Southeast Asia heat wave season forecasting method based on tropical waves and artificial intelligence
By extracting tropical wave signals using Kalman filtering and Fourier transform, and combining them with a conditional diffusion model based on a multi-scale feature fusion module, the error problem in the sub-seasonal forecast of Southeast Asian heat waves was solved, achieving more accurate forecast results.
Patent Information
- Application Number
- CN202511653190.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-06
AI Technical Summary
Existing technologies have significant errors in forecasting the second season of heat waves in Southeast Asia. Traditional linear statistical methods are ineffective in characterizing the interactions between tropical waves, which limits the improvement of forecasting skills.
Data fusion was performed using the Kalman filter method. Tropical wave signals were extracted through Fourier transform and bandpass filtering. Combined with the conditional diffusion model of the multi-scale feature fusion module, artificial intelligence was used to optimize model training and generate a spatiotemporal distribution forecast of heat wave risk.
It significantly improves the reliability and sophistication of subseasonal heatwave forecasts in Southeast Asia, enabling it to reflect the complex spatiotemporal characteristics of heatwaves and quantify uncertainties.
Smart Images
Figure CN121479265A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of heatwave subseason forecasting technology, specifically to a method for forecasting Southeast Asian heatwave subseasons based on tropical waves and artificial intelligence. Background Technology
[0002] Against the backdrop of global warming, accurate sub-seasonal (2 weeks to February) forecasts of heat waves are crucial for disaster prevention, mitigation, and risk management. However, current sub-seasonal heat wave forecasting techniques remain quite limited. Current sub-seasonal heat wave forecasts primarily rely on sub-seasonal numerical models, but due to the complexity of sub-seasonal physical processes, the diversity of predictability sources, and inherent limitations of numerical models, mainstream sub-seasonal numerical models exhibit significant errors in forecasting Southeast Asian heat waves. Therefore, it is necessary to correct the forecast results before using them. Fusion of multi-source models based on long-term forecasts and observational data is an effective means of improving the forecasting technique of numerical models. Besides further improving the forecasting technique of numerical models, diagnosing and analyzing the potential predictability sources of heat wave events within numerical models and constructing dynamic statistical models based on this analysis is another effective means of improving the sub-seasonal heat wave forecasting technique.
[0003] In recent years, studies have shown that tropical waves of different scales, such as the Madden-Julian Oscillation (MJO), Kelvin wave, and Rossby wave, play an important modulating role in the generation of heat waves in Southeast Asia. They can serve as an important predictability source for forecasting heat wave events in Southeast Asia. However, due to their structural complexity and multi-timescale interaction characteristics, traditional linear statistical methods are unable to effectively characterize and utilize the interactions between these wave signals, thus limiting the improvement of forecasting skills. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide a method for forecasting the subseasonal heatwave in Southeast Asia based on tropical waves and artificial intelligence, thereby solving the problems existing in the background technology.
[0005] Technical solution: The present invention provides a method for forecasting the second season of Southeast Asian heat waves based on tropical waves and artificial intelligence, comprising the following steps:
[0006] (1) After collecting and preprocessing multi-source meteorological forecast data and observation data, the forecast data are dynamically corrected and fused based on the Kalman filter method;
[0007] (2) Using Fourier transform and bandpass filtering, tropical wave signals of different scales are separated and extracted from the data, including Madden-Julian oscillation (MJO), Kelvin wave and Rossby wave, to construct a dataset reflecting the multi-scale characteristics of tropical waves;
[0008] (3) Establish a conditional diffusion model that integrates multi-scale feature fusion modules, combine tropical wave characteristics with regional meteorological elements, and optimize the model based on historical training sets;
[0009] (4) Input the meteorological forecast data at the preset time into the trained conditional diffusion model to generate the corresponding heat wave risk spatiotemporal distribution forecast product.
[0010] Furthermore, in step (1), the preprocessing is as follows: outlier removal and spatiotemporal interpolation of missing values; wherein, spatiotemporal interpolation includes linear interpolation in the time dimension and inverse distance weighted interpolation in the spatial dimension.
[0011] Furthermore, in step (1), the dynamic correction and fusion of forecast data based on the Kalman filter method is as follows: the weights of multi-source forecast data are dynamically adjusted by Kalman gain to realize real-time correction of forecast errors and fusion of multi-source data.
