A method and device for predicting extreme values ​​of dynamic submarine cable response

Abstract

This invention discloses a method and device for predicting the extreme values ​​of dynamic submarine cable response, integrating physical mechanisms and dynamic temporal characteristics. By deeply fusing the dynamic submarine cable control equations with data-driven training through a physical information neural network, the mechanical mechanisms of the dynamic submarine cable are transformed into residual constraint terms, overcoming the poor generalization of purely data-driven methods. An LSTM temporal prediction model extracts the amplitude features of the six-degree-of-freedom motion of the floating body, compressing long-term dynamic coupling effects into physically interpretable features. A structure type judgment module is designed, covering two mainstream submarine cable structures without model reconstruction, significantly enhancing engineering applicability. A time-varying weight strategy dynamically balances data fitting loss, physical residual loss, and motion prediction loss, accelerating convergence while ensuring the prediction results conform to physical laws. By outputting the extreme values ​​of key node responses and motion amplitude features in real time, dynamic submarine cable strength failure early warning is provided for offshore wind power, greatly reducing the operation and maintenance costs of offshore wind power.

Description

Technical Field

[0001] This invention belongs to the field of offshore wind power generation technology, and relates to a method and device for predicting the extreme values ​​of dynamic submarine cable response, and more particularly to a method and device for predicting the extreme values ​​of dynamic submarine cable response that integrates physical mechanisms and dynamic time-series characteristics. Background Technology

[0002] With the advancement of the offshore wind power industry, deep-sea floating wind power has become the mainstream direction of offshore wind power development and a research hotspot internationally. However, the development of offshore floating wind turbines still faces severe technical challenges, one of the key challenges being ensuring the reliability and integrity of the transmission cables. As a critical power transmission component of floating wind power systems, dynamic cables are subjected to various complex loads during their service life, including the movement of the floating body itself and marine environmental loads. They are prone to strength failure at critical nodes (such as the top suspension point and the buoy connection area) due to excessive tension or insufficient bending radius. Therefore, accurately predicting the extreme values ​​of the dynamic cable response is crucial for real-time safety early warning of offshore wind power.

[0003] However, existing technologies still have many shortcomings. Currently, traditional numerical simulation methods mainly involve establishing finite element models and calculating the extreme values ​​of the response under extreme loads through time-domain dynamic simulation. While this method offers high accuracy, it requires separate modeling and calculation for each sea state, resulting in long processing times and high computational resource consumption, failing to meet real-time early warning requirements. To improve computational efficiency, some studies employ neural networks (such as BP and GA-BP) to establish mapping models between sea state parameters and response extreme values. Although these methods can significantly accelerate prediction, they are essentially "black box models" with serious flaws, such as strong dependence on training samples, lack of physical laws, poor structural adaptability, and insufficient generalization. Furthermore, current research has failed to effectively extract the dynamic influence of the floating body's six-degree-of-freedom displacement time-series characteristics on the submarine cable response, making it difficult to guarantee the predictive reliability of motion amplitude characteristics (displacement standard deviation, maximum displacement amplitude).

[0004] Based on the above considerations, there is an urgent need to develop a dynamic submarine cable response extreme value prediction method that integrates physical mechanisms and dynamic temporal characteristics. By combining physical mechanisms, temporal dynamic characteristics and the advantages of data-driven approaches, this new method can overcome the bottlenecks of numerical simulation and neural network simulation, and provide a highly reliable early warning means for the strength failure of dynamic submarine cables for large floating wind turbines. Summary of the Invention

[0005] To address the challenges of time-consuming traditional numerical methods that fail to meet real-time early warning requirements, and the insufficient generalization ability of traditional pure data-driven neural networks, this invention provides a method and device for predicting extreme values ​​of dynamic submarine cable response by integrating physical mechanisms and dynamic temporal characteristics. This method is used to predict extreme loads and provide early warning of strength failure in dynamic submarine cables for large floating wind turbines. The method of this invention is more accurate in predicting extreme values ​​of dynamic submarine cable response for large floating wind turbines, and its iterative optimization is more efficient.

