Balanced regulation method of LNG ship formation combining virtual connection with distribution prediction

Abstract

The present application belongs to the field of shipping, and particularly relates to a kind of balanced regulation method of LNG ship formation fusing virtual connection and distribution prediction.The method first characterizes the multi-body system of LNG ship formation;The multi-body system sets up a virtual connection relationship between ships to represent mutual influence.Using the multi-body system, the risk state vector corresponding to the ship navigation state is generated, and then the risk prediction vector is generated in combination with the virtual connection relationship;The risk prediction vector of each ship is input into the pre-trained risk prediction model based on time series data, the local risk of the ship is predicted and fused into the global risk of the formation;According to the global state and risk level, the preset control strategy experience set is queried to obtain the optimal control strategy;And when the yaw angle or pitch angle of any ship exceeds the safety threshold, the balanced control instruction is issued to it.The present application solves the problem of lack of balanced control strategy at the ship formation level in the prior art, and ensures the navigation stability of LNG ship formation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of shipping, and particularly relates to a balance adjustment method for LNG ship formation by fusing virtual connection and distributed prediction, and a corresponding computer program product and ship formation balance control device. BACKGROUND

[0002] In the field of ship engineering technology, the stability and safety of ship navigation have always been the focus of research. With the development of trade demand, LNG ship formation navigation is increasing, and ensuring the stability and safety of LNG ship formation navigation can optimize shipping resource allocation, reduce transportation risk and improve shipping efficiency.

[0003] In traditional ship balance control research, most control methods mainly focus on the balance adjustment of a single ship itself, focusing on the structural design of the ship itself and the limited balance device control, but ignoring the importance of ship formation collaborative operation. When focusing on LNG ship formation, the limitations of the above traditional balance control methods are more prominent. During the navigation of LNG ships, the liquid in the tank has a large and random sloshing amplitude, which has a non-negligible impact on the balance state of the ship. However, the existing methods almost do not comprehensively and systematically integrate multi-ship collaborative control into the balance control system. At the same time, in LNG ship formation navigation, the active balance weight system as a key auxiliary facility, its optimization degree is directly related to the ability of the ship to cope with liquid sloshing.

[0004] However, current research on LNG ship active balance weight system pays little attention to the collaborative optimization of energy consumption and balance effect, including the comprehensive consideration of weight block movement strategy, overall balance state of the ship and energy consumption, and the comprehensive influence of these factors on dynamic balance control of the ship. This leads to the fact that the ship is difficult to efficiently cope with the unbalanced force caused by LNG liquid sloshing during navigation, greatly increasing the risk of navigation; the ship design and manufacturing department lacks scientific and effective balance control scheme support when designing LNG ships, making it difficult to reasonably allocate balance devices; and the shipping company also lacks precise balance control method when operating LNG ships, which cannot guarantee the safety and stability of navigation. SUMMARY

[0005] In order to solve the problem that the prior art lacks a balance adjustment method for ship formation, the present application provides a balance adjustment method for LNG ship formation by fusing virtual connection and distributed prediction, and a corresponding computer program product and ship formation balance control device.

[0006] The present application adopts the following technical solutions:

[0007] A method for balancing LNG vessel fleets by integrating virtual connectivity and distributed prediction, comprising:

[0008] A multibody system characterizing an LNG ship convoy is constructed based on rigid body dynamics theory. The multibody system determines whether there is a virtual connection relationship between any two ships by combining the distance between them, and generates the corresponding interaction forces and moments.

[0009] The system acquires the real-time navigation status of each vessel and uses a multibody system to generate a corresponding risk state vector based on the navigation status.

[0010] Collect data from each ship i angular velocity of sway Liquid level height h i and acceleration along the plane ( a xi , a yi ), and compare it with its own risk state vector X risk,i And other ships with virtual connections. j Risk state vector X risk,j Together as a risk prediction vector for current ships D i .in, i =1… n , n This indicates the number of ships in the formation.

