Typhoon prediction method based on improved diffusion model

Abstract

The application discloses a typhoon prediction method based on an improved diffusion model and belongs to the technical field of meteorological prediction and artificial intelligence. The method comprises the following steps: collecting a typhoon best path and ERA5 data and performing Z-Score standardization; constructing a sample through a sliding window, inputting 24h typhoon and environmental variables, outputting 1 / 6 / 12h variables, and layering according to individual cases; training a double-branch LSTM prior; constructing a conditional diffusion model, the conditional encoder and the prior have the same structure, the denoising network takes noise estimation, a time step, a conditional context and the prior as inputs, and the denoising network is trained under cosine scheduling by weighting mean square error and Haversine physical loss; generating a set through DDIM sampling and adding the set to the prior; searching for a scaling factor to perform probability calibration on a verification set; and evaluating the method by MAE, RMSE and CRPS. The method considers both prediction accuracy and probability reliability and is suitable for typhoon early warning services.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of weather forecasting and artificial intelligence, and in particular to a typhoon forecasting method based on an improved diffusion model. Background Technology

[0002] Therefore, accurate typhoon forecasting has always been a key challenge in meteorological science and disaster prevention and mitigation. Traditional typhoon forecasting mainly relies on numerical weather prediction models and statistical-dynamic models: the former calculates the future atmospheric state on large-scale supercomputers by solving partial differential equations of atmospheric motion, such as the European Centre for Medium-Range Weather Forecasts' global model and the US National Environmental Prediction Center's hurricane-specific model; the latter establishes regression or similarity forecasting relationships between climate persistence factors and the future state of typhoons based on historical data, such as the classic SHIPS statistical hurricane intensity forecasting scheme. However, numerical models are extremely sensitive to initial conditions and have huge computational costs, and parameterized schemes are difficult to accurately characterize the fine-grained convection processes in the typhoon's core region; statistical models are difficult to capture complex nonlinear evolution, especially in predicting extreme cases such as rapid intensification and abrupt path changes.

[0003] In recent years, deep learning technology has been widely applied to weather forecasting. Models such as convolutional neural networks, graph neural networks, long short-term memory networks, and attention-based Transformers have achieved a series of results in typhoon track and intensity prediction, automatically extracting spatiotemporal features from reanalysis data or satellite imagery. Building on this, diffusion models, as an emerging generative modeling paradigm, have begun to be applied to ensemble forecasting and spatiotemporal field reconstruction of meteorological elements due to their superior ability in modeling high-dimensional complex distributions. Examples include short-term precipitation forecasts, radar echo extrapolation, and ensemble generation of global meteorological fields. Diffusion models progressively add noise to the data through a forward process until it becomes a prior distribution, and then use a backward process to learn and denoise the data to generate samples that conform to the conditional distribution. This approach is naturally suitable for expressing the inherent uncertainties and multimodal evolution characteristics of typhoon forecasting. However, the application of diffusion models to typhoon forecasting is still in its early stages. Most of the existing preliminary attempts have directly transferred the diffusion framework used for general precipitation or circulation field generation to the typhoon scenario. The integration of information such as the vortex dynamic structure, boundary layer and eyewall strong gradient characteristics, and ocean-atmosphere coupling constraints unique to typhoons is still insufficient, resulting in problems such as blurred vortex morphology, increased center position drift, and underestimation of extreme intensity in long-term forecasts.