[0006] The technical solution adopted in this invention is as follows:

[0007] A dynamic submarine cable response extreme value prediction method integrating physical mechanisms and dynamic temporal characteristics includes the following steps:

[0008] Construct a multi-condition training set, which includes sea state parameter vectors and floating body six-degree-of-freedom displacement time series data as model inputs, and response extreme values ​​and motion amplitude characteristics of dynamic submarine cable key nodes as model outputs.

[0009] Construct an LSTM time series prediction model, with the input being the time series data of the six degrees of freedom displacement of the floating body, and the output being the motion amplitude features;

[0010] A physical information neural network is constructed, with sea state parameter vectors as input and response extrema as output. The output of geometric parameters is dynamically selected based on the number of buoys.

[0011] A structure type determination module is constructed to select the corresponding physical residual calculation path based on the real-time received float quantity parameter, and calculate the physical residual based on the output of the physical information neural network.

[0012] A composite loss function is constructed that integrates data fitting loss, physical residual loss, and motion prediction loss. Time-varying weight coefficients are used to dynamically balance the three losses, emphasizing data fitting in the early stage of training and strengthening physical constraints in the later stage.

[0013] The gradient of the composite loss function with respect to the network weights is calculated by backpropagation, and the network weights are updated using gradient descent until convergence.

[0014] The real-time sea state parameters are input into the trained physical information neural network model, which outputs the prediction results of the extreme values ​​of the response and the motion amplitude characteristics of the key nodes of the dynamic submarine cable.

[0015] Furthermore, the construction of the multi-condition training set specifically includes:

[0016] Spatial sampling methods are used to sample within a set range of sea state parameters to obtain the sea state parameter vector and the time series data of the six degrees of freedom displacement of the floating body as input to the model.

[0017] A parameterized dynamic submarine cable model is established using hydrodynamic analysis software. The sea state parameter vector is used as the model input of the dynamic submarine cable model, and the extreme values ​​of the response and motion amplitude characteristics of the key nodes of the dynamic submarine cable are output.

[0018] Furthermore, the sea state parameter vector includes meaningful wave height, wave period, wave direction, surface flow velocity, surface flow direction, wind speed, and wind direction; the response extrema include effective tension and bending radius; and the motion amplitude characteristics include displacement standard deviation and maximum displacement amplitude.

[0019] Furthermore, in the LSTM time series prediction model, the input layer receives the six-degree-of-freedom displacement time series data of the floating body, the hidden layer adopts a long short-time memory network, and the output layer generates motion amplitude feature vectors.

[0020] Furthermore, in the physical information neural network, the input layer receives the sea state parameter vector, the hidden layer adopts a fully connected neural network, and the output layer dynamically selects the output items according to the number of buoys: when the number of buoys is 0, only the response extreme value is output; when the number of buoys > 0, the response extreme value and geometric parameters are output.

[0021] Furthermore, the step of selecting the corresponding physical residual calculation path based on the real-time received float quantity parameter is as follows:

[0022] When the number of buoys is 0, the dynamic submarine cable is a catenary type, and the catenary control equation is selected to calculate the physical residual;

[0023] When the number of buoys is greater than 0, the dynamic submarine cable is of the gentle wave type, and the gentle wave type control equation is selected to calculate the physical residual.

[0024] Furthermore, the physical residuals are calculated using the catenary control equations as follows:

[0025]

[0026] in, For physical residuals, The dynamic submarine cable tension prediction value at the suspension point is output by a physical information neural network. The horizontal external force at the suspension point is input from the simulation. The mass per unit length of the dynamic submarine cable is a material constant. It is the acceleration due to gravity; This refers to the length of the catenary section, which is a design parameter.

[0027] The physical residuals are calculated using the catenary control equations as follows:

[0028]

[0029] in, The angle at the connection point between the pontoon section and the lower suspension chain section is output by a physical information neural network. The angle at the seabed contact point is output by a physical information neural network. This refers to the length of the lower suspension chain segment, which is a design parameter. The wet weight ratio of the catenary segment is equal to , The horizontal tension is input from the simulation. This is the wet weight per unit length of the submarine cable, which is a material constant. Water depth is an environmental input. The segment height is k=1, 2, 3, and is output by the physical information neural network.