[0011] The risk prediction vector for each vessel is input into a trained time-series-based risk prediction model, WXS, to generate the corresponding risk prediction result for each vessel. And utilize the fusion weights determined based on the entropy weight method. w i according to Generate global risk for the entire ship formation .

[0012] when If the preset safety threshold is exceeded, a risk is identified; at this point, the overall formation status is determined. s and risk level R c Query a preset set of experience for control strategies to obtain an optimal control strategy. B* , B* ={ b i}, i =1… n ; and when the yaw and / or pitch angle of any vessel exceeds the safety threshold, according to B*The balance control command is sent to each ship;

[0013] The global state of the formation includes the position s of each ship, x the speed y of the formation, and the risk prediction vector v of each ship. D i The risk level R c is evaluated according to the global risk predicted by the WXS model. b i represents the control action of the i-th ship, i ; Δ v i and represent the speed and displacement vectors of the counterweight in the ship, respectively.

[0014] The present application includes a computer program product comprising a computer program. When the computer program is executed by a processor, the balance regulation method of the LNG ship formation integrating virtual connection and distributed prediction as described above is realized; and the imbalance risk of the formation as a whole is predicted according to the real-time state of each ship in the formation, and adaptive control actions are generated to maintain the balance of the ships.

[0015] The present application includes a ship formation balance control device comprising a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the computer program, the balance regulation method of the LNG ship formation integrating virtual connection and distributed prediction as described above is realized; and the imbalance risk of the formation as a whole is predicted according to the real-time state of each ship in the formation, and adaptive control actions are generated to maintain the balance of the ships.

[0016] The technical solution provided by the present application has the following beneficial effects:

[0017] ​​The present application aims at the balance and control problems of each ship in the LNG ship formation caused by liquid sloshing in the storage tank and the influence of adjacent ships during navigation, constructs a virtual connection relationship between ships through high-precision communication and positioning technology, analyzes the dynamics of the formation as a whole, and realizes multi-ship cooperative control. In the imbalance prediction, the present application deploys a distributed imbalance prediction algorithm, uses sensor data of each ship and shared information of the formation to establish a local prediction model, and fuses to generate a global prediction to predict imbalance risks in advance. In generating the control strategy, the present application innovatively uses the DQN reinforcement learning algorithm to optimize the active balance weight system of the ship, takes the balance degree and energy consumption as the objective function to obtain the optimal weight movement strategy, and combines the pre-trained control strategy experience set to realize rapid response. Through the cooperation of the above three, the present application can automatically coordinate the propulsion system, balance device and weight block action when the ship is unbalanced, efficiently restore the balance of the formation, significantly improve the navigation safety and stability of the LNG ship formation, and provide an innovative solution for dynamic balance control. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The step flowchart of the balance adjustment method of the LNG ship formation provided in embodiment 1 of the present application is fused with virtual connection and distributed prediction.

[0019] Figure 2 The flowchart of predicting the imbalance risk of the ship formation by using the WXS model in embodiment 1 of the present application.

[0020] Figure 3 The flowchart of generating the imbalance control strategy of different consent states by using the reinforcement learning algorithm in embodiment 1 of the present application.

[0021] Figure 4 The time sequence curve of the LNG liquid sloshing additional force in the X direction under the wind and wave load of different schemes in the test experiment.

[0022] Figure 5 The time sequence curve of the LNG liquid sloshing additional force in the Y direction under the wind and wave load of different schemes in the test experiment.

[0023] Figure 6 The time sequence curve of the LNG liquid sloshing additional moment in the X direction under the wind and wave load of different schemes in the test experiment.

[0024] Figure 7 The time sequence curve of the LNG liquid sloshing additional moment in the Y direction under the wind and wave load of different schemes in the test experiment.

[0025] Figure 8 The time sequence change curve of the liquid sloshing additional force and the additional moment of different schemes in the test experiment.