[0004] Current existing technologies similar to this invention mainly fall into two categories: one is a typhoon prediction method based on deterministic deep learning, and the other is a general diffusion model for meteorological field generation. The first type of method is essentially a point-to-point deterministic mapping, which lacks the ability to characterize the probability distribution in typhoon prediction. It is difficult to reflect the set of possible paths in a single forecast sample, and it suffers from over-smoothing and amplitude saturation problems in capturing extreme intensity abrupt changes. The second type of method applies conditional denoising diffusion probability models to high-resolution radar echo prediction or probabilistic forecasting of global meteorological fields. However, when directly applied to typhoon prediction tasks, several significant shortcomings exist: the conditional encoding of general diffusion models typically only performs global pooling or simple splicing of historical environmental fields, without designing a spatially adaptive conditional mechanism for isolated vortex systems like typhoons with asymmetric and compact structures. This leads to the appearance of typhoon eye filling, eyewall structure breakage, or false circulation in the generated samples, resulting in a systematic underestimation of intensity. The backsampling process of conventional diffusion models requires hundreds or thousands of inference steps to obtain high-quality samples, making it difficult to meet the stringent computational efficiency requirements of real-time typhoon early warning operations. Furthermore, due to the lack of physical constraints in long-distance temporal evolution, the predicted trajectory is prone to deviating from the phase space constrained by the actual atmospheric motion equations. Existing methods generally only use mean square error or simple absolute error when defining noise scheduling and training objectives, lacking sensitivity to the sharp gradient changes in the peak region and core region of typhoons during their rapid development. They cannot actively enhance the fitting of high-speed typhoons, thus limiting the ability to rapidly enhance the early detection of events. Therefore, there is an urgent need to make targeted improvements to the existing diffusion model so that it can maintain the advantages of probabilistic and fine structure generation while conforming to the dynamic characteristics of typhoons, thereby improving the prediction accuracy, efficiency and reliability in extreme scenarios. Summary of the Invention

[0005] To address the problems in existing diffusion models for typhoon prediction, such as distorted generated sample structures due to the lack of specific design for conditional encoding for vortex structures, inconsistent path dynamics due to excessive backsampling steps and lack of physical constraints, and inadequate fitting of extreme intensity due to the insensitivity of global equilibrium loss to the typhoon peak region, this invention aims to provide a high-fidelity, high-efficiency, and physically consistent typhoon probability prediction method. Through a conditional diffusion model framework guided by deterministic priors, this method maintains the advantages of probabilistic generation while conforming to typhoon dynamic characteristics, thereby improving prediction accuracy, computational efficiency, and reliability in extreme scenarios.

[0006] To achieve the above objectives, this invention provides a typhoon forecasting method based on an improved diffusion model, comprising the following steps: S1. Obtain historical typhoon best path data and meteorological reanalysis data, unify spatial resolution and construct multivariate time series, and perform Z-Score standardization. S2. Construct samples using a sliding window based on individual typhoon cases. Input the typhoon and environmental variable sequences of the past 24 hours and output the typhoon variables at future times. Divide the samples into training set, validation set and test set according to individual cases. S3. Construct a dual-branch LSTM network. After the typhoon branch and the environment branch are fused by cross-attention, the time-sharing output head generates predictions of typhoon variable points for each time period and trains the prior model. S4. Construct a conditional diffusion model. Its conditional encoder uses the same dual-branch LSTM and cross-attention structure as S3 to output the conditional context vector. The denoising network takes the noise target estimation, time step, conditional context vector and the point prediction obtained from S3 as inputs and outputs the direct prediction of the original target. It adopts cosine noise scheduling and the loss function is composed of the mean square error and the Haversine spherical distance physical constraint loss weighted together. S5. Using the samples from S2 and the prior predictions from S3, train the S4 diffusion model. During training, all parameters of the prior model are frozen, and its output is used only as static conditional input, not participating in gradient backpropagation. The Adam optimizer, supplemented by learning rate decay and early stopping strategies, is used to obtain the conditional diffusion model. S6. During testing, a deterministic forecast is first generated by the S3 prior model, and then the S5 conditional diffusion model is randomly sampled multiple times. The samples are generated by DDIM fast sampling and denoising, and then added to the S3 prior prediction to obtain the ensemble forecast. S7. On the S2 validation set, search for a scaling factor for each forecast variable. Based on the width of the set quantile interval, determine the optimal scaling factor by minimizing the absolute error between the coverage and the target confidence level. Apply it to the test set to obtain the calibrated prediction interval. S8. Calculate the mean absolute error, root mean square error, and continuous graded probability score of the test set to evaluate the forecast performance.

[0007] Preferably, in S1, the historical typhoon best path dataset includes time, longitude, latitude, maximum sustained wind speed, and minimum central pressure; the meteorological grid reanalysis data for the target area includes ERA5 reanalysis data for the southeastern sea region of my country with a resolution of 0.25∘×0.25∘, specifically including horizontal wind components U10 and V10 at a height of 10m, temperature T2m at 2m, surface pressure SP, and relative humidity RH500, temperature T500, and horizontal wind components U500 and V500 on the 500hPa isobaric surface; the Z-Score standardization formula is: ; in, σ and σ are the mean and standard deviation of each variable on the training set, respectively. The standardized data are used for subsequent model training.

