Typhoon Wind Field Reconstruction Method, Device and Medium Based on Bayesian Model
Through the method of combining Bayesian model and feedforward neural network, the problem of incongruence of physical characteristics of typhoon wind farms is solved, and efficient and accurate wind farm risk assessment is achieved.
Patent Information
- Application Number
- CN202410928696.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2025-08-01
- Estimated Expiration
- 2044-07-11
AI Technical Summary
In the prior art, the inconsistency of physical characteristics of typhoon wind farms leads to inaccurate assessment of wind farm risks, and the instability of different data sources and the lack of interdependence of independent development models increase the uncertainty of the evaluation results.
Using a Bayesian model-based method, combined with a feedforward neural network, the weight parameters are estimated through the Bayesian hierarchical model, the wind field model is reconstructed using the typhoon optimal path dataset, and multi-source data is integrated to maintain the intrinsic consistency of the wind field parameters.
The accuracy of wind farm simulation and risk assessment is improved, the deviation of evaluation results is reduced, and efficient wind farm parameter solution and risk analysis are achieved.
Smart Images
Figure CN118898196B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind field reconstruction, and particularly to a typhoon wind field reconstruction method, device, equipment and medium based on a Bayesian model. Background Art
[0002] A tropical cyclone (hereinafter simply referred to as a typhoon), also known as a hurricane or a typhoon, is one of the natural disasters that pose a great threat to human life. A typhoon not only causes direct damage, but also leads to storm surges and waves, which may bring more disasters. Statistical data shows that the economic losses caused by typhoons in China rank first in the world every year, and the number of casualties ranks second. Therefore, it is very important to evaluate the risks brought by typhoons. Accurate and efficient estimation of typhoon hazards provides a scientific basis for risk assessment. The hazard of wind can be quantified by the probability that the wind exceeds a given speed threshold at a specific location.
[0003] There is currently a trend in the development of typhoon wind field hazard estimation methods: more and more independently developed modules and multi-source data are incorporated, including numerical model reanalysis data and remote sensing data. Although this can improve the ability to represent typhoon physical processes, it will also introduce inconsistencies between data sets, and thus the values of each parameter may lead to disharmony between the physical characteristics of the wind field. In addition, data from some sources may not be stably obtained. The needs for obtaining wind field parameters are implemented through different models, and the models of different modules are usually fitted separately without considering their mutual dependencies. These drawbacks may increase the uncertainty of the final hazard estimation results and lead to certain biases. From the user's perspective, the insurance / reinsurance or wind energy industries need to regularly incorporate newly obtained typhoon data to update their hazard assessments while maintaining the temporal consistency of the results. This requires stable data availability and computationally efficient algorithms to reduce the running time. Summary of the Invention
[0004] To overcome the problems existing in the related art, the present disclosure provides a typhoon wind field reconstruction method, device, equipment and medium based on a Bayesian model to solve the technical problem that the disharmony between the physical characteristics of the typhoon wind field in the related art leads to inaccurate typhoon wind field hazard assessment.
[0005] One or more embodiments of this specification provide a typhoon wind field reconstruction method based on a Bayesian model, including the steps of:
[0006] Obtain a typhoon best track data set, and the typhoon best track parameters include the longitude, latitude, maximum wind speed, translation speed, sine function of translation azimuth, cosine function of translation azimuth, and sine and cosine functions of the date of the typhoon center;
[0007] Construct an initial typhoon wind field model based on the modified Rankine vortex model, and convert the initial typhoon wind field model into an average wind speed model based on each quadrant;
[0008] Estimate the weight parameters through a Bayesian hierarchical model. Among them, based on the observable maximum wind radius, the data layer defines that the average wind speed and the maximum wind radius in each quadrant follow a normal distribution; the process layer defines the typhoon wind field model of the underlying physical process that generates the observable maximum wind radius, and a feedforward neural network model that generates the wind field parameters for predicting the typhoon wind field model. Its control parameter is the weight parameter w of the feedforward neural network model. The wind field parameters include the maximum wind radius, the shape parameter, the wave number - 1 asymmetry amplitude, and the initial azimuth angle of the wind field; the prior layer defines that the parameter w follows a normal distribution, and the parameter σ follows a lognormal distribution;
[0009] Obtain the posterior probability distribution of the weight parameters through the Monte Carlo method, and then determine the weight parameters through the MAP method to obtain the feedforward neural network model;
[0010] Based on the typhoon best track dataset and the feedforward neural network model, determine the wind field parameters, and determine the typhoon wind field model according to the wind field parameters.
[0011] One or more embodiments of this specification provide a typhoon wind field reconstruction device based on a Bayesian model, including:
[0012] A dataset acquisition module that acquires the typhoon best track dataset. Among them, the typhoon best track parameters include the longitude, latitude, maximum wind speed, translation speed, sine function of the translation azimuth angle, cosine function of the translation azimuth angle, and sine and cosine functions of the date;
[0013] A typhoon wind field model construction module that constructs an initial typhoon wind field model based on the modified Rankine vortex model and converts the initial typhoon wind field model into an average wind speed model based on each quadrant.