[0026] Figure 9 To test the time sequence comparison curve of local risk and global risk of ship formation in different schemes in the experiment.

[0027] Figure 10 To test the time sequence change curve of ship formation yaw angle and pitch angle in different schemes in the experiment.

[0028] Figure 11 To test the radar chart of comprehensive performance of different schemes in the experiment. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0030] Example 1

[0031] The technicians of the scheme of the present embodiment find that the traditional ship balance control technology focuses on single ship independent regulation and control, lacks consideration of overall cooperation of LNG ship formation, and is difficult to cope with the chain reaction caused by imbalance of a ship in the formation. In view of this problem, the scheme provided by the present embodiment combines the chain reaction between ships during navigation to construct a virtual connection relationship between the ships, and then performs dynamic analysis and cooperative control on all ships in the ship formation with virtual connection relationship as a whole. This scheme can quickly coordinate the multi-ship propulsion system and balance device when the ship is out of balance, greatly improving the balance recovery efficiency of the formation.

[0032] In terms of imbalance risk prediction, the existing technology relies on simple sensor monitoring or experience judgment, lacking systematic and forward-looking prediction mechanism. The technical scheme provided by the present embodiment acquires the data collected by the sensors deployed in all ships, and forms a global risk by fusing the local prediction results of each ship through a risk prediction model WXS based on time sequence data. This strategy not only can improve the prediction accuracy and misjudgment rate of the risk, but also can predict the imbalance risk 1-3 minutes in advance, so as to gain sufficient response time for control measures.

[0033] In the regulation of unbalanced risks, the embodiment adopts the existing ship active balancing weight system. Unlike the existing regulation strategy optimized in real time by algorithm; in order to improve the comprehensiveness and real-time performance of the optimization target, the embodiment adopts a brand-new strategy library matching strategy. In the generation process of each regulation strategy in the strategy library, the application adopts an optimization target considering the balancing effect and energy consumption, and adopts a reinforcement learning algorithm to optimize the regulation strategy of the counterweight system under different working conditions. The related regulation strategy can not only effectively compensate the unbalanced force generated by the LNG liquid sloshing of the ship, but also can reduce the energy consumption of the balancing weight system by 20%-30%, and the economic benefit is obvious.

[0034] Specifically, as Figure 1 indicated, the balancing regulation method of the LNG ship formation provided by the embodiment fuses virtual connection and distributed prediction, and includes the following processes:

[0035] I. Constructing a multi-body system of a ship formation

[0036] In order to accurately analyze the influence of the interaction between each ship in the formation on the unbalanced risk of the whole ship formation, the embodiment first constructs a multi-body system representing the LNG ship formation based on the rigid body dynamics theory. In the multi-body system constructed in the embodiment, whether there is a virtual connection relationship between any two ships is determined in combination with the distance between the two ships, and the corresponding interaction force and interaction torque are generated for any two ships with a virtual connection relationship. The multi-body system constructed in the embodiment can clearly determine the correlation between the interaction force between the ships and the motion parameters, and provide an accurate dynamics basis for subsequent cooperative control. i and j

[0037] In detail, in the embodiment, the generalized coordinate vector of each ship i is defined as ; wherein, i =1… n , n indicates the number of ships in the formation. Among them, ( x i , y i ) correspond to the plane coordinates of ship i ; x i and y i are the horizontal coordinate and the vertical coordinate of ship i in the plane coordinate system, respectively. In actual application, the corresponding plane coordinates can also be represented by the latitude and longitude of the ship. , , are the horizontal coordinate and the vertical coordinate of ship i ​The yaw, pitch, and heading angles are all measured. These data, characterizing the ship's real-time status, can be obtained through real-time monitoring using the ship's navigation and positioning system and various sensors.