[0012] Furthermore, in step (2), the Fourier transform and bandpass filtering methods are as follows: perform a two-dimensional fast Fourier transform on the data to obtain the meridional wavenumber-frequency spectrum, and extract the zonal wavenumber and frequency range tropical wave signal through bandpass filtering.
[0013] Furthermore, the extraction of tropical wave signals specifically involves selecting the eastward propagation component of the Madden-Julian Oscillation (MJO) with its zonal wavenumber and period range.
[0014] Furthermore, the extraction of tropical wave signals also includes combining the filtered zonal wind and geopotential height into new variables and projecting them onto meridional basis functions to identify Kelvin waves and Rossby waves.
[0015] Furthermore, in step (3), the conditional diffusion model integrating the multi-scale feature fusion module includes a forward noise addition process and a reverse noise reduction process. The reverse noise reduction process uses tropical wave signals and numerical model forecast products as conditional variables, and captures the long-range dependence and multi-scale interaction between conditional variables through the multi-scale feature fusion module.
[0016] Furthermore, the multi-scale feature fusion module includes convolutional layers, downsampling layers, a self-attention mechanism module, and a feature fusion and residual enhancement module, which are used to achieve spatial alignment of multi-scale features, capture of global dependencies, and cross-scale information interaction.
[0017] An electronic device according to the present invention includes a memory and a processor. The memory stores a computer program, and the processor executes the program to implement the steps of the method.
[0018] The present invention discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method.
[0019] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention fully considers the modulation effect of tropical waves on the generation and evolution of heat waves in Southeast Asia, using it as a key input feature of the model to ensure that the forecast model is built on a solid physical foundation. This invention characterizes the interaction of different tropical waves through a conditional diffusion model, reflecting the complex spatiotemporal characteristics of heat wave evolution and quantifying uncertainties through ensemble prediction, thereby significantly improving the reliability and sophistication of forecasts. Attached Figure Description
[0020] Figure 1 This is a flowchart of the present invention;
[0021] Figure 2 This is a schematic diagram of the model structure of the present invention;
[0022] Figure 3 This is the multi-scale feature fusion module of the present invention. Detailed Implementation
[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0024] like Figure 1 As shown, this embodiment of the invention provides a method for forecasting the second season of heat waves in Southeast Asia based on tropical waves and artificial intelligence, including the following steps:
[0025] (1) Collect multi-source meteorological forecast data and observation data, preprocess the data, and then dynamically correct and fuse the forecast data based on the Kalman filter method; including the following steps:
[0026] (11): Collect multi-source meteorological forecast data and observation data, remove outliers from the collected data, and perform spatiotemporal interpolation on missing data. The temporal interpolation method is linear interpolation, while the spatial interpolation method is inverse distance weighting, and its calculation formula is as follows:
[0027]
[0028] Formula (1) is the time interpolation method, where , For two moments when data is available, and For the data at these two times, For the time of missing measurement, The data is obtained by interpolation at the time of missing measurement.
[0029]
[0030] Formula (2) is a spatial interpolation method, where For missing measurement points, For the n points with data surrounding the missing measurement point, These are the data points. This indicates the distance from these points to the missing measurement points. The data obtained by interpolation for missing measurement points.
[0031] (12) The Kalman filter method is used to fuse multi-source numerical models to reduce the error of forecast products. The core calculation formula is as follows:
[0032]
[0033]
[0034] Where t represents time. H is the forecast matrix, containing the forecast results for all models, where f is the model value, N is the number of models, and M is the number of forecast grid points. W is a state vector containing model weight information, determined through iterative calculation during the training period.
[0035]
[0036]
[0037]
[0038]
[0039]
[0040] Where the subscript t indicates time t, f and a represent the forecast state and analysis state respectively, P represents the error covariance matrix, Q is the process noise covariance, R is the observation noise covariance, K is the Kalman gain, used to dynamically adjust the model weights, and Y is the actual observation result.
[0041] (2) Using Fourier transform and bandpass filtering methods, tropical wave signals of different scales, including Madden-Julian oscillations (MJO), Kelvin waves, and Rossby waves, are separated and extracted from the data to construct a dataset reflecting the multi-scale characteristics of tropical waves; including the following steps:
[0042] (21) Extract the global wind field (u, v) and geopotential height Z at 6-hour resolution for the tropical region (24°S–24°N) and interpolate them to a 1°×1° grid.