[0014] A Bayesian hierarchical model construction module that estimates the weight parameters through a Bayesian hierarchical model. Among them, based on the observable maximum wind radius, the data layer defines that the average wind speed and the maximum wind radius in each quadrant follow a normal distribution; the process layer defines the typhoon wind field model of the underlying physical process that generates the observable maximum wind radius, and a feedforward neural network model that generates the wind field parameters for predicting the typhoon wind field model. Its control parameter is the weight parameter w of the feedforward neural network model. The wind field parameters include the maximum wind radius, the shape parameter, the wave number - 1 asymmetry amplitude, and the initial azimuth angle of the wind field; the prior layer defines that the parameter w follows a normal distribution, and the parameter σ follows a lognormal distribution;
[0015] The weight parameter confirmation module obtains the posterior probability distribution of the weight parameters through the Monte Carlo method, and then determines the weight parameters through the MAP method to obtain a feedforward neural network model;
[0016] The wind field model confirmation module determines the wind field parameters based on the typhoon best track dataset and the feedforward neural network model, and determines the typhoon wind field model according to the wind field parameters.
[0017] One or more embodiments of this specification provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the typhoon wind field reconstruction method based on the Bayesian model as described above.
[0018] One or more embodiments of this specification provide a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the steps of the typhoon wind field reconstruction method based on the Bayesian model as described above.
[0019] A typhoon wind field reconstruction method, device, equipment, and medium based on the Bayesian model provided by this disclosure have the advantage that by applying machine learning algorithms and combining Bayesian hierarchical models, the problem of solving the unknown parameters of the typhoon wind field model is transformed into solving the weight parameters of the feedforward neural network model. Compared with the existing method of separately estimating the unknown wind field parameters of the wind field model through four models, it is difficult to consider the correlation and consistency between the wind field parameters. In this embodiment, the four wind field parameters are estimated through a feedforward neural network model combined with a Bayesian hierarchical model. The input variables of the feedforward neural network model include the best track data of the time, location, and movement information of the typhoon, making the model have spatio-temporal properties and maintaining the internal consistency of the characteristics of all aspects of the typhoon to the greatest extent. It solves the problem of large deviations in the hazard assessment results caused by the inconsistent data sources for typhoon simulation and hazard assessment. Brief Description of the Drawings
[0020] To more clearly illustrate the technical solutions in one or more embodiments of this specification or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in this specification. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0021] Figure 1 It is a flowchart of a typhoon wind field reconstruction method based on the Bayesian model provided by one or more embodiments of this specification;
[0022] Figure 2Block diagram of a typhoon wind field reconstruction device provided for one or more embodiments of this specification;
[0023] Figure 3 Schematic structural diagram of a computer device provided for one or more embodiments of this specification. Specific embodiments
[0024] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will clearly and completely describe the technical solutions in one or more embodiments of this specification with reference to the accompanying drawings in one or more embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this invention.
[0025] The following will make a detailed description of the present invention in combination with the specific embodiments and the drawings in the specification.
[0026] Method embodiments
[0027] According to an embodiment of the present invention, a typhoon wind field reconstruction method based on a Bayesian model is provided. As Figure 1 shown, it is the flowchart of the typhoon wind field reconstruction method based on the Bayesian model provided for this embodiment. The typhoon wind field reconstruction method based on the Bayesian model according to the embodiment of the present invention includes:
[0028] Step S1, obtain the typhoon best track data set. Among them, the typhoon best track data includes the longitude, latitude, maximum wind speed, translation speed, sine and cosine functions of the translation azimuth angle of the typhoon center, and the typhoon observation date (including the sine and cosine functions of the date); among them, the translation azimuth angle refers to the angle with the due north as 0° and clockwise with the typhoon movement direction. The sine and cosine functions of the translation azimuth angle characterize the typhoon movement direction characteristics. The sine and cosine of the date refer to the sine and cosine functions with the typhoon observation date (i.e., 365 days per year, excluding the year) as a periodic variable to characterize the seasonal characteristics of the typhoon occurrence.
[0029] Step S2, construct an initial typhoon wind field model based on the modified Rankine vortex model, and convert the initial typhoon wind field model into an average wind speed model for each quadrant;
[0030] Step S3: Estimate the weight parameters based on the Bayesian hierarchical model. Among them, based on the observable maximum wind radius, the data layer defines that the average wind speed and the maximum wind radius in each quadrant follow a normal distribution; the process layer defines the typhoon wind field model for the underlying physical process that generates the observable maximum wind radius, and a feedforward neural network model for generating the wind field parameters for predicting the typhoon wind field model. Its control parameter is the weight parameter w of the feedforward neural network model. The wind field parameters include the maximum wind radius, shape parameter, wave number - 1 asymmetry amplitude, and initial azimuth angle of the wind field; the prior layer defines the normal distribution of the parameter w and the lognormal distribution of the parameter σ.
