Ship air flow field early warning method based on shipborne wind profile radar data

Abstract

The invention discloses an over-ship flow field early warning method based on shipborne wind profile radar data, and the method comprises the following steps: obtaining wind profile radar data, and carrying out the two-dimensional convolution of the wind profile radar data through a CNN (convolutional neural network), and forming wind profile radar two-dimensional data with height features reserved; constructing an LSTM deep learning network, training the LSTM deep learning network by using the two-dimensional historical data of the wind profile radar, and optimizing the LSTM deep learning network to obtain a trained LSTM deep learning network; using the trained LSTM deep learning network to predict wind profile radar data at future moments, and recovering two-dimensional data obtained through prediction into three-dimensional wind field data with height information through a deconvolution method; calculating a vertical wind shear value and a horizontal wind shear value through the obtained three-dimensional wind field data, and judging whether to carry out early warning on the flow field above the ship or not through the values; according to the method, the operation efficiency is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of ship navigation technology, and in particular relates to a method for early warning of airflow over ships based on shipborne wind profiler radar data. Background Technology

[0002] With the continuous development of the marine exploration industry and artificial intelligence technology, intelligentization and automation will become the main development directions for marine environmental element observation and early warning. Research vessels are usually equipped with a variety of meteorological observation equipment, including wind profiler radar, radiosondes, and microwave radiometers, to obtain comprehensive marine meteorological information. With the development of marine meteorological observation capabilities, wind profiler radar is one of the important detection tools. In marine meteorological observation and research, wind profiler radar can help ships obtain accurate mid- and low-level wind field data, providing important support for weather forecasting and climate research.

[0003] Since wind shear is a meteorological phenomenon related to rapid changes in wind speed and direction, it can easily pose a danger to ship navigation. Shipborne wind profiler radar can acquire continuous wind speed and direction data, and capture the nonlinear variation law of wind speed and direction through deep learning models, enabling wind shear prediction in the short term and providing effective early warning information. Summary of the Invention

[0004] The present invention adopts the following technical solution:

[0005] A method for early warning of airspace over ships based on shipborne wind profiler radar data, the method includes the following steps:

[0006] Acquire wind profiler radar data and perform two-dimensional convolution on the wind profiler radar data through a CNN convolutional neural network to form two-dimensional wind profiler radar data that retains height features;

[0007] An LSTM deep learning network was constructed and trained using two-dimensional historical data from wind profiler radar. The LSTM deep learning network was then optimized using the Adam method to obtain the trained LSTM deep learning network.

[0008] The trained LSTM deep learning network is used to predict wind profile radar data at future moments, and the predicted two-dimensional data is restored into three-dimensional wind field data with height information by deconvolution.

[0009] By acquiring three-dimensional wind field data, the vertical wind shear value and the horizontal wind shear value are calculated, and the value is used to determine whether to issue a flow field warning for ships.

[0010] Furthermore, the method for optimizing LSTM deep learning networks using the Adam method is as follows:

[0011] First, determine the loss function L;

[0012] The exponential moving average of the gradient direction of the loss function is calculated using a momentum model. Noise is smoothed by modifying the optimization parameters, thus accelerating loss convergence. The method is as follows:

[0013] The optimization parameter is θ, and the loss gradient is... The momentum equation is then:

[0014] m t =β1·m t-1 +(1-β1)·g t

[0015] Where: m t Let β1 be the momentum value at step t, initially set to 0; β1 is the momentum decay rate, which controls the weight of the historical gradient contribution. In this invention, β1 = 0.9.

[0016] The adaptive learning rate is calculated by estimating the variance of the error gradient by adjusting the learning rate parameters, as follows:

[0017] v t =β2·v t-1 +(1-β2)·g t 2

[0018] In the formula: v t β1 is the exponential moving average of the gradient direction at step t, initially set to 0; β2 is the decay rate, which is 0.999 in this invention.