[0043] (22) Select a 90-day sliding window. For the tstep day forecast, the observations of the previous (90-tstep) days and the forecasts of the next tstep days constitute a 90-day sliding window.
[0044] (23) Perform a two-dimensional fast Fourier transform (FFT) to obtain the meridional wavenumber-frequency (k-ω) spectrum:
[0045]
[0046] in For input data, Indicates Fourier transform, For zonal wavenumber, Angular frequency, These are the spectral coefficients. Based on this, bandpass filtering is performed on the spectrum to extract the zonal wavenumber components in the range k=2~40, and according to... / The symbols separate the eastward and westward propagating waves. / >0 indicates the eastward transmission component. / <0 indicates the westward propagation component. Based on this, the components with zonal wavenumbers k=1–3 and periods of 30–90 days are selected as candidate signals for the MJO, and only the eastward propagation component is retained to conform to the spatial-temporal propagation characteristics of the MJO.
[0047] (24) Combine the filtered zonal wind and geopotential height into a new variable:
[0048]
[0049]
[0050] Where g is the acceleration due to gravity. The equivalent gravitational wave velocity, q and r are the transformed combined variables, with the dimension of velocity (m·s⁻¹), which corresponds to the Riemann invariant in shallow water wave theory.
[0051] (25) Define the meridional basis function of the equatorial shallow water wave, and project the field obtained in step 2.4 onto the meridional basis function. The formula for calculating the meridional basis function is as follows:
[0052]
[0053] in, Let n be the meridional characteristic function. Let y be the latitude coordinate of the polynomial related to the Hermite polynomial. The equatorial notch scale is calculated using the following formula:
[0054]
[0055] in Here is the equatorial β parameter. The formula for calculating the projection is as follows:
[0056]
[0057] in This refers to the spectral variables after filtering and separating the east-west direction in step 2.4. These are the projection coefficients on the nth mode, representing the strength of the mode's contribution to the disturbance.
[0058] (26) Based on the combination of characteristic coefficients of different wave modes, the specific tropical wave type is identified. Among them, the Kelvin wave is mainly composed of... Characterized by, Rossby waves are mainly composed of and Common characteristics.
[0059] (3) Establish a conditional diffusion model that integrates a multi-scale feature fusion module, combining tropical wave characteristics with regional meteorological elements, and optimize the model based on historical training sets; specifically as follows: Figure 2 As shown, the conditional diffusion model includes forward denoising and backward denoising processes. The core formula for the forward denoising process is:
[0060]
[0061] in, This is the initial image. This is the image after adding noise for t steps. represent The cumulative product, It is standard Gaussian noise. Specifically, the actual heat wave observation is the original image ( By continuously adding noise, the noise product is obtained. Through this process, the heat wave observation is gradually diffused from a deterministic field into pure noise, forming a controllable degradation process, providing a training target for reverse denoising learning.
[0062] In the reverse denoising process, the invention uses the noise generated during the noise addition process, the original forecast products of the numerical model, and the tropical wave signals (MJO, Rossby wave, Kelvin wave) extracted in step 2 as conditional variables. After feature extraction and encoder compression, these are processed by a multi-scale feature fusion module (whose structure is as follows). Figure 3(As shown) This invention aims to capture long-range dependencies and multi-scale interactions between different conditional variables and spatial-temporal features, ultimately obtaining actual heat wave observations through a decoder. The multi-scale feature fusion module in this invention mainly includes: multi-layer input features obtained through downsampling, convolutional layers, downsampling layers, a self-attention mechanism module, and a feature fusion and residual enhancement module. Specifically, the multi-scale features obtained by the encoder undergo feature extraction and channel alignment through a 3×3 convolutional layer, and spatial alignment of the multi-scale features is achieved through downsampling. Subsequently, self-attention calculation is performed independently on the features at each scale to capture global dependencies in the spatial and channel dimensions. Simultaneously, feature interaction connections are established between different scales to achieve cross-scale information association and sharing, thereby improving the expressive power and contextual consistency of the features. Finally, the multi-scale features processed by the attention module are concatenated along the channel dimension, and the fused features are added to the input features to form a residual connection, outputting a comprehensive multi-scale attention feature, achieving information fusion and stable gradient propagation.