[0030] Furthermore, the composite loss function for:

[0031]

[0032] For data fitting loss term, it represents the mean square error between the predicted extreme value of the response output by the physical information neural network and the simulated value; The physical residual loss term is the physical residual obtained from the structure type determination module. Calculated, i.e. ; For motion prediction loss, it represents the mean square error between the predicted values ​​and simulated values ​​of the motion amplitude features output by the LSTM time series prediction model. , and , respectively, are the time-varying weight coefficients of the data fitting loss term, the physical residual loss term, and the motion prediction loss term, where t is the training round.

[0033] Furthermore, the time-varying weighting coefficients , and The time-varying weight coefficients are dynamically adjusted according to the training round t. , and The specific calculation formula is as follows:

[0034]

[0035]

[0036]

[0037] in, The initial weights are used to fit the data to the loss. The maximum weight is assigned to the physical residual loss. The weight that maximizes the motion prediction loss; The rate of increase of physical residual loss. To fit the loss decay rate to the data, To predict the rate of loss growth during motion, satisfy This results in the data fitting loss being dominant in the early stages of training, the physical residual loss being dominant in the middle stages of training, and the motion prediction loss continuously increasing to the preset maximum value in the later stages of training.

[0038] A computer device, the computer device comprising:

[0039] One or more processors;

[0040] Memory, used to store one or more programs;

[0041] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned dynamic submarine cable response extreme value prediction method that integrates physical mechanisms and dynamic timing characteristics.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] This invention deeply integrates the dynamic submarine cable control equations with data-driven training through a physical information neural network, transforming the mechanical mechanism of the dynamic submarine cable into residual constraint terms. This significantly improves the predictive reliability of the model under out-of-sample conditions such as extreme sea states, overcoming the poor generalization of pure data-driven methods. By extracting the amplitude features of the six-degree-of-freedom motion of the floating body through an LSTM time-series prediction model, the long-term dynamic coupling effect is compressed into physically interpretable features, addressing the shortcomings of traditional methods that ignore the real-time interaction between the floating body and the submarine cable, thus improving the reliability of motion amplitude prediction. An innovative structure type judgment module is designed, covering two mainstream submarine cable structures without model reconstruction, greatly enhancing engineering applicability. A time-varying weight strategy is used to dynamically balance data fitting loss, physical residual loss, and motion prediction loss, accelerating convergence while ensuring that the prediction results conform to physical laws. By outputting the extreme values ​​of key node responses and motion amplitude features in real time, dynamic submarine cable strength failure early warning is provided for offshore wind power, reducing the risk of cable breakage and significantly reducing the operation and maintenance costs of offshore wind power. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall process of the method in the embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram illustrating the principle of a catenary-type dynamic submarine cable structure in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram illustrating the principle of the wave-slowing dynamic submarine cable structure in an embodiment of the present invention.

[0047] Figure 4This is a schematic diagram illustrating the principle of the LSTM time series prediction model in an embodiment of the present invention.

[0048] Figure 5 This is a schematic diagram of the physical information neural network topology in an embodiment of the present invention.

[0049] Figure 6 and Figure 7 The image shows a comparison between the prediction method and the prediction of the effective extreme value of dynamic submarine cable under the BP neural network model in the embodiments of the present invention, demonstrating that the method can accurately predict the effective tension extreme value of dynamic submarine cable under extreme sea conditions. Detailed Implementation

[0050] The technical solution of the present invention will be further described clearly and in detail below with reference to the accompanying drawings and specific embodiments.

[0051] A dynamic submarine cable response extreme value prediction method integrating physical mechanisms and dynamic temporal characteristics includes the following steps:

[0052] (1) Spatial sampling method is used to sample within the set sea state parameter range to obtain the sea state parameter vector as the model input. And the time series data of the six degrees of freedom displacement of the floating body (sway, roll, heave, pitch, pitch, bow roll); among which For the sake of righteousness, the waves rise high. For wave cycles, The direction of the wave. For surface flow rate, For surface flow direction, For wind speed, Wind direction;

[0053] A parametric dynamic submarine cable model is established using hydrodynamic analysis software (such as OrcaFlex, Sesam, WAMIT, etc.). Sea state parameter vectors are used as inputs to the dynamic submarine cable model. Dynamic analysis of the model yields the extreme values ​​of the responses at key nodes of the dynamic submarine cable under the corresponding sea states. (Effective tension, bending radius) and motion amplitude characteristics (displacement standard deviation) Maximum displacement amplitude The simulation covers two typical structures: catenary (without pontoons) and wave-damping (with pontoons), and ensures diverse operating conditions through variable parameter scanning.