[0031] Step S4: Obtain the posterior probability distribution of the weight parameter w through the Monte Carlo method, and then determine the weight parameter w through the MAP (Maximum A posteriori Estimation) method to obtain the feedforward neural network model.
[0032] Step S5: Determine the wind field parameters based on the typhoon best track dataset and the feedforward neural network model, and determine the typhoon wind field model according to the wind field parameters.
[0033] The method provided in this embodiment applies machine learning algorithms and combines the Bayesian hierarchical model to transform the solution of the unknown parameters of the typhoon wind field model into the solution of the weight parameter w of the feedforward neural network model. Compared with the existing method of separately estimating the unknown wind field parameters of the wind field model through four models, it is difficult to consider the correlation and consistency between the wind field parameters (that is, the values of each parameter may lead to the incoordination between the physical characteristics of the wind field). The method in this embodiment estimates the four wind field parameters through the feedforward neural network model combined with the Bayesian hierarchical model. The input variables of the feedforward neural network model include the best track data of the typhoon's time, position, and movement information, making the model have spatio-temporal properties and maintaining the internal consistency of the typhoon's various characteristics to the greatest extent. It solves the problem of large deviation in the hazard assessment results caused by the non-uniform data source of typhoon simulation and hazard assessment.
[0034] The method in this embodiment combines a neural network within the Bayesian hierarchical modeling framework, which can adapt to various wind field models. All unknown parameters are finally represented by a set of weight parameters of the feedforward neural network model, making the solution process of the wind field parameters more efficient and intuitive. With the help of modern machine learning software, the entire process from model development to hazard analysis is highly computationally efficient and highly automated.
[0035] In this embodiment, a typhoon full-path simulation algorithm is constructed using the typhoon best track set data and a main model, generating a sufficient amount of synthetic typhoon data. The technical solution mainly refers to "Stochastic Simulation of Tropical Cyclones for Risk Assessment at One Go: A Multivariate Functional PCA Approach" published by Yang Chi et al. in the 8th issue of the journal "Earth and Space Science" in 2021. This method proposes a new model for risk assessment of full-path simulation. Through machine learning methods: multivariate functional principal component analysis, high-performance synthetic typhoons can be generated in a fully automated manner, with almost no manual intervention, which greatly improves the modeling efficiency and reduces the running time, especially when newly available typhoon data is regularly incorporated into the model.
[0036] In this embodiment, in step S2, an initial typhoon wind field model is constructed based on the modified Rankine vortex model, specifically as follows.
[0037] The initial typhoon wind field model in this embodiment uses a modified Rankine vortex model with wavenumber-1 asymmetry:
[0038]
[0039]
[0040] where V R (r,θ) is the tangential wind speed as a function of the wind field radius r and the wind field azimuth angle θ. For the development of this model, a relative coordinate system of typhoon movement is used, setting the right direction of the typhoon movement vector by 90° as the 0° direction of the wind field, and measuring the wind field azimuth angle counterclockwise.
[0041] The formula contains known parameters (V m ) and four unknown parameters (R m , x, a, and θ0). Among them, V m is the maximum wind speed (MWS), R m is the radius of maximum wind (RMW), x is a shape parameter taking values in the interval [0,1], a is the wavenumber-1 asymmetry amplitude, and θ0 is the initial azimuth angle of the wind field. These four free parameters are usually linearly regressed to the maximum wind speed V m , the typhoon center position, and the translation speed (SPD).
[0042] In this embodiment, the weight parameter w of the neural network can be estimated by minimizing the error function based on the dataset of input and output variables. However, for the initial feedforward neural network model that predicts the wind field parameters of the typhoon wind field model in this embodiment, this process cannot be directly carried out because there is only one output variable R m There are observed values (i.e., RMW), and the other three are unobservable (latent variables). In addition, as can be seen from Equation (1), calculating V R (r,θ) requires the known wind field radius r and the wind field azimuth angle θ, and the θ direction is not provided in the best track dataset, so Equation (1) cannot be directly used. One way to solve this problem is to examine the average wind speed in each quadrant instead of V R (r,θ). Therefore, the initial typhoon wind field model is converted into an average wind speed model based on each quadrant, as follows:
[0043]
[0044] In Equation (2), θ1 and θ2 are the lower and upper bounds of the wind field azimuth angle of a quadrant respectively. Except for the shape parameter x, the wavenumber -1 asymmetry amplitude a, and the initial wind field azimuth angle θ0, there are no other unobserved parameters in Equation (2), so it can be used for neural network parameter estimation.
