CN-121980183-A - High-frequency long-time-sequence short-time prediction method and system based on generated AI

Abstract

The invention relates to the technical field of precipitation prediction, in particular to a high-frequency long-time sequence short-term prediction method and a system based on generation type AI, wherein the method comprises the following steps of compressing an original radar reflectivity sequence into a compact hidden variable Token sequence by using a pre-trained variation self-encoder; in the compressed hidden space, taking the hidden variable Token sequence as a condition, and gradually generating a predicted hidden variable Token sequence of a future time period in a probability generation mode; reconstructing the predicted hidden variable Token sequence into a predicted radar reflectivity sequence of a future time period using a decoder of the variance self-encoder. The method realizes the high-resolution and long-time-effect radar proximity prediction, improves the accuracy of rainfall prediction in a short-time prediction gray zone time window of 2-6 hours, and provides solid support for early warning and accurate decision making of extreme weather construction.

Inventors

• SUN HAOFEI
• HAN WEI
• HUANG WEI
• GAO ZHIQIU
• YANG YUNFAN

Assignees

• 中国气象局上海台风研究所(上海市气象科学研究所)

Dates

Publication Date

20260505

Application Date

20260122

Claims (10)

1. 1. The high-frequency long-time sequence short-term forecasting method based on the generated AI is characterized by comprising the following steps: Compressing the original radar reflectivity sequence into a compact hidden variable Token sequence using a pre-trained variational self-encoder; in the compressed hidden space, taking the hidden variable Token sequence as a condition, and gradually generating a predicted hidden variable Token sequence of a future time period in a probability generation mode; Reconstructing the predicted hidden variable Token sequence into a predicted radar reflectivity sequence of a future time period using a decoder of the variance self-encoder.
2. 2. The high-frequency long-time-sequence short-term forecasting method based on the generated AI according to claim 1, wherein the method is characterized by comprising the following steps: The variation self-encoder is a causal variation self-encoder.
3. 3. The high-frequency long-time-sequence short-term forecasting method based on the generated AI according to claim 2, wherein the method is characterized by comprising the following steps of: the causal variation self-encoder uses causal 3D convolution.
4. 4. The high-frequency long-time-sequence short-term forecasting method based on the generated AI according to claim 1, wherein the method is characterized by comprising the following steps: And generating the prediction hidden variable Token sequence by using a diffusion generation model based on the noise Token sequence under the condition of the hidden variable Token sequence.
5. 5. The high-frequency long-time-sequence short-term forecasting method based on the generated AI of claim 4, wherein the method comprises the following steps of: the noise Token sequence is randomly sampled from Gaussian noise.
6. 6. The high-frequency long-time-sequence short-term forecasting method based on the generated AI of claim 5, wherein the method comprises the following steps: The number of frames of the noise Token sequence is consistent with the length and time step of the predicted radar reflectivity sequence.
7. 7. The high-frequency long-time-sequence short-term forecasting method based on the generated AI of claim 6, wherein the method is characterized by: the diffusion generation model is built based on RECTIFIED FLOW framework using diffusion transducers.
8. 8. The high-frequency long-time-sequence short-term forecasting method based on the generated AI of claim 7, wherein the method comprises the following steps of: The diffusion transducer uses a 3D causal self-attention mechanism to capture both local advection and non-local convection triggers.
9. 9. The high-frequency long-time-sequence short-term forecasting method based on the generated AI of claim 8, wherein the method comprises the following steps of: the diffusion transducer injects time coordinates into each layer of the network through adaptive layer normalization.
10. 10. The high-frequency long-time-sequence short-time forecasting system based on the generated AI is characterized by comprising a data input device, a data output device, a processor and a storage, wherein the storage comprises a computer readable storage medium, a computer program is stored in the computer readable storage medium, and the computer program comprises program instructions, and the program instructions when being executed by the processor cause the processor to realize the high-frequency long-time-sequence short-time forecasting method based on the generated AI according to any one of claims 1-9.