[0008] Preferably, in S2, the input of the sliding window is the typhoon variables and environmental field variables of the past 24 hours (24 steps), and the output is the typhoon variables at the next 1 hour, 6 hours and 12 hours. The typhoon variables include longitude, latitude and maximum wind speed. The training set is data from 2021 to 2023, the validation set is data from 2024, and the test set is data from 2025 to ensure that there is no crossover of samples of the same typhoon between different sets.

[0009] Preferably, in S3, the typhoon branch uses a two-layer LSTM with 64 hidden units each, and the environment branch uses a two-layer LSTM with 128 hidden units each; a cross-attention mechanism fuses the output sequences of the two branches to capture the interaction between the typhoon and the environmental field; the loss function uses weighted mean square error, with latitude and longitude components given greater weight to emphasize location accuracy; training uses an adaptive learning rate optimizer, with learning rate decay and early stopping strategies; the trained model is denoted as... Its output point prediction is denoted as This information is provided as prior information to subsequent diffusion models.

[0010] Preferably, in S4, the conditional diffusion model is an improved denoised diffusion probability model DDPM; the conditional encoder takes a historical observation sequence x as input and outputs a fixed-length conditional context vector c; the denoising network is a multilayer perceptron, and its input includes the current noisy target estimate. The diffusion time step t, the conditional context vector c, and the complete prediction provided by the prior model. The output is the original standardized target. The direct prediction form; the total number of diffusion steps T is set to 50 steps, and cosine noise scheduling is used; the forward process noise addition is based on the formula Execution, in which , Noise scheduling coefficients; denoising network by( ( ) is the input, and the output is the predicted original target. ; The loss function is: ; in, For mean square error loss, For physical constraint loss, weights To balance the two terms; the physical constraint loss uses the Haversine spherical distance, and for each forecast time k, the true longitude is set to... latitude is The predicted longitude is latitude is All values ​​are expressed in radians, representing the Haversine distance at that time. The formula for calculation is: ; Where R = 6371 km, , Physical constraint loss Defined as the average Haversine distance across all forecast times, the expression is: Where H is the number of forecast times.

[0011] Preferably, in S5, training samples are used ( ) and the corresponding prior prediction Simultaneously, the diffusion model is trained using the Adam optimizer with an initial learning rate of... The batch size is 128; during training, the validation set loss is monitored. If there is no improvement after 10 consecutive epochs, the learning rate is halved; if there is no improvement after 20 consecutive epochs, early stopping is triggered and the optimal model weights are restored; after training, the conditional diffusion model is obtained. .

[0012] Preferably, in S6, for each sample x in the test set, the prior model is first used. Deterministic forecasts Then, using the conditional diffusion model Perform N random samplings, each starting with isotropic Gaussian noise, and iterate using the DDIM fast sampling algorithm. Denoising is performed to generate multiple possible forecast samples. The DDIM sampling iteration formula is: ; in =( The original target predicted by the denoising network is the standard deviation of random noise. The expression is: ; Setting randomness parameters during sampling The value range is from 0.5 to 1.5, preferably, =1.0 to stimulate diversity in the set; the generated samples are added to the prior predictions to obtain the final set.

[0013] Preferably, in S7, a scaling factor is independently searched for for each forecast variable on the validation set. This is used for post-calibration of the prediction interval; the specific method is as follows: calculate the 25th and 75th quantiles of the validation set, and calculate the difference between the two to obtain the interval width. Preliminary analysis on the validation set showed that when the scaling factor was below 0.5, the interval was too narrow and the coverage was severely low; when it was above 5.0, the interval width was excessively expanded and the sharpness was significantly deteriorated. Therefore, the search was limited to the range [0.5, 5.0]. This makes the scaled interval The coverage of the true values ​​reached the target confidence level of 75%, of which The target value is the median of the set. The search process is achieved by minimizing the absolute error between the coverage and the target value. The optimal scaling factor obtained from the search is saved and applied to the set prediction of the test set to obtain the calibrated prediction interval.