[0054] A multi-condition training set was constructed using sea state parameter vectors, time series data of six-degree-of-freedom displacement of the floating body, and response extreme values ​​and motion amplitude characteristics of key nodes of the dynamic submarine cable.

[0055] (2) Construct an LSTM time series prediction model. Its input layer receives the six-degree-of-freedom displacement time series data of the floating body, the hidden layer adopts a long short-time memory network, and the output layer generates the motion amplitude feature vector of the key nodes of the dynamic submarine cable. .

[0056] (3) Construct a Physical Information Neural Network (PINN) with the following topology: The input layer receives the sea state parameter vector. The hidden layer uses a fully connected neural network, with ReLU being the preferred activation function. The output layer is determined based on the number of pontoons. Dynamically selectable output: When the number of floats is 0, only the extreme values ​​of the response are output. When the number of pontoons > 0, the extreme value of the output response is... and geometric parameters , , .

[0057] (4) Construct a structure type judgment module. After forward propagation calculation in the physical information neural network, the module determines the structure type based on the number of floats received in real time. Select the corresponding physical residual calculation path and calculate the physical residual based on the output of the physical information neural network. .

[0058] When the number of buoys is 0, the dynamic submarine cable is a catenary type. The catenary control equation is selected to calculate the physical residuals:

[0059] in, The dynamic submarine cable tension prediction value at the suspension point is output by a physical information neural network. The horizontal external force at the suspension point is input from the simulation. The mass per unit length of the dynamic submarine cable is a material constant. It is the acceleration due to gravity; This refers to the length of the catenary section, which is a design parameter.

[0060] When the number of buoys > 0, the dynamic submarine cable is of the wave-slowing type. The wave-slowing type governing equation is selected to calculate the physical residuals:

[0061]

[0062] in, The angle at the connection point between the pontoon section and the lower suspension chain section is output by a physical information neural network. The angle at the seabed contact point is output by a physical information neural network. This refers to the length of the lower suspension chain segment, which is a design parameter. The wet weight ratio of the catenary segment is equal to , The horizontal tension is input from the simulation. This is the wet weight per unit length of the submarine cable, which is a material constant. Water depth is an environmental input. The segment height is k=1, 2, 3, and is output by the physical information neural network.

[0063] (5) Construct a composite loss function that integrates data fitting loss, physical residual loss, and motion prediction loss using a time-varying weighting strategy. :

[0064]

[0065] in, For data fitting loss term, it represents the mean square error between the predicted extreme value of the response output by the physical information neural network and the simulated value; The physical residual loss term is the physical residual obtained from the structure type determination module. Calculated, i.e. ; For motion prediction loss, it represents the mean square error between the predicted values ​​and simulated values ​​of the motion amplitude features output by the LSTM time series prediction model. , and , respectively, are the time-varying weight coefficients of the data fitting loss term, the physical residual loss term, and the motion prediction loss term, where t is the training round.

[0066] The data fitting loss term The calculation equation is as follows:

[0067]

[0068] in, This represents the number of samples used to train PINN; For the neural network to the first Predicted values ​​for each sample; For the first Simulated values ​​for each sample.

[0069] The motion prediction loss term The calculation equation is as follows:

[0070]

[0071] in, This represents the number of time-series samples used to train the LSTM. For the first The standard deviation of displacement predicted by the LSTM model in each time series sample segment; For the first The standard deviation of the simulated displacement corresponding to each time series sample segment; For the first The maximum displacement amplitude predicted by the LSTM model in each time series sample segment; For the first The simulated maximum displacement amplitude corresponding to each time series sample segment.