[0045] In this embodiment, in step S3, the method of embedding the neural network in the Bayesian hierarchical model transforms the parameter solving process of the initial typhoon wind field model into the solving process of the neural network weight w, solving the problem of large deviation in the risk assessment results caused by the non-uniform data sources of typhoon simulation and risk assessment. In this embodiment, a multi-output feedforward neural network model is used to predict four parameters (R m , x, a, and θ0). The feedforward neural network is the basic form of the neural network and consists of multiple layers of units. The first layer is the input layer, the last layer is the output layer, and the intermediate layers are called hidden layers. A fully connected feedforward neural network has unidirectional connections from the units of each layer to all the units of the subsequent layer. At each unit, the output is the weighted sum of the inputs from the previous layer plus a constant (bias), and then transformed by a non-linear function called the activation function:
[0046]
[0047] where, is the output of the k-th unit in the l-th layer, n l-1 is the number of neurons in the (l - 1)-th layer, is the weight of the connection between the j-th unit in the (l - 1)-th layer and the k-th unit in the l-th layer, b l is the bias of the l-th layer, is an activation function. For the input layer (l = 1), and b 1 = 0. The activation function of the hidden units is usually chosen as a sigmoid function, such as the "logistic" or "tanh" function. The activation of the output units (l = L) is determined by the nature of the data and the distribution of the target variables set, to give a set of network outputs This results in a series of function transformations. By setting b l = 1 for l > 1, the bias parameter can be absorbed into a set of weight parameters. Therefore, the neural network model can be simply regarded as a non-linear function, from a set of input variables {x i} to a set of output variables {y k}, controlled by a set of weight vectors w.
[0048] This embodiment uses an initial feedforward neural network model with two hidden layers to represent four wind field parameters. For simplicity, each hidden layer has the same number K of units, and will be optimized through test data. The training data is the typhoon best track dataset. The input vector x consists of 8 variables: x = (LON, LAT, MWS, SPD, COSB, SINB, COSD, SIND), that is, longitude, latitude, maximum wind speed, translation speed, cosine and sine of the translation azimuth angle, and cosine and sine of the date. The activation function of the hidden units is chosen as the tanh function:
[0049]
[0050] For the output layer of the initial feedforward neural network model, specify y1 = R m , y2 = x, y3 = a, and y4 = θ0. In this embodiment, for simplicity, the layer index is removed; according to the nature and range of the output parameters, the activation function is chosen as: is the standard normal cumulative distribution function that maps real numbers to [0, 1], maps real numbers to the range [-180, 180].
[0051] In this embodiment, referring to Equation 2, the process of estimating the weight parameter w involves a probability, which can be implemented by a Bayesian model. Generally, given an input, the evaluation probability model allows sampling of the output corresponding to the input (e.g., probabilistically sampling), and / or providing a probability distribution for the output corresponding to the input. The input is typically a vector including one or more values. In some embodiments, the output is a value; in other embodiments, the output includes multiple values. The probability model typically includes one or more random variables, such as latent variables and / or output variables. The latent variables can be defined by a probability distribution or by a random process.
[0052] In this embodiment, the neural network model is embedded as a latent process into the Bayesian hierarchical model and is constrained by its parent process, i.e., the modified Rankine vortex model. The parameter w can be estimated using Bayesian inference algorithms. Throughout the research process, from parameter estimation to wind field reconstruction, the Bayesian model averaging (BMA) technique based on the maximum a posteriori estimation ensemble of the weight parameter w is adopted. This wind field reconstruction method constitutes a two-stage method for typhoon wind field hazard analysis based solely on the best track dataset (one stage being the typhoon full path simulation method for the typhoon wind field of the above-mentioned best track dataset). Each of the two stages only contains one main model, and the estimation has better generalization performance compared to estimating each parameter through several independently developed modules. Therefore, the intrinsic consistency among various aspects of the typhoon can be maximally maintained.
[0053] In a specific embodiment, in step S3, the weight parameter w is estimated based on the Bayesian hierarchical model, and the Bayesian hierarchical model is specifically as follows.
[0054] Bayesian hierarchical modeling is based on Bayes' rule:
[0055]
[0056] where θ = {θ j} is a set of parameters, and y = {y i} is a set of observation results. P(θ) is the prior distribution of θ. When y is given, the sampling distribution P(y|θ) is also called the likelihood (function). P(y) is called the evidence. Therefore, P(θ|y) is the posterior distribution of θ given y. Given a new input x, the predictive distribution of the output is calculated as:
[0057]
[0058] Equation (6) is called Bayesian model averaging (BMA), which performs a weighted average over all possible models (distinguished by different settings of the parameter θ). For the typhoon wind field, the Bayesian hierarchical model can be defined as three layers:
[0059] 1) Data layer:
[0060]
[0061] R m ~N(y1,σ2)(8);
[0062] 2) Process layer:
[0063]
[0064] [y1,y2,y3,y4] = Φ(x; w) (10);
[0065] 3) Prior layer:
[0066] w ~ N(0, 10) (11);
[0067] σ ~ LogN(0, 1) (12);
[0068] In equations (7)-(12), the subscript indices of the input and output data and the subscript indices of the vector elements w and σ are omitted in this embodiment.
[0069] The data layer of the Bayesian model defines the distribution of the model observables including the mean wind speed in each quadrant and the distribution of R m and they are both normally distributed.