[0014] Preferably, in S8, the deterministic indicators include Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), and the probabilistic indicators include Continuous Graded Probability Score (CRPS); the formula for calculating CRPS is: ; The results of the calculations are compared with those of common time series prediction models. The results are visualized and evaluated.

[0015] The advantages and beneficial effects of this invention compared to the prior art are: 1. Significantly improved forecast accuracy. The prior LSTM network fully extracts the typhoon's own evolution characteristics and environmental field interaction information through a dual-branch architecture and cross-attention mechanism. The conditional diffusion model is generated on this basis, effectively correcting the systematic bias of the single model. As a result, the path and intensity errors of the ensemble forecast mean are lower than those of conventional deep learning models across multiple forecast lead times.

[0016] 2. Physically consistent and highly reliable probability output. By introducing a differentiable Haversine spherical distance physical constraint loss during diffusion model training and adaptively scaling and calibrating the generated set, the problem of underestimated forecast variance in diffusion models with moderate amounts of data is solved, providing calibrated and reliable probability data for typhoon risk warnings.

[0017] 3. High computational efficiency and easy engineering implementation. By adopting a low diffusion step count and a fast DDIM sampling strategy, the computational cost of training and inference is significantly reduced while ensuring generation quality. This allows the method to be stably trained on datasets with tens of thousands of samples, making it suitable for practical meteorological operational applications.

[0018] 4. Enhanced Prediction Sensitivity in Extreme Scenarios. Through the synergistic effect of physical constraint loss and probabilistic calibration mechanism, the model's ability to capture rare but high-impact events such as rapid typhoon intensification is improved, and the fitting effect at abrupt gradient changes in the peak and core regions is enhanced.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart of a typhoon prediction method based on an improved diffusion model, as described in an embodiment of the present invention. Figure 2 This is the result of a one-hour multi-model comparison test of Typhoon 1 in an embodiment of the present invention; Figure 3 This is the result of a six-hour multi-model comparison test of Typhoon 2 in an embodiment of the present invention; Figure 4 This is the result of a 12-hour multi-model comparison test of Typhoon 3 in an embodiment of the present invention. Detailed Implementation

[0021] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] like Figures 1-4As shown, this invention provides an overall technical solution for a typhoon forecasting method based on an improved diffusion model. The core innovation of this method lies in coupling deterministic deep learning priors with a conditional diffusion model, constructing a complete typhoon forecasting technology chain from data preprocessing, prior modeling, probability generation to post-calibration. The diffusion model is a generative neural network that progressively adds Gaussian noise to the data through a forward process until the signal is annihilated, and then recovers samples conforming to the target distribution from the pure noise through a reverse denoising process. In typhoon forecasting scenarios, this model can naturally characterize the uncertainty of typhoon path and intensity evolution. However, the pure diffusion model suffers from problems such as unstable training and underestimated variance of generated samples under medium-sized sample conditions. Therefore, this solution introduces a pre-trained LSTM deterministic model as a prior guide, injecting point predictions into the denoising network, allowing the diffusion process to generate micro-perturbations near the precise mean, preserving the dispersion capability of the probability set while ensuring that the forecast trajectory does not deviate from the actual dynamic constraints.

[0024] In some embodiments, historical typhoon best track data are derived from typhoon yearbooks published by meteorological service agencies, including typhoon center longitude, latitude, maximum sustained wind speed, and minimum central pressure observed every 6 hours. Meteorological reanalysis data uses the ERA5 dataset, which was obtained by the European Centre for Medium-Range Weather Forecasts (ECMWF) through reanalysis of historical observation data using numerical weather prediction models, providing globally uniformly gridded three-dimensional atmospheric state variables. This scheme selects key areas along the southeast coast of my country, with a spatial resolution of 0.25°×0.25°. The vertical layers include horizontal wind components U10 and V10 at a height of 10m, temperature T2m at 2m, surface pressure SP, and relative humidity RH500, temperature T500, and horizontal wind components U500 and V500 at the 500hPa isobaric surface. Since the dimensions and numerical ranges of the variables differ significantly, directly inputting them into the neural network would lead to gradient update imbalance; therefore, Z-Score standardization is used: the mean of each variable is calculated on the training set. with standard deviation Convert the original value X to This makes each variable approximately follow a standard normal distribution with zero mean and unit variance, thereby accelerating model convergence and eliminating the influence of dimensions.

