parameters and extreme precipitation identification conditions into the generator, and optimize the parameters of the generator by back propagation using the generator loss function; the generator loss function introduces a physical equation regularization term, which encodes the physical laws of water mass conservation and energy conservation into the loss function, forcing the generated data to satisfy the physical constraints; Step 3: Generate a required number of new precipitation samples using the trained Time GAN model.

2. The method of claim 1, wherein, The method further comprises the following steps: in the discriminator, when extreme precipitation is identified according to the extreme precipitation identification conditions, increase the weight of the false precipitation sequence term to make the model focus more on generating high-quality extreme precipitation data; In the generator, some parameters are designated as extreme precipitation parameters; when extreme precipitation is identified according to the extreme precipitation identification conditions, the extreme precipitation parameters are enabled to enhance the modeling capability of strong convective processes; if it is not extreme precipitation, the extreme precipitation parameters are not enabled.

3. The method of claim 1, wherein, The extreme precipitation identification conditions include convective available potential energy (CAPE); if the convective available potential energy (CAPE) is greater than a set convective available potential energy threshold, it is identified as extreme precipitation.

4. The method of claim 3, wherein, The extreme precipitation identification conditions are: C ext = [CAPE, |Av 850-200 |, RH BL ] T where C ext is a vector composed of extreme precipitation identification conditions; CAPE is the convective available potential energy, |Δv 850-200 | is the vertical wind shear modulus, RH BL is the boundary layer relative humidity.

5. The method of claim 2, wherein, The discriminator loss function is: where, is the discriminator loss; E[] is the expectation operation; D() is the discriminator function, outputting a probability value between 0 and 1; X real is the real precipitation sequence; G(Z) is the precipitation sequence generated by the generator, Z is random noise, G is the generator network; γ is a weighting coefficient; when it is determined that it is extreme precipitation, γ selects γ1, otherwise γ selects γ2, γ1> γ2.

6. The method of claim 1, wherein, The determination method of the climate scenario parameters is: The large-scale climate driving signals include global temperature rise ΔΤ glob , regional sea surface temperature anomaly ΔSST reg , and Arctic sea ice change ΔIce arc ; Map the three large-scale climate driving signals to a fixed-dimensional vector of climate scenario parameters through a function f to respond to different climate scenarios.

7. The method of claim 1, wherein, The generator loss function introduces a physical equation regularization term L phy is: The first term in the formula Corresponding to the water vapor mass conservation, ensure that the precipitation generation and water vapor transport matching; wherein, P gen For the amount of precipitation, ε is the condensation efficiency coefficient, q is the specific humidity, v is the wind speed vector, ΔS is the loss term, Indicates the vertical integration first, then time integration; t0-Δt, t0 is the start and end of the time integration interval; z s , z c The lower boundary and the upper boundary of the vertical integration interval are determined according to the spatial range involved in the precipitation, and ∇· represents the divergence calculation; The second term in the formula Corresponding to the energy conservation, the correlation between precipitation and atmospheric energy change; wherein, The equivalent potential temperature, The latent heat flux; The conditional approximate proportional coefficient is obtained from the basic equation of thermodynamics; α and β are weight coefficients, respectively.

8. The method of claim 7, wherein, The weight coefficients α and β are dynamic weight coefficients: a = 0.2 + 0.6 x min(l, P obs / RES obs ) β = 0.3 + 0.4 x min(l, P obs / RES obs ) where P obs is the observed precipitation, RES obs is the threshold of daily heavy rain precipitation; min(,) means taking the minimum value; when P obs ≥ RES obs , heavy rain occurs, and the physical constraint is strengthened by increasing a and β; when it is not heavy rain or no rain, the physical constraint is weakened by decreasing a and β; when it is raining but not heavy rain, the more the precipitation, the stronger the physical constraint.

9. The method of claim 8, wherein, When heavy rain occurs, α = 0.8 and β = 0.

7.

10. The method of claim 1, wherein, In step 3, the generated samples are obtained using the generator, the generated samples and the original samples are combined, and a balanced data set with a required proportion is constructed.

Images