Description

High-frequency long-time-sequence short-time prediction method and system based on generated AI Technical Field The invention relates to the technical field of precipitation prediction, in particular to a high-frequency long-time-sequence short-time prediction method and system based on a generated AI. Background The accurate short-term prediction early warning for extreme precipitation is the key of global disaster prevention and reduction, however, a difficult-to-surmount 'predictability obstacle' exists for a long time in the prediction time period of 2-6 hours, namely a so-called 'short-term prediction Gray Zone'. During this time window, traditional observation-based extrapolation methods fail due to the accumulation of nonlinear errors, while Numerical Weather Prediction (NWP) is limited by computational efficiency, which has been less than resolving the dynamics of storm scale before a storm occurs. Despite the recent advances in some AI shortfall approaches, the problem is still not completely solved. Wherein the deterministic model is targeted at minimizing mean error (MSE), systematically limited by the phenomenon of "Regression-to-mean", resulting in a forecast image blur and loss of extrema. In contrast, the generation of the countermeasure network (GAN), while preserving the sharpness of the image, is prone to "Mode Collapse" and produces a "false echo (Hallucinatory Echoes)" that is not acceptable for traffic. To alleviate these problems, a recent decomposition paradigm (e.g., nowcastNet, diffCast, cascast, alphaPre) attempts to explicitly separate weather evolution into additive components, deterministic optical flow for advection and random residuals for detail. However, this manual separation is fundamentally flawed because it cuts off the causal chain which is tightly entangled during the atmosphere. In the real atmosphere, kinematics (KINEMATICS, such as movement) and Dynamics (Dynamics, such as growth) are coupled together by thermodynamic feedback loops, for example, the kinematic advection of outflow boundaries formed by rain-cooling, often is a direct kinetic mechanism that triggers new convection generation. A model that structurally strips "motion" from "growth" is blind to this mechanism. If extrapolation is made in the "gray zone", this lack of modeling ability to model the coupling evolution will directly lead to structural disintegration of the storm system. Therefore, it is necessary to explore an accurate and efficient short-term prediction method of precipitation. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a high-frequency long-time-sequence short-time forecasting method and a system based on the generated AI. In order to achieve the aim, the invention provides a high-frequency long-time sequence short-term prediction method based on generation type AI, which comprises the following steps of compressing an original radar reflectivity sequence into a compact hidden variable Token sequence by using a pre-trained variable self-encoder, gradually generating a predicted hidden variable Token sequence of a future time period in a compressed hidden space by using the hidden variable Token sequence as a condition through a probability generation mode, and reconstructing the predicted hidden variable Token sequence into the predicted radar reflectivity sequence of the future time period by using a decoder of the variable self-encoder. The invention realizes the high-resolution and long-time-effect radar proximity prediction, improves the accuracy of rainfall prediction in a time window of short-time prediction gray areas, and lays a solid foundation for constructing digital twinning of an earth system. Optionally, the variation self-encoder is a causal variation self-encoder. Optionally, the causal variation self-encoder uses causal 3D convolution. Optionally, the prediction hidden variable Token sequence is generated using a diffusion generation model based on a noise Token sequence, on the condition of the hidden variable Token sequence. Optionally, the noise Token sequence is randomly sampled from gaussian noise. Optionally, the number of frames of the noise Token sequence corresponds to the length and time step of the predicted radar reflectivity sequence. Optionally, the diffusion generation model is built using a diffusion transducer based on RECTIFIED FLOW framework. Optionally, the diffusion transducer captures both local advection and non-local convection triggers using a 3D causal self-attention mechanism. Optionally, the diffusion transducer injects time coordinates into each layer of the network by adaptive layer normalization. In a second aspect, the present invention provides a high frequency long-time-sequence short-time prediction system based on a generated AI, the high frequency long-time-sequence short-time-prediction system based on a generated AI comprising a data input device, a data output device, a processor and a sto