[0025] In some embodiments, training samples are generated using a sliding window mechanism, with each typhoon's lifecycle as an independent unit. The window length is fixed at the past 24 hours, corresponding to 24 time steps. Each step includes the typhoon variables (longitude, latitude, maximum wind speed) and environmental field variables (the aforementioned ERA5 elements) at that moment. The window slides forward in fixed steps (1 hour), outputting the typhoon variables for the next 1, 6, or 12 hours. To prevent information leakage and ensure the independence of the test, the dataset is hierarchically divided by typhoon cases rather than by time steps: all typhoon cases from 2021 to 2023 are included in the training set, cases from 2024 are included in the validation set, and cases from 2025 are included in the test set, ensuring that consecutive time steps of the same typhoon do not appear simultaneously in both the training and test sets. This division method simulates the forecasting scenario of unknown typhoons in real forecasting operations, making the evaluation of the model's generalization performance more reliable.

[0026] In some embodiments, Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network that effectively alleviates the vanishing gradient problem in long sequences by introducing gating mechanisms (input gate, forget gate, output gate), making them particularly suitable for handling meteorological sequences like typhoons with significant time-series dependencies. This scheme designs a dual-branch architecture: the typhoon branch only inputs the typhoon's own variable sequence, employing two LSTM layers, each with 64 hidden units, to capture the internal evolution rhythm of the typhoon; the environment branch inputs a large-scale environmental field sequence, employing two LSTM layers, each with 128 hidden units, to extract external forcing information such as steering airflow and sea surface temperature. The hidden state sequences of the two branches are fused through a cross-attention mechanism: cross-attention allows the hidden state of the typhoon branch at each time step to "query" the hidden states of the environment branch at each time step, and calculates attention weights for weighted aggregation, thereby explicitly modeling the nonlinear interaction between the typhoon and the environmental field. The fused features are compressed into a fixed-dimensional vector by global average pooling, then mapped through a fully connected layer, and finally output by a time-division output head for 1-hour, 6-hour, and 12-hour forecast values, respectively. During training, a weighted mean squared error loss method was used, with higher weights assigned to latitude and longitude components, as path error is a core indicator for typhoon forecasting. The optimization process employed an adaptive learning rate strategy, coupled with an early stopping mechanism to prevent overfitting. After training, the model was denoted as [model name missing]. Its output This serves as a conditional prior for subsequent diffusion models.

[0027] In some embodiments, an improved Denoising Diffusion Probability Model (DDPM) is employed, consisting of a conditional encoder and a denoising network. The conditional encoder reuses the dual-branch LSTM with cross-attention architecture from S3, encoding the historical observation sequence x into a fixed-length conditional context vector c, ensuring that the conditional information and the prior model reside in the same semantic space. The denoising network employs a multilayer perceptron (MLP), whose input dimension includes the current noisy target estimate z. tThe diffusion time step t (embedded via sinusoidal position encoding), the conditional context vector c, and the prior predictions. prior The output is a direct prediction of the original standardized objective y0. This prediction method is more conducive to stable training of regression tasks compared to prediction noise. The total number of diffusion steps T is set to 50 steps, and cosine noise scheduling is used to make the signal-to-noise ratio transition smoother. The forward noise addition process follows... ,in The loss function is a weighted average of a data fidelity term and a physical constraint term. The physical constraint term uses Haversine spherical distance, a formula based on spherical trigonometry that directly penalizes latitude and longitude errors by calculating the great circle distance between the predicted and actual points. In the formula, R = 6371 km is the average radius of the Earth, and Δ... Δλ and Δt are the latitude difference and longitude difference, respectively, both expressed in radians; this differentiable loss forces the diffusion model to learn spatial positional relationships that conform to Earth's geometric constraints during gradient backpropagation, rather than simply fitting values ​​in Euclidean coordinates.