[0072] The time-varying weighting coefficient , and The time-varying weight coefficients are dynamically adjusted according to the training epoch t, ​​ensuring that data fitting is the primary focus in the early stages of training, while physical constraints and temporal characteristics are emphasized in the mid-to-late stages, balancing data fitting with adherence to physical laws. , and The specific calculation formula is as follows:

[0073]

[0074]

[0075]

[0076] in, The initial weights are used to fit the data to the loss. The maximum weight is assigned to the physical residual loss. The weight that maximizes the motion prediction loss; The rate of increase of physical residual loss. To fit the loss decay rate to the data, To predict the rate of loss growth during motion, satisfy This results in the data fitting loss being dominant in the early stages of training, the physical residual loss being dominant in the middle stages of training, and the motion prediction loss continuously increasing to the preset maximum value in the later stages of training.

[0077] (6) Calculate the composite loss function using the backpropagation algorithm. Network weights gradient :

[0078]

[0079] in, This is the weight matrix of the neural network; The data fitting loss gradient is calculated by applying the data fitting loss function to the weight matrix. The set of partial derivatives of each element in the set; The physical residual loss gradient is the result of applying the physical residual loss function to the weight matrix. The set of partial derivatives of each element in the set; The gradient of the motion prediction loss is the result of applying the motion prediction loss function to the weight matrix. The set of partial derivatives of each element in the set.

[0080] Update network weights using gradient descent:

[0081]

[0082] in, This is the updated weight matrix; Let be the weight matrix for the k-th step; k is the current iteration step number. The learning rate is used to adjust the degree to which the gradient affects the update of the weight matrix.

[0083] Iterate through the training until convergence is determined when any of the following conditions are met:

[0084] or

[0085] in, This is the weighted convergence threshold; This is the loss convergence threshold.

[0086] (7) Input the real-time sea state parameters and floating body motion time series data into the trained fusion network, and output the key node response extreme value motion amplitude characteristics and motion amplitude characteristics.

[0087] The gradient is calculated through backpropagation, and the network weights are updated using gradient descent until convergence.

[0088] The real-time sea state parameters are input into the trained physical information neural network model, which outputs the extreme value prediction results of the response of key nodes of the dynamic submarine cable.

[0089] Example

[0090] like Figure 1 The diagram shows a flowchart of the dynamic submarine cable response extreme value prediction method that integrates physical mechanisms and dynamic temporal characteristics according to the present invention. This method is implemented based on the core framework of a physical information neural network and an LSTM temporal prediction module, forming a complete "AI-enhanced physical mechanism fusion method".

[0091] Using the Latin hypercube sampling method, sampling is performed within a set range of sea state parameters to obtain the meaningful wave height as an input parameter. Wave cycle Wave direction Surface flow rate Surface flow direction Wind speed ,wind direction Combined with the time-series data of the six-DOF displacements of the floating body (swell, roll, heave, pitch, pitch, bow roll); a numerical model was established in marine engineering dynamics analysis software (such as OrcaFlex, Sesam, WAMIT, etc.), with sea state parameter vectors as input. The output is the extreme values ​​of the dynamic submarine cable critical node response. (Effective tension, bending radius) and motion amplitude characteristics (displacement standard deviation) Maximum displacement amplitude ).

[0092] An LSTM time series prediction model is constructed. The input layer receives the six-DOF displacement time series data of the floating body, the hidden layer adopts a long short-time memory network, and the output layer generates the motion amplitude feature vectors of key nodes of the dynamic submarine cable. .

[0093] Construct a physical information neural network with the following topology: the input layer receives a vector of sea state parameters. The hidden layer uses a fully connected neural network (with ReLU activation function); the output layer is based on the number of pontoons. Dynamically select output items.

[0094] Design a structural type determination module based on the number of pontoons. Activate the corresponding physical residual term If the number of buoys is 0, then the residual terms of the control equations for the catenary dynamic submarine cable are activated. Its structural principle diagram is as follows Figure 2 As shown, the physical residual term for the catenary type is obtained based on the vector synthesis of the dynamic cable tension at the suspension point as a result of the horizontal external force and the cable weight; if the number of buoys is greater than 0, then the residual term of the control equation for the wave-slowing type dynamic cable is activated. Its structural principle diagram is as follows Figure 3 As shown, the wave-type physical residual term is obtained based on the angular continuity constraint and the geometric compatibility constraint.