[0038] In order to establish a system that includes the mass of each ship m i and moments of inertia in three dimensions I xi , I yi , I zi In this embodiment, considering the additional inertial forces and torques generated by the sloshing of liquid within the storage tanks during the LNG vessel's navigation, innovative additional dynamic parameters are introduced to construct the dynamic model of each vessel in the multibody system corresponding to the ship formation. In this embodiment, to effectively assess the vessel's imbalance risk, the vessel's dynamic model consists of two parts: the ship's translational dynamics equations and rotational dynamics equations.

[0039] Specifically, in a multi-body system, any first... i The translational dynamics equations of a ship can be expressed as follows:

[0040] ;

[0041] In the above formula, j Indicates with ships i Other vessels with virtual connection relationships; m i Indicates a ship i quality; F ij For ships j For ships i The interaction force vector; F li F is the vector of the additional force generated by the sloshing of LNG liquid. wi r is the environmental load force vector that includes wind, waves, and current. i For ships i The position vector, .

[0042] In this embodiment, the environmental load force vector F wi The Morrison formula can be used for calculation as follows:

[0043] ;

[0044] In the above formula, Indicates the density of seawater. C D Indicates the drag coefficient. C M Indicates the additional quality coefficient. Adenotes the wetted surface area of the ship, V denotes the displacement volume of the ship, ri denotes the velocity vector of the ship relative to the water flow; denotes the first derivative of v ri , i.e. the acceleration vector of the ship relative to the water flow.

[0045] The rotational dynamics equation of any i th ship is:

[0046] ;

[0047] wherein, is the interaction torque vector of the ship j on the ship i ; is the additional torque vector generated by the LNG liquid sloshing; is the environmental load torque vector; I i is the inertia matrix of the ship i , I i =diag( I xi , I yi , I zi ); is the angular velocity vector of the ship i , .

[0048] In the rotational dynamics equation of the ship constructed in the embodiment, the skilled person also adopts an innovative micro-element integration strategy to quantitatively analyze the additional force vector F li and the additional torque vector generated by the LNG liquid sloshing carried in the ship, and the corresponding calculation formula is as follows:

[0049] ;

[0050] In the above formula, is the LNG liquid density, V l is the LNG liquid volume in the ship tank; v l is the liquid micro-element velocity vector, r cl is the position vector of the liquid micro-element relative to the center of mass of the ship; denotes the material derivative.

[0051] wherein the liquid micro-element velocity vector v l can be obtained by solving the Navier-Stokes equation considering the free surface sloshing:

[0052] ;

[0053] In the above formula, p l is the liquid pressure; u l is the liquid dynamic viscosity, g is the gravity acceleration vector; represents the gradient calculation symbol.

[0054] In the translation dynamics equation and rotation dynamics equation of the ship constructed in the embodiment, the interaction force vector F ij and the interaction torque vector reflect the virtual connection relationship between any two ships in the ship formation; the virtual connection relationship needs to be determined comprehensively in combination with the relative position and attitude of any two ships in the formation, and in actual application, the interaction force vector F i and the interaction torque vector j between any two ships ij and can be calculated through the following virtual force field model:

[0055] ;

[0056] ;

[0057] In the above formula, k ij and c ij are the elastic coefficient and the damping coefficient of the virtual connection respectively; k tij and c tij are the torsional elastic coefficient and the torsional damping coefficient of the virtual connection respectively; r i and r j are the position vectors of the ships i and j respectively; and are the yaw angles of the ships i and j respectively; d ij is the actual distance between the two ships; d th is the preset effective action distance threshold.

[0058] The analysis of the above process shows that in the multi-body system of the embodiment, when the distance between any two ships exceeds the threshold value, it is considered that the influence between the two is small, and a virtual connection relationship does not need to be established, and the two belong to two independent single-body systems. When the distance between the two ships is less than the preset distance threshold, it is indicated that the sailing and action of the two will interact through the water body, and the embodiment establishes a corresponding "virtual connection" between the two and generates a corresponding interaction force vector F ij and an interaction torque vector to jointly analyze the states of the two.