[0019] Because m t and v t Initially set to 0, the model will be biased towards 0 during the initial training phase. Correction is needed to eliminate the influence of this initial bias on the model and make the estimated values ​​closer to the true distribution. The correction formula is as follows:

[0020]

[0021] In the formula: The direction of momentum is indicated by the value; a larger value indicates a more pronounced convergence phenomenon. For adaptive learning rate, A larger value indicates that the gradient direction oscillation can be effectively and consistently maintained.

[0022] pass and The optimization parameter is calculated as θ. t+1 The formula is as follows:

[0023]

[0024] In the formula: η is the initial learning rate, which is 0.001 in this invention; ∈ is the minimum value to prevent the denominator from being 0, which is 1 × 10 in this invention. -8 ;

[0025] Ultimately, θ t+1 It is applied to the input gate weights of LSTM deep learning networks to optimize LSTM deep learning networks.

[0026] Furthermore, the method for recovering the predicted two-dimensional wind field data with height information using deconvolution is as follows:

[0027] The input two-dimensional data in deconvolution recovery is represented as follows:

[0028]

[0029] In the formula:

[0030] F out Two-dimensional data predicted by a trained LSTM deep learning network;

[0031] H′ and W′ represent the number of horizontal and vertical grid points of the deconvolution data;

[0032] The deconvolution kernel is represented as follows:

[0033]

[0034] In the formula:

[0035] W (l) It is a deconvolution kernel;

[0036] K d K h K w These are the indexes of the data in the height, horizontal, and vertical directions, respectively.

[0037] The deconvolution formula is as follows:

[0038]

[0039] In the formula:

[0040] d′, h′, w′ are the height, horizontal, and vertical indices of the output 3D data;

[0041] c, c′ are the input channel indices;

[0042] s d s h s w s is the step size in this invention. d s h s w =1.

[0043] Furthermore, the method for calculating the vertical wind shear and horizontal wind shear values ​​using the acquired three-dimensional wind field data is as follows: Vertical wind shear value C z The calculation formula is as follows:

[0044]

[0045] In the formula:

[0046] V x (h1), V x (h2) represents the wind component in the x direction at heights h1 and h2, respectively;

[0047] V y (h1), V y (h2) represents the wind component in the y direction at heights h1 and h2, respectively;

[0048] Horizontal wind shear represents the difference in wind speed or direction between two points at the same altitude. It is the composite value of horizontal wind shear in the x and y directions, and the formula is as follows:

[0049] Calculate the horizontal wind shear C in the x-direction. x :

[0050]

[0051] In the formula:

[0052] V y1 (h), V y2 (h) represents the wind component in the y direction at two adjacent moments at the same altitude;

[0053] V z1 (h), V z2 (h) represent the wind components in the z-direction at two adjacent moments at the same altitude;

[0054] Calculate the horizontal wind shear C in the y-direction. y :

[0055]

[0056] In the formula:

[0057] V x1 (h), V x2 (h) represents the wind component in the x direction at two adjacent moments at the same altitude;

[0058] The horizontal wind shear in the x and y directions is combined to obtain the horizontal wind shear value C measured by the wind profiler radar. xy :

[0059]

[0060] Furthermore, the method for determining whether to issue a flow field warning for ships using numerical data is as follows:

[0061] Based on the vertical and horizontal wind shear values, wind shear intensity is divided into four levels: mild, moderate, severe, and serious. An early warning is issued when the shear intensity reaches moderate or above.

[0062] Furthermore, the determined loss function L is:

[0063]

[0064] In the formula:

[0065] x i This represents the actual wind speed.

[0066] To forecast wind speed;

[0067] N is the number of samples.