[0028] In some embodiments, training data is organized as triples (x, y0, ..., y0, ...). prior ), where x is the historical sequence and y0 is the true target. prior It is derived from forward inference of the frozen S3 prior model. The optimizer used is Adam, which combines momentum and adaptive learning rate adjustment mechanisms, with an initial learning rate set to 1×10⁻⁶. -4 The batch size is 128. During training, the loss is evaluated on the validation set after each epoch. If the validation loss does not decrease for 10 consecutive epochs, the learning rate is halved to escape local optima. If there is still no improvement after 20 consecutive epochs, early stopping is triggered, and the historically optimal model weights are restored to avoid degradation in generation quality due to overfitting. This combination of gradual learning rate decay and early stopping strategy is particularly important in scenarios with limited typhoon sample sizes (tens of thousands), ensuring that the model maintains its generalization ability to unknown typhoons while fully fitting the training data. After training, the conditional diffusion model is obtained. .

[0029] In some embodiments, for the test samples to be predicted, the prior model f is first used. prior Generate a deterministic forecast trajectory prior Then, the stochastic generation process of the conditional diffusion model is initiated. Each generation starts from standard Gaussian noise z. tStarting with the DDIM (Denoising Diffusion Implicit Model) fast sampling algorithm, iterative denoising is performed. The core idea of ​​DDIM is to treat the diffusion process as a non-Markov chain, allowing for deterministic or stochastic inferences in fewer steps than the original: its iterative formula involves estimating the current noise z... t Compared with the original target Combined with 0, and introducing a controllable random noise standard deviation σ t This achieves a reverse transition from step t to step t-1. The randomness parameter η controls the random amplitude of the sampling process; this scheme sets η=1.0 to preserve sufficient ensemble diversity, generating multiple possible paths and intensity evolutions. Finally, the predicted samples obtained from each sampling are compared with the prior predictions. prior The summation yields a priori-corrected ensemble forecast. The essence of this addition operation is to use a diffusion model to generate a residual distribution and then superimpose probability perturbations on the prior mean, so that the ensemble members both cluster around high-skill deterministic forecasts and possess reasonable dispersion.

[0030] In some embodiments, the original set generated by the diffusion model often suffers from underestimation of variance, meaning the prediction interval is too narrow, resulting in actual observations falling within the prediction interval less frequently than the nominal confidence level. To correct this bias, this scheme independently searches for the optimal scaling factor for each forecast variable (longitude, latitude, maximum wind speed) on the validation set. Specifically, for each sample in the validation set, the 25th and 75th quantiles are calculated for the N set members, and the difference is used to obtain the interval width. ; using the median of the set Centered on [0.5, 5.0], traverse the range [0.5, 5.0]. Construction interval The coverage rate of this interval to the true observations is calculated. The optimal coverage rate is determined by minimizing the absolute error between the coverage rate and the 75% target confidence level. This calibration process does not require retraining the neural network; it adjusts the empirical coverage to the nominal level simply by linearly scaling the interval width. The computational cost is extremely low, yet the effect is significant. Saving the optimal scaling factor obtained through the search and applying it to the test set yields a reliable probability prediction interval.

[0031] In some embodiments, deterministic indicators include mean absolute error (MAE) and root mean square error (RMSE): MAE reflects the average deviation between forecast and observed values ​​and is insensitive to outliers; RMSE amplifies larger errors through squaring and focuses more on extreme deviations. The probability indicator uses the Continuous Graded Probability Score (CRPS), which comprehensively measures the calibration and sharpness of ensemble forecasts: its calculation formula is the average absolute deviation between ensemble members and actual values ​​minus half the average absolute deviation among ensemble members. When the ensemble is a single-point forecast, it degenerates into absolute error; it reaches its minimum value when the ensemble is perfectly calibrated and the dispersion is appropriate. To verify the superiority of this method, RNN, LSTM, BiLSTM, GRU, BiGRU, Transformer, and quantile regression were selected as comparative models. Parallel experiments were conducted on the same dataset and partitioning strategy, comprehensively comparing wind speed and trajectory dimensions, and forecast lead times of 1, 6, and 12 hours.

[0032] In summary, this invention significantly improves the forecast accuracy of typhoon paths and intensities by employing a collaborative architecture of a deterministic prior model and a conditional diffusion model, while maintaining probabilistic generation capabilities. The dual-branch LSTM and cross-attention mechanism effectively extract the interaction features between typhoon evolution and the environmental field. The diffusion model performs residual correction based on priors, resulting in a lower ensemble mean error than conventional deep learning models. During training, a differentiable Haversine spherical distance physical constraint loss is introduced, and the prediction interval is post-calibrated on the validation set using an adaptive scaling factor. This effectively solves the variance underestimation problem of the diffusion model under moderate datasets, providing a calibrated and reliable probabilistic basis for typhoon risk warnings. Furthermore, the low diffusion step count and DDIM fast sampling strategy significantly reduce inference time, enabling stable training of the model with tens of thousands of samples. This demonstrates good engineering feasibility and business application prospects, while also enhancing the sensitivity to capturing extreme evolution scenarios such as rapid enhancement.