[0063] The core formula for reverse noise reduction is as follows:
[0064]
[0065] Where C is the set of input conditions. The mean value predicted by the neural network. This is the variance term, which can take a fixed value. This distribution characterizes the probability of generating a sample at time t−1 given a condition set C, thus enabling a gradual reconstruction of real heatwave observations. During model training, the corresponding output is the actual heatwave observation. After model training, multiple samplings can directly generate ensemble forecast products of the heatwave field, which improves the model's generalization ability and also characterizes atmospheric uncertainties.
[0066] (4) Input the meteorological forecast data at the preset time into the trained conditional diffusion model, and generate the corresponding heat wave risk spatiotemporal distribution forecast product by ensemble prediction to quantify the forecast uncertainty.
Claims
1. A method for predicting the subseasonal heatwave in Southeast Asia based on tropical waves and artificial intelligence, characterized in that, Includes the following steps: (1) After collecting and preprocessing multi-source meteorological forecast data and observation data, the forecast data are dynamically corrected and fused based on the Kalman filter method; (2) Using Fourier transform and bandpass filtering, tropical wave signals of different scales are separated and extracted from the data, including Madden-Julian oscillation (MJO), Kelvin wave and Rossby wave, to construct a dataset reflecting the multi-scale characteristics of tropical waves; (3) Establish a conditional diffusion model that integrates multi-scale feature fusion modules, combine tropical wave characteristics with regional meteorological elements, and optimize the model based on historical training sets; (4) Input the meteorological forecast data at the preset time into the trained conditional diffusion model to generate the corresponding heat wave risk spatiotemporal distribution forecast product.
2. The method for forecasting the subseasonal heatwave in Southeast Asia based on tropical waves and artificial intelligence according to claim 1, characterized in that, In step (1), the preprocessing is as follows: outlier removal and spatiotemporal interpolation of missing values; wherein, spatiotemporal interpolation includes linear interpolation in the time dimension and inverse distance weighted interpolation in the spatial dimension.
3. The method for subseasonal forecasting of Southeast Asian heat waves based on tropical waves and artificial intelligence according to claim 1, characterized in that, In step (1), the dynamic correction and fusion of forecast data based on the Kalman filter method is as follows: the weights of multi-source forecast data are dynamically adjusted by Kalman gain to realize real-time correction of forecast errors and fusion of multi-source data.
4. The method for subseasonal forecasting of Southeast Asian heat waves based on tropical waves and artificial intelligence according to claim 1, characterized in that, In step (2), the Fourier transform and bandpass filtering methods are as follows: perform a two-dimensional fast Fourier transform on the data to obtain the meridional wavenumber-frequency spectrum, and extract the zonal wavenumber and frequency range tropical wave signal through bandpass filtering.
5. The method for subseasonal forecasting of Southeast Asian heat waves based on tropical waves and artificial intelligence according to claim 4, characterized in that, The extraction of tropical wave signals specifically involves selecting the eastward propagation component of the zonal wavenumber and period range of the Madden-Julian Oscillator (MJO).
6. The method for subseasonal forecasting of Southeast Asian heat waves based on tropical waves and artificial intelligence according to claim 4, characterized in that, Extracting tropical wave signals also involves combining filtered zonal wind and geopotential height into new variables and projecting them onto meridional basis functions to identify Kelvin and Rossby waves.
7. The method for subseasonal forecasting of Southeast Asian heat waves based on tropical waves and artificial intelligence according to claim 1, characterized in that, In step (3), the conditional diffusion model integrating the multi-scale feature fusion module includes a forward noise addition process and a reverse noise reduction process. The reverse noise reduction process uses tropical wave signals and numerical model forecast products as conditional variables and captures the long-range dependence and multi-scale interaction between conditional variables through the multi-scale feature fusion module.
8. The method for forecasting the subseasonal heatwave in Southeast Asia based on tropical waves and artificial intelligence according to claim 7, characterized in that, The multi-scale feature fusion module includes convolutional layers, downsampling layers, a self-attention mechanism module, and a feature fusion and residual enhancement module, which are used to achieve spatial alignment of multi-scale features, capture of global dependencies, and cross-scale information interaction.
9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the program to implement the steps of the method according to claims 1-8.
10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the method described in claims 1-8.