[0068] The beneficial technical effects of this invention are as follows:

[0069] The method of this invention has significant advantages and differences compared to existing technologies: First, in terms of computational efficiency, existing small- and medium-scale wind field forecasting methods generally rely on numerical models, requiring complex equation integration and large-scale iterative calculations, which are time-consuming and difficult to meet real-time requirements. In contrast, this invention, through the combination of convolutional neural networks (CNN) and long short-term memory networks (LSTM), only needs to process single-point data from shipborne wind profiler radar to quickly complete the forecast, significantly improving computational efficiency. Second, in terms of data dependence, traditional methods usually require the introduction of a large range of background field data (temperature, humidity, air pressure, etc.) as support; otherwise, accurate results are difficult to obtain. This invention, however, relies entirely on time-series data acquired by shipborne radar to achieve forecasting, reducing dependence on external data and better matching actual navigation scenarios. Third, in terms of real-time performance and applicability, this invention can achieve rolling forecasts with a minute-level sampling period, quickly outputting vertical and horizontal wind shear values, possessing the forecast characteristics of "short, flat, and fast," and can provide immediate early warning information, thereby effectively improving the safety assurance capability of ship navigation, demonstrating stronger practicality and innovation than existing technologies. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the rolling forecast provided in an embodiment of the present invention;

[0071] Figure 2 A flowchart illustrating a method for early warning of airspace over ships based on shipborne wind profiler radar data, provided in an embodiment of the present invention. Detailed Implementation

[0072] The following, in conjunction with the accompanying drawings, provides a more clear and complete description of a flexible electricity capacity market clearing method and apparatus provided by the present invention:

[0073] like Figure 2 As shown, this embodiment provides a method for early warning of the airspace flow field over ships based on shipborne wind profiler radar data. The method includes the following steps:

[0074] Acquire wind profiler radar data and perform two-dimensional convolution on the wind profiler radar data through a CNN convolutional neural network to form two-dimensional wind profiler radar data that retains height features;

[0075] An LSTM deep learning network was constructed and trained using two-dimensional historical data from wind profiler radar. The LSTM deep learning network was then optimized using the Adam method to obtain the trained LSTM deep learning network.

[0076] The trained LSTM deep learning network is used to predict wind profile radar data at future moments, and the predicted two-dimensional data is restored into three-dimensional wind field data with height information by deconvolution.

[0077] By acquiring three-dimensional wind field data, the vertical wind shear value and the horizontal wind shear value are calculated, and the value is used to determine whether to issue a flow field warning for ships.

[0078] To improve the accuracy of predicted wind profiler radar data, the data sampling period needs to be as short as possible. When the forecast duration exceeds the data sampling period, simply forecasting one step forward is insufficient. To meet the requirements for forecast duration, this invention employs a rolling forecasting approach for multi-step forecasting in each forecast, such as... Figure 1 The diagram illustrates a two-step forecast. The dashed line represents the current time, the solid blocks to the left of the dashed line represent historical data, and the numbered hollow blocks to the right of the dashed line represent the data to be forecasted. In the first step of the forecast, the sample input X1 composed of the current time data is used to forecast the value for the next sampling period, i.e., the value at the hollow block with the number 1. In the second step of the forecast, the forecasted value 1 is combined with historical data to form the sample input X2, which forecasts the value for the second calculation period.

[0079] Because LSTM requires rolling forecasts for each height layer of the wind field when processing wind profiler radar data, over long time scales, this can lead to significant differences between different height layers, increasing the calculated low-level wind shear error. To extract high- and low-level wind field features, this invention uses a CNN convolutional neural network to perform two-dimensional convolution on the wind profiler radar data, forming two-dimensional data that retains height features. LSTM deep learning is then used for forecasting to generate the next time-series two-dimensional data. Deconvolution is then used to restore the two-dimensional data to three-dimensional data with height information, and the forecast model is optimized using the Adam method.

[0080] Specifically,

[0081] When performing two-dimensional convolution on wind profiler radar data using a CNN convolutional neural network to form two-dimensional wind profiler radar data that retains height features, by setting the convolution kernel parameters, all depth layer information at each spatial location is integrated into a scalar value. This effectively solves the problem that excessive errors may occur between each height layer when LSTM performs extrapolation calculations. The method is as follows:

[0082] The input tensor for a single time-series data is:

[0083]

[0084] In the formula: D, H, and W represent the number of height grid points, horizontal grid points, and vertical grid points of the data;

[0085] C in This refers to the number of data input channels under a single grid condition; in this item, C... in =3 represents the u, v, and w components of the wind, where the u component is the left-right component facing the bow, the v component is the forward-backward component facing the bow, and the w component is the component perpendicular to the ship.