[0070] The distribution parameters of the normal distribution are the mean μ and the standard deviation σ, where the mean μ is determined by equation (9) of the process layer, and the standard deviation σ is defined by equation (12) of the prior layer. The process layer defines the underlying physical process for generating the observable values: including the typhoon wind field model related to four wind field parameters (refer to equation 9), and the feedforward neural network model for generating the wind field parameters (refer to equation 10). The control parameter of the feedforward neural network model is the weight w.
[0071] The prior layer defines the prior distributions of w and σ: assuming that each element of the vector is independent and identically distributed, w and σ follow a normal distribution and a lognormal distribution respectively.
[0072] Based on the established Bayesian hierarchical model above, in this embodiment, the posterior probability distribution of the weight parameter w is obtained by the Monte Carlo method, and then the weight parameter w is determined by the MAP method to obtain the feedforward neural network model; finally, based on the typhoon best track dataset and the feedforward neural network model, the wind field parameters are determined, and the typhoon wind field model is determined according to the wind field parameters.
[0073] In one embodiment, when using the typhoon wind field model obtained above for typhoon strong wind hazard assessment at a fixed location, it is the vector addition of the tangential wind speed V R corresponding to the fixed location and the typhoon translation speed V T as follows:
[0074]
[0075] In this embodiment, the purpose of hierarchical modeling is to infer the weight parameter w of interest through its posterior distribution. The point estimate of the weight parameter w can be achieved by maximizing the posterior distribution of the weight parameter w, which is called the maximum a posteriori probability estimation (MAP). However, the posterior distribution of the weight parameter w is usually multimodal: there are many local maxima on the surface of the distribution function. If the set of MAP estimates of the weight parameter w is available, then the ensemble average of the wind fields reconstructed by each weight parameter w is actually the BMA prediction. Using modern machine learning software, such as the TensorFlow deep learning framework, the estimate of each weight parameter w can be obtained through a stochastic optimization method.
[0076] In the method provided in this embodiment, the Bayesian model has a complete theoretical framework and a complete set of calculation methods. For specific details, reference needs to be made to the existing relevant technical content. When the Bayesian model infers unknown parameters, it is first necessary to obtain its posterior probability distribution, which is mainly obtained through the Monte Carlo method. Then, the mean or maximum value of the posterior distribution is used as the inference of the unknown parameter. The unknown parameter to be inferred here is the neural network weight w, and the sample is the observable quantity, and the source is the typhoon best track IBTrACS dataset. This is the entire set of methods and the only dataset used in the first-stage model and the existing model. Since the neural network does not directly train to obtain the weight w, but first uses the Bayesian method to obtain the posterior distribution of the weight w, and then obtains its estimated value. After obtaining the estimated value of the weight w, the neural network model is determined. At this time, when the variable values of the input vector x (LON, LAT, MWS, SPD, COSB, SINB, COSD, SIND) are input, the values of four wind field parameters can be output, and thus the typhoon wind field can be constructed according to Equation (1). These input quantities are in the IBTrACS dataset. The input quantities simulating the typhoon path are input into the determined neural network model, and the corresponding 4 simulated typhoon wind field parameters are output, and then the wind field on the simulated typhoon path is reconstructed. Reconstructing the wind fields of all long-sequence typhoon simulated paths (more than ten thousand years) can be used for typhoon wind field hazard analysis, obtaining the distribution of the recurrence period of strong typhoon winds on the ocean surface (such as once in a hundred years, etc.), and further used for offshore engineering risk assessment.
[0077] The advantages of this embodiment are illustrated below through specific comparison cases.
[0078] To verify the rationality of this method, this case analyzed the once-in-a-hundred-year wind speed V for 10 coastal cities affected by typhoons in China 100(Unit: m s-1), and compared with the existing research results (Table 1, the existing research results are cited from the literature). As can be seen from Table 1, except that Xiao et al. (2010) may be higher than all other results due to systematic deviation, the results of this study are within a reasonable range compared with other results, but show more obvious spatial differences than other results, indicating that this method can better describe the spatial variation of typhoon wind field hazards in the Northwest Pacific Ocean and its impact on coastal cities in China.
[0079] Table 1. Typhoon wind speed V with a return period of 100 years in 10 coastal cities 100 Comparison of analysis results
[0080]
[0081] a Recommended industrial standard of the People's Republic of China, "Code for Wind Resistance Design of Highway Bridges" (JTG / T D60-01-2004). b National standard of the People's Republic of China, "Code for Loads on Building Structures" (GB 50009-2012).
[0082] c This method is the typhoon wind field reconstruction method based on the Bayesian model proposed in this patent.
[0083] Except for this method, all other data and literature are cited from: Fang, G, Zhao, L, Cao, S., Zhu, L, & Ge, Y. (2020). Estimation of tropical cyclone wind hazards in coastal regions of China (Estimation of tropical cyclone wind hazards in coastal regions of China). Natural Hazards and Earth System Sciences (Natural Hazards and Earth System Sciences), 20, 1617-1637. https: / / doi.org / / 10.5194 / nhess-20-1617-2020.