[0059] In addition, the dynamic model of the multi-body system constructed in the embodiment not only considers the traditional rigid body dynamics factors, but also innovatively introduces the LNG liquid sloshing additional dynamics parameters and the ship interaction model based on the virtual connection, so that the dynamic characteristics of the formation containing multiple LNG ships under complex working conditions can be more accurately described.

[0060] II. Prediction of imbalance risk

[0061] Based on the constructed multi-body system of the ship formation, as shown in the figure, the embodiment can obtain the sailing state of each ship Figure 2 in real time, and then generate a corresponding risk state vector i according to the sailing state by using the multi-body system X risk,i . The data format of the risk state vector i of any i-th ship is as follows: X risk,i

[0062]

[0063] In actual application, the translational dynamics equation and the rotational dynamics equation of the ship can be directly obtained; E li , which represents the sloshing energy of the ship, and the calculation formula is as follows: Finally, the embodiment can extract the yaw rate and the trim acceleration processed by the fourth-order Butterworth low-pass filter in the dynamic model.

[0064] In the embodiment, in order to more accurately realize the analysis and prediction of the ship state and the imbalance risk, the state data (i.e. i X risk,i ) of each ship, the state of the formation, and the state of all associated ships having a virtual connection relationship with i are collectively used as input data for evaluating the imbalance risk of the ship.​​​

[0065] Specifically, the embodiment collects the sway angular velocity i , liquid level height , and in-plane acceleration h i of each ship a xi , a yi , and takes it as the risk prediction vector of the current ship together with the risk state vector X risk,i of itself and the risk state vectors X risk,j of other ships with virtual connection relationship D i , wherein .

[0066] Next, the embodiment inputs the risk prediction vector of each ship into a trained risk prediction model WXS based on time series data to generate the corresponding risk prediction result of each ship.

[0067] In the scheme of the embodiment, the risk prediction model is the key to realize local analysis of the unbalanced risk of each ship. The risk prediction model (WXS model) of the embodiment is a network model trained based on machine learning, which can establish the mapping relationship between “dynamic parameters-risk level” through massive historical data of ship operation. In actual application, the embodiment can adopt a cross-entropy comprehensive prediction model based on wavelet analysis, which can predict the current risk prediction result D i of the ship x t according to the time series data i of the input risk prediction vector

[0068] In actual application, the risk prediction model will first perform wavelet decomposition on the input original collected data, and then perform inference analysis according to the decomposition result to obtain the prediction result. In the embodiment, the wavelet decomposition formula used by the risk prediction model is as follows:

[0069] ;

[0070] In the above formula, t denotes time; and are the scale function and wavelet function, respectively; a j,k is the first j ​​approximate coefficients of the layer at the k-th moment; d j,k approximate coefficients of the layer at the k-th moment; j approximate coefficients of the layer at the k-th moment; k detail coefficients of the layer at the k-th moment, k representing a time axis index of the time series data.

[0071] In the further optimized scheme of the embodiment, the information ship transmission can be realized by the special communication module carried by each ship. C i i =1, 2, …, n) In order to ensure the data accuracy, the collected data is first processed by a fourth-order Butterworth low-pass filter, and the transfer function thereof is:

[0072]

[0073] In the above formula, s representing the frequency domain characteristics of the data used; s k are poles of the filter, which removes high-frequency noise interference and ensures the accuracy and stability of the data. Subsequently, a virtual connection relationship between ships is constructed based on a self-organizing network protocol, forming a dynamic topology network, and the communication delay between nodes in the network meets the real-time requirement. On this basis, d j,k The following dynamic threshold function is used for processing to enhance the feature signal-to-noise ratio:

[0074]

[0075] wherein, representing the enhanced detail coefficient; representing a dynamic threshold parameter, is a standard deviation estimate of the detail coefficient, N is the data length of the time series data.