[0086] The shape of the 3D convolution kernel is:

[0087]

[0088] In the formula:

[0089] K is the kernel size, and in this invention, K = 3;

[0090] C out To determine the number of output channels, this invention uses C. out =H×W.

[0091] To ensure that the convolutional kernel can cover the entire height in one slide, the depth stride is set to D, and the convolutional kernels do not overlap in the depth direction. The convolution process formula is as follows:

[0092]

[0093] In the formula: F out [1,i,j,C out [] represents the two-dimensional output result after convolution; i and j are the relative positions;

[0094] s h s w In this embodiment, s represents the spatial dimension step size. h ,s w =1;

[0095] d, c, k1, k2 are indices; W is the convolution kernel tensor.

[0096] The single-time result output after convolution is:

[0097]

[0098] in:

[0099]

[0100] Using the above methods, the depth information of the 3D fusion data is encoded into the 2D feature map, which reduces the computational complexity while preserving the spatial structure.

[0101] The batch processing method uses a CNN convolutional neural network to transpose the 3D fusion data of multiple time series into 2D data of multiple time series. Then, deep learning methods are used to extract features based on previous time series and extrapolate them to generate 2D data of the next time series.

[0102] This invention utilizes LSTM deep learning for time series extrapolation. The LSTM algorithm consists of an input gate, a forget gate, and an output gate. The first step is to decide what information to discard from the cell state. This decision is made by a sigmoid layer called the "forget gate." It looks at h... t-1 (Previous output) and x t (Current input), and set the cell state to C. t-1 For each number in the (previous state), output a number between 0 and 1. 1 represents keeping it completely, while 0 represents deleting it completely. The module formula is:

[0103] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0104] In the formula: f t This is the forgetting gate vector, with values ​​ranging from [0,1], representing the proportion of the previous memory retained.

[0105] σ is the activation function.

[0106] W f Forget gate weight matrix, dimension n h ×(n h +K), n h As the dimension of memory units, this invention takes n. h =128, K is the kernel size, and in this invention, K=3;

[0107] b f The forget gate bias term is a term of length n. h The vector, b in this invention f The initial value is 1.

[0108] The second step is to decide what information to store in the cell state, called the "cell state update layer." First, the input gate determines the range of data to be updated, consistent with the two-dimensional data after CNN convolution. The cell state update layer then creates candidate vectors C. t This vector will be added to the cell's state. In the next step, we will combine these two vectors to create the update value, using the following module formula:

[0109] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0110]

[0111] In the formula: i t This is the input gate, which determines the proportion of the current input written into memory; The candidate memory vector has a value range of [-1, 1].

[0112] W i W c , is the weight matrix of the input gate and the candidate state, with the same dimension as W. f b i b C For the corresponding bias term, the length is the same as b. f The initial value of this invention is 0.

[0113] The third step is to change the previous state value C. t-1 Updated to C t Multiply the previous state value by f t This expresses the part that one hopes to forget, followed by... The module formula is:

[0114]

[0115] Finally, the model's output is determined, based on the cell states. First, the sigmoid layer is run to determine the output cell states; then, the cell states are passed through the tanh function (normalizing the values ​​to between -1 and 1) and multiplied by the output of the sigmoid gate. This yields the output of one layer of the network. The module formula is as follows:

[0116] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0117] h t =o t *tanh(C t )

[0118] In the formula:

[0119] o t The activation value of the output gate controls how much of the current unit's memory is exposed to the outside; it is a dynamic weighting factor. In this invention, when the wind speed is stable and changes slowly... t =0.4, when wind speed changes abruptly or wind shear increases. t =0.8, when the observation noise is relatively large. t =0.3;

[0120] W o Output gate weight matrix, with the same dimensions as W. f ;

[0121] b o The output should include a bias term with the same length as b. f The initial value of this invention is 0.