[0095] Define the composite loss function It integrates data fitting loss, physical residual loss, and motion prediction loss, with time-varying weight coefficients. , and The training round t is dynamically adjusted to ensure that data fitting is the primary focus in the early stages of training, while physical constraints and temporal characteristics are strengthened in the middle and later stages, thus balancing data fitting with adherence to physical laws.

[0096] Calculate the composite loss using the backpropagation algorithm. Network weights gradient :

[0097]

[0098] Update network weights using gradient descent:

[0099]

[0100] Iterate through the training until convergence is determined when any of the following conditions are met:

[0101] or

[0102] Real-time sea state parameters and floating body motion time series data are input into the trained fusion network, and the output results of extreme value prediction of key node response and motion amplitude characteristics are output.

[0103] Figure 4 This is a schematic diagram of the LSTM timing prediction model in the method of this invention, used to process the six-degree-of-freedom motion timing data of a floating body. The LSTM timing prediction model consists of an input layer, an LSTM hidden layer, and a fully connected output layer. The system receives the six-degree-of-freedom displacement timing data of the floating body. , ,in The displacement vector of the floating body at time step t (6-dimensional); the LSTM hidden layer adopts a multi-layer LSTM cell structure ( Figure 4 Taking a single-layer example, it can actually be stacked. The core processing involves three key steps: first, the gating mechanism is computed, and then the forget gate... The percentage of historical information retained is determined (0 = complete forgetting, 1 = complete retention), among which It is the Sigmoid activation function. This is the weight matrix of the forget gate (f is a special parameter indicating that the parameter belongs to the forget gate). This is the hidden state from the previous moment. Let t be the displacement vector of the floating body at the t-th time step (6-dimensional). The bias vector for the forget gate; the vector input gate. Controlling the degree of updating of new information, among which is the weight matrix of the vector input gate (i is a special parameter indicating that the parameter belongs to the vector input gate). The bias vector of the input gate; candidate state t Where tanh is the hyperbolic tangent activation function. is the weight matrix of the candidate states (c is a special parameter indicating that the parameter belongs to the candidate state). This is the bias vector for the candidate state. Next is the state update, the cell state. t ,in For the forget gate of the current time step, This represents the cell state at the previous time step. This is the input step for the current time step. t Given the candidate cell state at the current time step; integrate historical and current information, output gate. = , (This is the weight matrix of the output gate, and o is a special parameter indicating that this parameter belongs to the output gate). The bias vector of the output gate; hidden state Generate the current output (including compressed memory information), where This is the output gate for the current time step. The cell state at the current time step; the first layer of the LSTM outputs the hidden state. This process is repeated as the input to the second layer until the temporal features at the final time step T are obtained. ,in This represents the hidden state of the final layer (L) at the last time step (T). This is the compressed global temporal feature vector. Finally, the motion amplitude feature vector of the key nodes of the dynamic submarine cable is output at the fully connected output layer. ,in The calculation formula is as follows:

[0104]

[0105] in It is a non-linear activation function. This is the weight matrix of the output layer (usually a two-dimensional matrix). To represent the hidden state at the current time step, This is the bias vector (one-dimensional vector). Finally, the displacement standard deviation is obtained. and maximum displacement amplitude It is then passed to the composite loss function for calculation.

[0106] Figure 5 This is a schematic diagram of the physical information neural network topology in the method of this invention. The Physical Information Neural Network (PINN) adopts a layered topology architecture, consisting of an input layer, hidden layers, and a dynamic output layer. The input layer receives a sea state parameter vector. Each input node corresponds to a specific environmental variable; the hidden layer, as the core feature extraction module, adopts a two-layer fully connected structure, with each layer containing several neuron nodes; the hidden layer achieves high-dimensional feature transformation through a non-linear activation function (ReLU).

[0107]

[0108] in, For the first The output vector of the hidden layer; For the first Layer weight matrix; For the first The output vector of the layer (i.e., the input of the current layer); For the first Layer bias vector; The activation function is defined as follows: z is the linear output value of a single neuron (from...) ), max is the function to find the maximum value.