[0076] In the training phase of the risk prediction model constructed in the embodiment, the cross-entropy loss L CE is used as the loss function for model optimization:

[0077]

[0078] In the above formula, p c representing the true probability distribution of the ship imbalance risk; is the class probability distribution predicted by the model; C is the total number of risk categories.

[0079] ​​​​​The risk prediction model in this embodiment consists of multiple base models. During the training phase, the risk prediction model uses dynamically updated adaptive weights. The weight update formula is as follows:

[0080] ;

[0081] In the above formula, T Indicates the number of iterations; and They are the first T and T The risk prediction model included in the +1 round of iterations m The weights of each base model. This represents the learning rate. In this embodiment, an adaptive learning rate related to the prediction error volatility is used in practical applications:

[0082] ;

[0083] In the above formula, Based on the learning rate, Var ( Error ) represents the variance of the prediction error.

[0084] In practical applications, each vessel can upload its own status data to a central management module on a command ship. The central management module then assesses the imbalance risk of each vessel based on data from different sources. The prediction task can also be performed by embedded modules deployed in each ship, which can then combine their own state data to predict their respective imbalance risks. The prediction results are then uploaded to the central management module, whereby the command ship performs a global analysis of the risk of imbalance in the ship formation.

[0085] Finally, this embodiment utilizes the fusion weights determined based on the entropy weight method. w i according to Generate global risk for the entire ship formation The calculation formula is as follows:

[0086] ;

[0087] III. Regulation of Imbalance Risk

[0088] Imbalance risk of individual ships predicted by the WXS model and the overall risk of formation These respectively reflect the probability distributions of the local and global prediction results for ship imbalance risk. When If the preset safety threshold is exceeded, the ship formation is deemed to have an imbalance risk. Specifically, in practical applications, the safety threshold set in this embodiment is 0.4. >0.4 indicates that there is a risk, and when ≤0.4, it is determined that there is no risk.

[0089] On this basis, in order to facilitate management and regulation, the embodiment can further discretize the global risk predicted by the WXS model, and further determine the risk level of the imbalance risk of the formation R c In a typical scheme provided by the embodiment, the mapping relationship between the global risk and the risk level R c is as follows:

[0090] When 0.5≥ >0.4, R c =R1, indicating low risk;

[0091] When 0.6≥ >0.5, R c =R2, indicating medium risk;

[0092] When >0.6, R c =R3, indicating high risk.

[0093] When the prediction result indicates that there is a risk, the embodiment first converts the prediction result of the global risk into the corresponding risk level R c , and then according to the global state s of the formation and the risk level R c , queries a preset regulation strategy experience set to obtain an optimal regulation strategy B* , B* ={ b i}, i =1… n ; and when any of the yaw angle and the pitch angle of any ship exceeds the safety threshold value, balance regulation instructions are issued to each ship according to B* In the embodiment, the safety threshold values of the yaw angle and the pitch angle of the ship are both set to 3°. In the regulation process, each ship also feeds back the attitude data in real time to form a closed loop control, so as to ensure that the LNG ship formation balance state is restored within 30 seconds, and the safe and stable navigation under complex sea conditions is ensured.

[0094] In the technical solution provided in this embodiment, the control strategy experience set is generated by a reinforcement learning algorithm based on the historical navigation data of the ship formation, representing the global state of any formation. s and risk level R c With regulation strategy B* A dataset containing mapping relationships. This mapping relationship is represented as: s + R c → B* .

[0095] Among them, the global state of the formation s Including formation position ( x , y Formation speed v and risk prediction vectors for each vessel D i The formation position can be determined by the position of the command ship within the formation, or by the approximate center of the positions of all ships in the formation. Risk Level R c Global risk predicted by the WXS model Evaluation and generation. Under the optimal control strategy. B* middle, b i Indicates the first i The control actions of individual ships, ;Δ v i and These represent the velocity and displacement vectors of the counterweight in the ship, respectively.