[0122] W o Output gate weight matrix, with the same dimensions as W. f ;

[0123] h t This is the output of the LSTM at the current time.

[0124] The system takes two-dimensional data from the current and past periods as input, processed by a CNN convolutional neural network, and uses an LSTM deep learning method to extrapolate and generate two-dimensional data for the next time period. Based on the above algorithm principle, the system takes two-dimensional data from the current and past periods as input, processed by a CNN convolutional neural network, and then passes it through the model network to obtain two-dimensional data for the future time period. Based on the data results, further deconvolution can be performed.

[0125] The method for optimizing LSTM deep learning networks using the Adam approach is as follows:

[0126] First, determine the loss function L:

[0127]

[0128] In the formula:

[0129] x i This represents the actual wind speed.

[0130] To forecast wind speed;

[0131] N is the number of samples.

[0132] The exponential moving average of the gradient direction of the loss function is calculated using a momentum model. Noise is smoothed by modifying the optimization parameters, thus accelerating loss convergence. The method is as follows:

[0133] The optimization parameter is θ, and the loss gradient is... The momentum equation is then:

[0134] m t =β1·m t-1 +(1-β1)·g t

[0135] Where: m t Let β1 be the momentum value at step t, initially set to 0; β1 is the momentum decay rate, which controls the weight of the historical gradient contribution. In this invention, β1 = 0.9.

[0136] The adaptive learning rate is calculated by estimating the variance of the error gradient by adjusting the learning rate parameters, as follows:

[0137] v t =β2·v t-1 +(1-β2)·g t 2

[0138] In the formula: v t β1 is the exponential moving average of the gradient direction at step t, initially set to 0; β2 is the decay rate, which is 0.999 in this invention.

[0139] Because m t and v t Initially set to 0, the model will be biased towards 0 during the initial training phase. Correction is needed to eliminate the influence of this initial bias on the model and make the estimated values ​​closer to the true distribution. The correction formula is as follows:

[0140]

[0141] In the formula: The direction of momentum is indicated by the value; a larger value indicates a more pronounced convergence phenomenon. For adaptive learning rate, A larger value indicates that the gradient direction oscillation can be effectively and consistently maintained.

[0142] pass and The optimization parameter is calculated as θ. t+1 The formula is as follows:

[0143]

[0144] In the formula: η is the initial learning rate, which is 0.001 in this invention; ∈ is the minimum value to prevent the denominator from being 0, which is 1 × 10 in this invention. -8 ;

[0145] Ultimately, θ t+1 It is applied to the input gate weights of LSTM deep learning networks to optimize LSTM deep learning networks.

[0146] The method for restoring the predicted two-dimensional wind field data with height information using deconvolution is as follows:

[0147] The input two-dimensional data in deconvolution recovery is represented as follows:

[0148]

[0149] In the formula:

[0150] F out Two-dimensional data predicted by a trained LSTM deep learning network;

[0151] H′ and W′ represent the number of horizontal and vertical grid points of the deconvolution data, respectively, where H′ = H and W′ = W.

[0152] The deconvolution kernel is represented as follows:

[0153]

[0154] In the formula:

[0155] W (l) It is a deconvolution kernel;

[0156] K d K h K w These are the indexes of the data in the height, horizontal, and vertical directions, respectively.

[0157] The deconvolution formula is as follows:

[0158]

[0159] In the formula:

[0160] d′, h′, w′ are the height, horizontal, and vertical indices of the output 3D data;

[0161] c, c′ are the input channel indices;

[0162] s d s h s w s is the step size in this invention. d s h s w =1.