[0084] Device embodiment
[0085] According to an embodiment of the present invention, a typhoon wind field reconstruction device based on a Bayesian model is provided, as Figure 2 shown, which is a block diagram of the typhoon wind field reconstruction device based on the Bayesian model provided in this embodiment. The typhoon wind field reconstruction device based on the Bayesian model according to the embodiment of the present invention includes:
[0086] A data set acquisition module 10, which acquires a typhoon best track data set, where the typhoon best track includes longitude, latitude, maximum wind speed, translation speed, sine and cosine of translation azimuth angle, and sine and cosine of date.
[0087] The typhoon wind field model construction module 20 constructs an initial typhoon wind field model based on the modified Rankine vortex model, and converts the initial typhoon wind field model into an average wind speed model based on each quadrant.
[0088] The Bayesian hierarchical model construction module 30 estimates the weight parameters based on the Bayesian hierarchical model. Among them, based on the observable maximum wind radius, the data layer defines that the average wind speed and the maximum wind radius on each quadrant follow a normal distribution; the process layer defines the typhoon wind field model of the underlying physical process that generates the observable maximum wind radius, and the initial feedforward neural network model used to predict the wind field parameters of the typhoon wind field model. Its control parameter is the weight parameter w of the feedforward neural network model. The wind field parameters include the maximum wind radius, the shape parameter, the wave number - 1 asymmetry amplitude, and the initial azimuth angle of the wind field; the prior layer defines that the parameters follow a normal distribution and the parameter σ follows a lognormal distribution.
[0089] The weight parameter confirmation module 40 obtains the posterior probability distribution of the weight parameter w through the Monte Carlo method, and then determines the weight parameter w through the MAP (Maximum a posteriori estimation) method to obtain the feedforward neural network model.
[0090] The wind field model confirmation module 50 determines the wind field parameters based on the typhoon best track data set and the feedforward neural network model, and determines the typhoon wind field model according to the wind field parameters.
[0091] The device provided in this embodiment applies a machine learning algorithm and combines a Bayesian hierarchical model to convert the solution of the unknown parameters of the typhoon wind field model into the solution of the weight parameters of the feedforward neural network model. Compared with the existing method of separately estimating the unknown wind field parameters of the wind field model through four models, it is difficult to consider the correlation and consistency between the wind field parameters (that is, the values of each parameter may lead to the incoordination of the physical characteristics of the wind field). The method of this embodiment estimates the four wind field parameters through the feedforward neural network model combined with the Bayesian hierarchical model. The input variables of the feedforward neural network model include the best track data of the time, position, and movement information of the typhoon, making the model have spatio-temporal properties and maintaining the internal consistency of the characteristics of all aspects of the typhoon to the greatest extent. It solves the problem of large deviation in the risk assessment results caused by the non-uniform data sources of typhoon simulation and risk assessment.
[0092] In this embodiment, the typhoon wind field model construction module 20 constructs an initial typhoon wind field model based on the modified Rankine vortex model, as follows.
[0093] The initial typhoon wind field model in this embodiment uses a modified Rankine vortex model with wave number - 1 asymmetry:
[0094]
[0095]
[0096] Among them, V R (r, θ) is the tangential wind speed as a function of the wind field radius r and the wind field azimuth angle θ. For the development of this model, a relative coordinate system of typhoon movement is used. The right direction of the typhoon movement vector by 90° is set as the 0° direction of the wind field, and the wind field azimuth angle is measured counterclockwise.
[0097] The initial typhoon wind field model is converted into an average wind speed model based on each quadrant as follows:
[0098]
[0099] In Equation (15), θ1 and θ2 are respectively the lower bound and the upper bound of the wind field azimuth angle of a quadrant. Except for the shape parameter x, the wave number -1 asymmetry amplitude a, and the initial wind field azimuth angle θ0, there are no other unobserved parameters in Equation (15), so it can be used for neural network parameter estimation.
[0100] In this embodiment, an initial feedforward neural network model with two hidden layers is used to represent four wind field parameters. Each hidden layer has the same number of K units and will be optimized through test data. The training data is the typhoon best track dataset, and the input vector x consists of 8 variables: x = (LON, LAT, MWS, SPD, COSB, SINB, COSD, SIND).
[0101] The activation function of the hidden unit is selected as the tanh function:
[0102]
[0103] For the output layer of the initial feedforward neural network model, y1 = R m , y2 = x, y3 = a, and y4 = θ0;
[0104] The activation function is selected as: The standard normal cumulative distribution function that maps real numbers to [0, 1], maps real numbers to the range [-180, 180].
[0105] In this embodiment, the neural network model is embedded as a potential process into the Bayesian hierarchical model and is constrained by its parent process, that is, the modified Rankine vortex model. The weight parameter w can be estimated using the Bayesian inference algorithm.