[0033] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0034] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A typhoon forecasting method based on an improved diffusion model, characterized in that, Includes the following steps: S1. Obtain historical typhoon best path data and meteorological reanalysis data, unify spatial resolution and construct multivariate time series, and perform Z-Score standardization. S2. Construct samples using a sliding window based on individual typhoon cases. Input the typhoon and environmental variable sequences of the past 24 hours and output the typhoon variables at future times. Divide the samples into training set, validation set and test set according to individual cases. S3. Construct a dual-branch LSTM network. After the typhoon branch and the environment branch are fused by cross-attention, the time-sharing output head generates predictions of typhoon variable points for each time period and trains the prior model. S4. Construct a conditional diffusion model. Its conditional encoder uses the same dual-branch LSTM and cross-attention structure as S3 to output the conditional context vector. The denoising network takes the noise target estimation, time step, conditional context vector and the point prediction obtained from S3 as inputs and outputs the direct prediction of the original target. It adopts cosine noise scheduling and the loss function is composed of the mean square error and the Haversine spherical distance physical constraint loss weighted together. S5. Use the S2 samples and S3 prior predictions to train the S4 diffusion model. During the training process, all parameters of the prior model are frozen, and its output is only used as static condition input and does not participate in gradient backpropagation. The Adam optimizer is used with learning rate decay and early stopping strategies to obtain the conditional diffusion model. S6. During testing, a deterministic forecast is first generated by the S3 prior model, and then the S5 conditional diffusion model is randomly sampled multiple times. The samples are generated by DDIM fast sampling and denoising, and then added to the S3 prior prediction to obtain the ensemble forecast. S7. On the S2 validation set, search for a scaling factor for each forecast variable. Based on the width of the set quantile interval, determine the optimal scaling factor by minimizing the absolute error between the coverage and the target confidence level. Apply it to the test set to obtain the calibrated prediction interval. S8. Calculate the mean absolute error, root mean square error, and continuous graded probability score of the test set to evaluate the forecast performance.

2. The typhoon forecasting method based on an improved diffusion model according to claim 1, characterized in that, In S1, the historical typhoon best track dataset includes time, longitude, latitude, maximum sustained wind speed, and minimum central pressure; the target area meteorological grid reanalysis data includes ERA5 reanalysis data of the southeastern sea region of my country with a resolution of 0.25∘×0.25∘, specifically including horizontal wind components U10 and V10 at 10m height, temperature T2m at 2m, surface pressure SP, and relative humidity RH500, temperature T500, and horizontal wind components U500 and V500 on the 500hPa isobaric surface; the Z-Score standardization formula is: ； in, The original meteorological variable values ​​to be standardized. These are the meteorological variable values ​​after Z-Score standardization. σ and σ are the mean and standard deviation of each variable on the training set, respectively. The standardized data are used for subsequent model training.

3. The typhoon forecasting method based on an improved diffusion model according to claim 1, characterized in that, In S2, the input to the sliding window is the typhoon variables and environmental field variables from the past 24 hours (24 steps), and the output is the typhoon variables for the next 1 hour, 6 hours, and 12 hours. These typhoon variables include longitude, latitude, and maximum wind speed. The training set consists of data from 2021 to 2023, the validation set consists of data from 2024, and the test set consists of data from 2025 to ensure that there is no overlap of samples of the same typhoon between different sets.

4. The typhoon forecasting method based on an improved diffusion model according to claim 1, characterized in that, In S3, the typhoon branch uses a two-layer LSTM with 64 hidden units in each layer, and the environment branch uses a two-layer LSTM with 128 hidden units in each layer. The cross-attention mechanism fuses the output sequences of the two branches to capture the interaction between the typhoon and the environment field. The loss function uses weighted mean square error, in which the latitude and longitude components are given a large weight to emphasize the location accuracy. Training employs an adaptive learning rate optimizer, along with learning rate decay and early stopping strategies; The trained model is denoted as Its output point prediction is denoted as This information is provided as prior information to subsequent diffusion models.