[0163] The method for calculating the vertical and horizontal wind shear values ​​using the acquired three-dimensional wind field data is as follows:

[0164] Vertical wind shear value C z The calculation formula is as follows:

[0165]

[0166] In the formula:

[0167] V x (h1), V x (h2) represents the wind component in the x direction at heights h1 and h2, respectively;

[0168] V y (h1), V y (h2) represents the wind component in the y direction at heights h1 and h2, respectively;

[0169] Horizontal wind shear represents the difference in wind speed or direction between two points at the same altitude. It is the composite value of horizontal wind shear in the x and y directions, and the formula is as follows:

[0170] Calculate the horizontal wind shear C in the x-direction. x :

[0171]

[0172] In the formula:

[0173] V y1 (h), V y2 (h) represents the wind component in the y direction at two adjacent moments at the same altitude;

[0174] V z1 (h), V z2 (h) represent the wind components in the z-direction at two adjacent moments at the same altitude;

[0175] Calculate the horizontal wind shear C in the y-direction. y :

[0176]

[0177] In the formula:

[0178] V x1 (h), V x2 (h) represents the wind component in the x direction at two adjacent moments at the same altitude;

[0179] The horizontal wind shear in the x and y directions is combined to obtain the horizontal wind shear value C measured by the wind profiler radar. xy :

[0180]

[0181] Referring to the International Civil Aviation Organization's classification standards for wind shear, wind shear intensity is divided into four levels: mild, moderate, severe, and serious; the warning level standards are shown in the table below. This provides commanders with auxiliary decision support for aircraft takeoff and landing safety.

[0182] Table 1 Wind Shear Classification Standards

[0183]

[0184] When the wind shear intensity reaches moderate or above, an early warning will be issued, and an early warning indicator of the corresponding level will automatically pop up. The information pushed includes the wind shear warning level, location, and time. Hovering the mouse over the indicator will display detailed information.

[0185] In this embodiment, existing small- and medium-scale wind field forecasting methods generally rely on numerical model calculations, requiring massive amounts of background field data (such as temperature, humidity, and pressure fields) as input, and obtaining results through complex numerical integration and boundary condition iteration. Although these methods have high accuracy, they generally suffer from long computation time, high dependence on external data, and insufficient real-time performance, making them unsuitable for direct application to short-term early warning for single-ship carriers.

[0186] In contrast, the method of this invention is entirely based on shipborne wind profiler radar data, utilizes a CNN convolutional network to extract vertical structure features, and combines LSTM time series extrapolation and Adam optimization to directly complete prediction calculations on the ship. Since the input only requires single-point radar time series data and does not rely on a large-scale external background field, the calculation steps are greatly simplified, enabling short-term wind shear forecasts to be completed in a very short time, achieving the characteristics of short, flat, and fast forecasts. Therefore, from the perspective of feasibility, this invention is significantly superior to existing methods, providing an efficient, fast, and operable means of early warning of overhead flow fields for ship navigation.

[0187] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for early warning of airflow field over ships based on shipborne wind profiler radar data, characterized in that, The method includes the following steps: Acquire wind profiler radar data and perform two-dimensional convolution on the wind profiler radar data through a CNN convolutional neural network to form two-dimensional wind profiler radar data that retains height features; An LSTM deep learning network was constructed and trained using two-dimensional historical data from wind profiler radar. The LSTM deep learning network was then optimized using the Adam method to obtain the trained LSTM deep learning network. The trained LSTM deep learning network is used to predict wind profile radar data at future moments, and the predicted two-dimensional data is restored into three-dimensional wind field data with height information by deconvolution. By acquiring three-dimensional wind field data, the vertical wind shear value and the horizontal wind shear value are calculated, and the value is used to determine whether to issue a flow field warning for ships.