[0106] In this embodiment, the Bayesian hierarchical model construction module 30 is defined as three levels:
[0107] 1) Data layer:
[0108]
[0109] R m ~N(y1,σ2)(18); 2) Process layer:
[0110]
[0111] [y1,y2,y3,y4]=Φ(x;w) (20); 3) Prior layer:
[0112] w~N(0,10)(21);
[0113] σ~LogN(0,1) (22);
[0114] In equations (17)-(22), the subscript indices of the input and output data and the subscript indices of the vector elements w and σ are omitted in this embodiment.
[0115] The data layer of the Bayesian model defines the distribution of the model observables i.e., the average wind speed on each quadrant and R m The distributions of both are normal distributions, with the mean μ defined by the process layer and the standard deviation σ defined by the prior layer.
[0116] The process layer defines the underlying physical process for generating observable values, including the typhoon wind field model (refer to equation 9) related to four wind field parameters, and the feedforward neural network model for generating wind field parameters (refer to equation 10). The control parameter of the feedforward neural network model is the weight w.
[0117] The prior layer defines the prior distributions of the parameters w and σ: Assuming that each element of the vector is independent and identically distributed, it is defined that the parameter w follows a normal distribution and the parameter σ follows a lognormal distribution.
[0118] The embodiment of the present invention is the corresponding device embodiment to the above method embodiment. The specific operations of each module processing step can be understood by referring to the description of the method embodiment, and will not be elaborated here.
[0119] As Figure 3 shown, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the typhoon wind field reconstruction method based on the Bayesian model in the above embodiment, or when the computer program is executed by a processor, it implements the typhoon wind field reconstruction method based on the Bayesian model in the above embodiment.
[0120] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in this application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.
[0121] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. The key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the method embodiments. The device and system embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.
[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and the content not described in detail in the specification of the present invention belongs to the well-known technology in the art.
Claims
1. A typhoon wind field reconstruction method based on a Bayesian model, characterized in that, Including the steps: Obtain the typhoon best track dataset, where the typhoon best track parameters include the longitude, latitude, maximum wind speed, translation speed, sine function of the translation azimuth angle, cosine function of the translation azimuth angle, and sine and cosine functions of the date; Construct an initial typhoon wind field model based on the modified Rankine vortex model, and convert the initial typhoon wind field model into an average wind speed model based on each quadrant; Estimate the weight parameters through a Bayesian hierarchical model. Among them, based on the observable maximum wind radius, it is defined at the data layer that the average wind speed and the maximum wind radius on each quadrant follow a normal distribution; at the process layer, a typhoon wind field model for the underlying physical process that generates the observable maximum wind radius and a feedforward neural network model for generating the wind field parameters of the predicted typhoon wind field model are defined. Its control parameter is the weight parameter of the feedforward neural network model, and the wind field parameters include the maximum wind radius, shape parameter, wave number - 1 asymmetry amplitude, and initial wind field azimuth angle; at the prior layer, it is defined that the parameter w follows a normal distribution and the parameter σ follows a lognormal distribution; Obtain the posterior probability distribution of the weight parameters through the Monte Carlo method, and then determine the weight parameters through the MAP method to obtain the feedforward neural network model; Based on the typhoon best track dataset and the feedforward neural network model, determine the wind field parameters, and determine the typhoon wind field model according to the wind field parameters.
2. The typhoon wind field reconstruction method based on the Bayesian model according to claim 1, wherein The initial typhoon wind field model uses a modified Rankine vortex model with wave number - 1 asymmetry. The specific form of the initial typhoon wind field model is as follows: Among them, V R (r,θ) is the tangential wind speed as a function of the wind field radius r and the wind field azimuth angle θ, and V m is the maximum wind speed, R m is the radius of the maximum wind, x is the shape parameter taking values in the interval [0,1], a is the wave number -1 asymmetry amplitude, and θ0 is the initial azimuth angle of the wind field; among them, the right direction of the typhoon movement vector by 90° is set as the 0° direction of the wind field, and the wind field azimuth angle is measured counterclockwise; Convert the initial typhoon wind field model into an average wind speed model based on each quadrant, as shown in the following formula: In the formula, θ1 and θ2 are respectively the lower and upper bounds of the wind field azimuth angle of the quadrant.
3. The typhoon wind field reconstruction method based on the Bayesian model according to claim 2, wherein, The initial feedforward neural network model is set with two hidden layers, and each hidden layer has the same number K of units; The activation function of the hidden unit is selected as the tanh function: The output layer of the initial feedforward neural network model is \(y_1 = R\) m , \(y_2 = x\), \(y_3 = a\) and \(y_4 = \theta_0\), \(R\) m is the maximum wind radius, \(x\) is the shape parameter taking values in the interval \([0, 1]\), \(a\) is the wavenumber - 1 asymmetry amplitude, and \(\theta_0\) is the initial azimuth angle of the wind field; The activation function is: The standard normal cumulative distribution function that maps real numbers to [0, 1], maps real numbers to the range [-180, 180], is the activation function of the k-th unit in the l-th layer, where k = 1, 2, 3, 4; The input is the typhoon best track dataset.