[0109] After the features extracted from the hidden layer are input, they are determined according to the number of pontoons. Dynamically select output items: If the number of pontoons is 0, output the extreme values ​​of the predicted response at critical nodes. If the number of pontoons is greater than 0, output the extreme values ​​of the predicted response at critical nodes. and geometric parameters ( , , ).

[0110] Figure 6 The test set difference and absolute percentage error (effective tension extreme value) are compared between the BP neural network and the hybrid network in this embodiment. Figure 7 This section compares the prediction dispersion (effective tension extreme values) of the BP neural network and the hybrid network in this embodiment on the test set sample points. The input values ​​of the hybrid network model in this embodiment are seven specific environmental parameters and the time-series displacement data of the six degrees of freedom of the floating body. The output values ​​are the maximum effective tension, minimum bending radius, and motion amplitude characteristics of the dynamic submarine cable under specific environmental parameters. The ratio of the training set to the test set is set to 8:2. The performance of the two models is compared based on their prediction performance on the test set. Figure 6 and Figure 7 It can be seen that the BP neural network has a large prediction error at some sample points, with the average error significantly exceeding that of the hybrid network prediction model. As the response value increases, its predicted value gradually decreases, and the difference from the true value gradually increases. This means that when more severe sea conditions occur, the BP neural network will predict a more dangerous effective tension extreme value, which is not conducive to timely and effective hazard warnings for dynamic submarine cables. In this embodiment, the hybrid network model can effectively solve the above problems. Even at locations with large response values, it can still make relatively accurate predictions, and the dispersion of its test set sample points is significantly lower than that of the BP neural network. Therefore, in this embodiment, the hybrid network model has better prediction performance for effective tension extreme values.

[0111] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0112] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0113] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0114] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0115] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.

Claims

1. A method for predicting the extreme values ​​of dynamic submarine cable response by integrating physical mechanisms and dynamic temporal characteristics, characterized in that, Includes the following steps: Construct a multi-condition training set, which includes sea state parameter vectors and floating body six-degree-of-freedom displacement time series data as model inputs, and response extreme values ​​and motion amplitude characteristics of dynamic submarine cable key nodes as model outputs. Construct an LSTM time series prediction model, with the input being the time series data of the six degrees of freedom displacement of the floating body, and the output being the motion amplitude features; A physical information neural network is constructed, with sea state parameter vectors as input and response extrema as output. The output of geometric parameters is dynamically selected based on the number of buoys. A structure type determination module is constructed to select the corresponding physical residual calculation path based on the real-time received float quantity parameter, and calculate the physical residual based on the output of the physical information neural network. A composite loss function is constructed that integrates data fitting loss, physical residual loss, and motion prediction loss, and the three losses are dynamically balanced using time-varying weight coefficients; the data fitting loss represents the mean square error between the predicted extreme value of the response output of the physical information neural network and the simulated value. The gradient of the composite loss function with respect to the network weights is calculated by backpropagation, and the network weights are updated using gradient descent until convergence. Real-time sea state parameters and floating body motion time series data are input into the trained fusion network, which outputs the prediction results of the extreme values ​​of the response and motion amplitude characteristics of key nodes of the dynamic submarine cable.

2. The dynamic submarine cable response extreme value prediction method integrating physical mechanisms and dynamic temporal characteristics according to claim 1, characterized in that, The construction of the multi-condition training set specifically includes: Spatial sampling methods are used to sample within a set range of sea state parameters to obtain the sea state parameter vector and the time series data of the six degrees of freedom displacement of the floating body as input to the model. A parameterized dynamic submarine cable model is established using hydrodynamic analysis software. The sea state parameter vector is used as the model input of the dynamic submarine cable model, and the extreme values ​​of the response and motion amplitude characteristics of the key nodes of the dynamic submarine cable are output.

3. The dynamic submarine cable response extreme value prediction method integrating physical mechanisms and dynamic temporal characteristics according to claim 2, characterized in that, The sea state parameter vector includes meaningful wave height, wave period, wave direction, surface flow velocity, surface flow direction, wind speed, and wind direction; the response extrema include effective tension and bending radius; the motion amplitude characteristics include displacement standard deviation and maximum displacement amplitude.