2. The method for early warning of ship-borne airspace flow field based on shipborne wind profiler radar data according to claim 1, characterized in that: The method for optimizing LSTM deep learning networks using the Adam approach is as follows: First, determine the loss function L; The exponential moving average of the gradient direction of the loss function is calculated using a momentum model. Noise is smoothed by modifying the optimization parameters, thus accelerating loss convergence. The method is as follows: The optimization parameter is θ, and the loss gradient is... The momentum equation is then: m t = β1 · m t-1 + (1 - β1) · g t In the formula, m t is the momentum value of the tth step, and is initially 0; β1 is a momentum attenuation rate, which controls the weight of the historical gradient contribution, and in the present application, β1 = 0.9; The adaptive learning rate is calculated by estimating the variance of the error gradient by adjusting the learning rate parameters, as follows: v t = β2· v t-1 + (1 - β2)· g t 2 In the formula, v t is the exponential moving average of the gradient direction of the tth step, and is initially 0; β2 is a decay rate, and in the present application, β2 = 0.999; Since m t and v t Initial 0, the initial stage of training will be biased to 0, and correction is needed to eliminate the influence of initial bias on the model and make the estimated value closer to the true distribution. The correction formula is as follows: In the formula: The direction of momentum is indicated by the value; a larger value indicates a more pronounced convergence phenomenon. For adaptive learning rate, A larger value indicates that the gradient direction oscillation can be effectively and consistently maintained. pass and The optimization parameter is calculated as θ. t+1 The formula is as follows: In the formula: η is an initial learning rate, in the application, η = 0.001; ∈ is a minimum value, preventing the denominator from being 0, in the application, ∈ = 1 x 10 -8 ; Finally, θ t+1 The input gate weight applied to the LSTM deep learning network is optimized.

3. The method for early warning of ship-borne airspace flow field based on shipborne wind profiler radar data according to claim 1, characterized in that: The method for restoring the predicted two-dimensional wind field data with height information using deconvolution is as follows: The input two-dimensional data in deconvolution recovery is represented as follows: In the formula: F out Two-dimensional data predicted by the trained LSTM deep learning network; H′ and W′ represent the number of horizontal and vertical grid points of the deconvolution data; The deconvolution kernel is represented as follows: In the formula: W (l) is the deconvolution kernel; K d K h K w These are the indexes of the data in the height, horizontal, and vertical directions, respectively. The deconvolution formula is as follows: In the formula: d′, h′, w′ are the height, horizontal, and vertical indices of the output 3D data; c, c′ are the input channel indices; s d s h s w s is the step size in this invention. d s h s w =1.

4. The method for early warning of ship-borne airspace flow field based on shipborne wind profiler radar data according to claim 1, characterized in that: The method for calculating the vertical and horizontal wind shear values ​​using the acquired three-dimensional wind field data is as follows: Vertical wind shear value C z The calculation formula is as follows: In the formula: V x (h1), V x (h2) represents the wind component in the x direction at heights h1 and h2, respectively; V y (h1), V y (h2) represents the wind component in the y direction at heights h1 and h2, respectively; Horizontal wind shear represents the difference in wind speed or direction between two points at the same altitude. It is the composite value of horizontal wind shear in the x and y directions, and the formula is as follows: Calculate the horizontal wind shear C in the x-direction. x : In the formula: V y1 (h), V y2 (h) represents the wind component in the y direction at two adjacent moments at the same altitude; V z1 (h), V z2 (h) represent the wind components in the z-direction at two adjacent moments at the same altitude; Calculate the horizontal wind shear C in the y-direction. y : In the formula: V x1 (h), V x2 (h) represents the wind component in the x direction at two adjacent moments at the same altitude; The horizontal wind shear in the x and y directions is combined to obtain the horizontal wind shear value C measured by the wind profiler radar. xy :

5. The method for early warning of ship-borne airspace flow field based on shipborne wind profiler radar data according to claim 4, characterized in that: The method for determining whether to issue a flow field warning for ships using numerical data is as follows: Based on the vertical and horizontal wind shear values, wind shear intensity is divided into four levels: mild, moderate, severe, and serious. An early warning is issued when the shear intensity reaches moderate or above.

6. The method for early warning of ship-borne airspace flow field based on shipborne wind profiler radar data according to claim 2, characterized in that: The determined loss function L is: In the formula: x i This represents the actual wind speed. To forecast wind speed; N is the number of samples.

Images