4. The typhoon wind field reconstruction method based on the Bayesian model according to claim 3, characterized in that, The Bayesian hierarchical model is defined as the following three layers: The data layer defines the average wind speed on each quadrant. and R m is normally distributed with a mean μ defined by the process layer and a standard deviation σ defined by the prior layer; The process layer defines the underlying physical process that generates the observable values: including a typhoon wind field model related to four wind field parameters and a feedforward neural network model for generating the wind field parameters. The control parameter of the feedforward neural network model is the weight parameter w; The prior layer defines the prior distributions of the parameters w and σ. The parameter w follows a normal distribution and the parameter σ follows a lognormal distribution.
5. A typhoon wind field reconstruction device based on a Bayesian model, characterized in that, Including: A dataset acquisition module that obtains the typhoon best track dataset, where the typhoon best track parameters include the longitude, latitude, maximum wind speed, translation speed, sine function of the translation azimuth angle, cosine function of the translation azimuth angle, and sine and cosine functions of the date; A typhoon wind field model construction module that constructs an initial typhoon wind field model based on the modified Rankine vortex model and converts the initial typhoon wind field model into an average wind speed model based on each quadrant; The Bayesian hierarchical model construction module estimates the weight parameters based on the Bayesian hierarchical model. Among them, based on the observable maximum wind radius, the data layer defines that the average wind speed and the maximum wind radius in each quadrant follow a normal distribution; the process layer defines the typhoon wind field model of the underlying physical process that generates the observable maximum wind radius, and a feed-forward neural network model for generating the wind field parameters of the predicted typhoon wind field model. Its control parameter is the weight parameter w of the feed-forward neural network model. The wind field parameters include the maximum wind radius, shape parameter, wave number -1 asymmetry amplitude, and initial wind field azimuth angle; the prior layer defines that the parameter w follows a normal distribution, and the parameter σ follows a lognormal distribution. The weight parameter confirmation module obtains the posterior probability distribution of the weight parameters through the Monte Carlo method, and then determines the weight parameters through the MAP method to obtain the feed-forward neural network model. The wind field model confirmation module determines the wind field parameters based on the typhoon best track dataset and the feed-forward neural network model, and determines the typhoon wind field model according to the wind field parameters.
6. The typhoon wind field reconstruction device based on the Bayesian model according to claim 5, characterized in that The typhoon wind field model construction module constructs the initial typhoon wind field model using the modified Rankine vortex model with wave number -1 asymmetry. The initial typhoon wind field model is as follows: Among them, V R (r,θ) is the tangential wind speed as a function of the wind field radius r and the wind field azimuth angle θ, and V m is the maximum wind speed, R m is the maximum wind radius, x is the shape parameter taking values in the range [0,1], a is the wave number -1 asymmetry amplitude, and θ0 is the initial azimuth angle of the wind field; among them, the right direction of the typhoon movement vector by 90° is set as the 0° direction of the wind field, and the wind field azimuth angle is measured counterclockwise; Convert the initial typhoon wind field model into an average wind speed model based on each quadrant, as follows: In the formula, θ1 and θ2 are the lower and upper bounds of the wind field azimuth angle of the quadrant, respectively.
7. The typhoon wind field reconstruction device based on the Bayesian model according to claim 6, wherein The feed-forward neural network model constructed in the Bayesian hierarchical model construction module is set with two hidden layers, and each hidden layer has the same number of K units. The activation function of the hidden unit is selected as the tanh function: The output layer of the initial feedforward neural network model is \(y_1 = R\) m , \(y_2 = x\), \(y_3 = a\) and \(y_4 = \theta_0\), where \(R\) m is the maximum wind radius, \(x\) is the shape parameter taking values in the interval \([0, 1]\), \(a\) is the wavenumber - 1 asymmetry amplitude, and \(\theta_0\) is the initial azimuth angle of the wind field; The activation function is: The standard normal cumulative distribution function that maps real numbers to [0, 1], maps real numbers to the range [-180, 180], is the activation function of the k-th unit in the l-th layer, where k = 1, 2, 3, 4; The input is the typhoon best track dataset.
8. The typhoon wind field reconstruction device based on the Bayesian model according to claim 7, characterized in that, The Bayesian hierarchical model is defined as the following three layers: The data layer defines the average wind speed on each quadrant and R m is normally distributed with the mean μ defined by the process layer and the standard deviation σ defined by the prior layer; The process layer defines the underlying physical process that generates the observable values: including the typhoon wind field model related to four wind field parameters, and a feed-forward neural network model for generating wind field parameters. The control parameter of the feed-forward neural network model is the weight w. The prior layer defines the prior distributions of the parameters w and σ. The parameter w follows a normal distribution and the parameter σ follows a lognormal distribution.
9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the Bayesian model-based typhoon wind field reconstruction method according to any one of claims 1 to 4.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the Bayesian model-based typhoon wind field reconstruction method according to any one of claims 1 to 4.