4. The method for predicting the extreme values ​​of dynamic submarine cable response by integrating physical mechanisms and dynamic temporal characteristics according to claim 1, characterized in that, In the LSTM time series prediction model, the input layer receives the six-degree-of-freedom displacement time series data of the floating body, the hidden layer adopts a long short-time memory network, and the output layer generates motion amplitude feature vectors.

5. The dynamic submarine cable response extreme value prediction method integrating physical mechanisms and dynamic temporal characteristics according to claim 1, characterized in that, In the physical information neural network, the input layer receives the sea state parameter vector, the hidden layer adopts a fully connected neural network, and the output layer dynamically selects the output items according to the number of buoys: when the number of buoys is 0, only the response extreme value is output; when the number of buoys > 0, the response extreme value and geometric parameters are output.

6. The method for predicting the extreme values ​​of dynamic submarine cable response by integrating physical mechanisms and dynamic temporal characteristics according to claim 1, characterized in that, The method for selecting the corresponding physical residual calculation path based on the real-time received float quantity parameter is as follows: When the number of buoys is 0, the dynamic submarine cable is a catenary type, and the catenary control equation is selected to calculate the physical residual; When the number of buoys is greater than 0, the dynamic submarine cable is of the gentle wave type, and the gentle wave type control equation is selected to calculate the physical residual.

7. The dynamic submarine cable response extreme value prediction method integrating physical mechanisms and dynamic temporal characteristics according to claim 6, characterized in that, The physical residuals are calculated using the catenary control equations as follows: ， in, For physical residuals, The dynamic submarine cable tension prediction value at the suspension point is output by a physical information neural network. The horizontal external force at the suspension point is input from the simulation. The mass per unit length of the dynamic submarine cable is a material constant. It is the acceleration due to gravity; This refers to the length of the catenary section, which is a design parameter. The physical residuals are calculated using the catenary control equations as follows: ， in, The angle at the connection point between the pontoon section and the lower suspension chain section is output by a physical information neural network. The angle at the seabed contact point is output by a physical information neural network. This refers to the length of the lower suspension chain segment, which is a design parameter. The wet weight ratio of the catenary segment is equal to , The horizontal tension is input from the simulation. This is the wet weight per unit length of the submarine cable, which is a material constant. Water depth is an environmental input. The segment height is k=1, 2, 3, and is output by the physical information neural network.

8. The method for predicting the extreme values ​​of dynamic submarine cable response by integrating physical mechanisms and dynamic temporal characteristics according to claim 1, characterized in that, The composite loss function for: ， For data fitting loss term, it represents the mean square error between the predicted extreme value of the response output by the physical information neural network and the simulated value; The physical residual loss term is the physical residual obtained from the structure type determination module. Calculated, i.e. ; For motion prediction loss, it represents the mean square error between the predicted values ​​and simulated values ​​of the motion amplitude features output by the LSTM time series prediction model. , and , respectively, are the time-varying weight coefficients of the data fitting loss term, the physical residual loss term, and the motion prediction loss term, where t is the training round.

9. The method for predicting the extreme values ​​of dynamic submarine cable response by integrating physical mechanisms and dynamic temporal characteristics according to claim 8, characterized in that, The time-varying weighting coefficient , and The time-varying weight coefficients are dynamically adjusted according to the training round t. , and The specific calculation formula is as follows: ， ， ， in, The initial weights are used to fit the data to the loss. The maximum weight is assigned to the physical residual loss. The weight that maximizes the motion prediction loss; The rate of increase of physical residual loss. To fit the loss decay rate to the data, To predict the rate of loss growth during motion, satisfy This results in the data fitting loss being dominant in the early stages of training, the physical residual loss being dominant in the middle stages of training, and the motion prediction loss continuously increasing to the preset maximum value in the later stages of training.

10. A computer device, characterized in that, The computer device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the dynamic submarine cable response extreme value prediction method that integrates physical mechanisms and dynamic temporal characteristics as described in any one of claims 1 to 9.

Images