5. The typhoon forecasting method based on an improved diffusion model according to claim 1, characterized in that, In S4, the conditional diffusion model is the improved denoised diffusion probability model DDPM; the input of the conditional encoder is the historical observation sequence x, and the output is a fixed-length conditional context vector c. The denoising network is a multilayer perceptron, and its input includes the current noisy target estimate. The diffusion time step t, the conditional context vector c, and the complete prediction provided by the prior model. The output is the original standardized target. The direct prediction form; The total number of diffusion steps T is set to 50 steps, and cosine noise is used for scheduling; the noise addition in the forward process is based on the formula. Execution, in which , This is the noise dispatch coefficient; Denoising Network by( ( ) is the input, and the output is the predicted original target. ; The loss function is: ； in, For mean square error loss, For physical constraint loss, weights To balance the two terms; the physical constraint loss uses the Haversine spherical distance, and for each forecast time k, the true longitude is set to... latitude is The predicted longitude is latitude is All values ​​are expressed in radians, representing the Haversine distance at that time. The formula for calculation is: ； in, The average radius of the Earth is taken as 6371 km. , ; The predicted latitude of the typhoon center. For the actual latitude of the typhoon center, physical constraint loss Defined as the average Haversine distance across all forecast times, the expression is: Where H is the number of forecast times.

6. The typhoon forecasting method based on an improved diffusion model according to claim 1, characterized in that, In S5, training samples are used ( ) and the corresponding prior prediction Simultaneously, the diffusion model is trained using the Adam optimizer with an initial learning rate of... The batch size is 128; during training, the validation set loss is monitored. If there is no improvement after 10 consecutive epochs, the learning rate is halved. If there is no improvement after 20 consecutive epochs, early stopping is triggered and the optimal model weights are restored. After training, the conditional diffusion model is obtained. .

7. The typhoon forecasting method based on an improved diffusion model according to claim 1, characterized in that, In S6, for each sample x in the test set, the prior model is first used. Deterministic forecasts Then, using the conditional diffusion model Perform N random samplings, each starting with isotropic Gaussian noise, and iterate using the DDIM fast sampling algorithm. Denoising is performed to generate multiple possible forecast samples. The DDIM sampling iteration formula is: ； in, The standard Gaussian random noise term introduced in step t, =( The original target predicted by the denoising network is the standard deviation of random noise. The expression is: ； Setting randomness parameters during sampling The value range is from 0.5 to 1.5, preferably, =1.0 to stimulate diversity in the set; the generated samples are added to the prior predictions to obtain the final set.

8. The typhoon forecasting method based on an improved diffusion model according to claim 1, characterized in that, In S7, a scaling factor is independently searched for for each forecast variable on the validation set. This is used for post-calibration of the prediction interval; the specific method is as follows: calculate the 25th and 75th quantiles of the validation set, and calculate the difference between the two to obtain the interval width. ; Preliminary experiments on the validation set showed that when the scaling factor was below 0.5, the interval was too narrow and the coverage was severely low; when it was above 5.0, the interval width was excessively expanded and the sharpness was significantly deteriorated. Therefore, the search range was [0.5, 5.0]. This makes the scaled interval The coverage of the true values ​​reached the target confidence level of 75%, of which The median of the set; The search process is achieved by minimizing the absolute error between the coverage and the target value; the optimal scaling factor obtained from the search is saved and applied to the ensemble forecast of the test set to obtain the calibrated prediction interval.

9. The typhoon forecasting method based on an improved diffusion model according to claim 1, characterized in that, In S8, the deterministic indicators include Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), and the probabilistic indicators include Continuous Graded Probability Score (CRPS); the formula for CRPS is: ； in, Let be the continuous hierarchical probability score corresponding to the d-th forecast variable, and M be the total number of sample members in the ensemble forecast. For the m-th set of predicted sample values ​​of the d-th predicted variable, For the nth set of predicted sample values ​​of the d-th predicted variable, This represents the actual observed value corresponding to the d-th forecast variable; The results of the calculations are compared with those of common time series prediction models. The results are visualized and evaluated.

Images