[0096] For ship active balancing counterweight systems, such as Figure 3 As shown, in this embodiment, when using the DQN reinforcement learning algorithm to optimize the ballast strategy and then construct the control strategy, the ship's pitch angle is used as the reference. Heading angle The sum of absolute values and energy consumption of counterweight movement E Minimization is used as the optimization objective function. In the reinforcement learning framework, the state space is defined to include the ship's position, attitude, and risk, and the action space is defined to include the three-dimensional movement parameters of the counterweight. A reward function is designed to balance the control effect and energy consumption. Through experience playback and deep neural network training, the optimal counterweight movement strategy is output to ensure that the heel and roll angles are controlled within ±3° and energy consumption is reduced by 20%-30%.

[0097] State space:

[0098] In the reinforcement learning framework of constructing the regulation strategy experience set, the state space is the basis for the DQN algorithm to perceive the environment, and needs to fully reflect the current balance state and risk characteristics of the ship. This embodiment contains three dimensions of information state space for each global state s The definition of each global state in the three-dimensional information state space is as follows:

[0099] ;

[0100] Wherein, , .

[0101] Action space and adaptation mechanism:

[0102] The action space defines the adjustment operations that can be performed by the counterweight system. In the scheme provided in this embodiment, the granularity of the regulation action will be dynamically adjusted according to the output risk level R c , realizing the response logic of “the higher the risk, the more decisive the action”. The mathematical definition is , the moving unit of the counterweight block is m, and the speed unit is m / s.

[0103] When R c =R1, it belongs to a low-risk scenario. At this time, the ship's attitude fluctuates less, and the probability of predicting future imbalance is reduced. The action space adopts a fine adjustment mode. The purpose of this design is to maintain balance while minimizing energy consumption and avoiding unnecessary frequent adjustment. The action space used in this state is:

[0104] .

[0105] When R c =R2, it belongs to a medium-risk scenario. At this time, the probability of predicting imbalance is higher, and a balance needs to be struck between response speed and energy consumption; the adjustment range of the action space is expanded. The action space used in this state is:

[0106] .

[0107] When R c =R3, it belongs to a high-risk scenario. At this time, the ship has approached the imbalance threshold, and the model predicts that the probability of imbalance in the near future is higher. The action space switches to the emergency adjustment mode, prioritizing rapid recovery of balance and appropriately relaxing the energy consumption limit. The action space used in this state is:

[0108] .

[0109] In the adaptive action space designed in this embodiment, the total dimension of the action space is 3 3 x 3 = 81 (covering all combinations of 3 risk levels), ensuring that the algorithm has sufficient adjustment freedom in different risk scenarios.

[0110] Reward function:

[0111] The reward function guides policy optimization by quantifying the pros and cons of actions. The reward function of this design fully integrates parameters to achieve multi-objective collaborative optimization. Specifically, the reward function is:

[0112] ;

[0113] In the above formula, R balance and R energy represent the balance reward and energy consumption reward, respectively; α and β represent the weights of the balance reward and energy consumption reward, respectively; and represent the ship roll angle and pitch angle after executing the control action, respectively; and represent the penalty terms determined according to the standard deviation of the roll angle and pitch angle; for example, in high sea conditions (Δ is large), even if the inclination angle is small, rapid attitude change will be given negative reward to suppress violent shaking. E is the incremental energy consumption of the control action, ; wherein, E and are the energy consumptions of the ship's active balance weight system before and after control. represents the risk adaptation coefficient determined according to the risk level. In this embodiment, for the three risk levels, the values of are 0.8, 1.0 and 1.2, respectively, to appropriately relax the energy consumption limit in high-risk states.

[0114] The neural network used in the reinforcement learning framework of this embodiment consists of an input layer, a risk feature layer, a fusion layer, and an output layer. The number of neurons in the input layer corresponds to the dimension of the state vector s , and the original data is standardized and transmitted into the network. The risk feature layer contains 32 neurons, which are specifically designed to process R c, the risk features are extracted by LeakyReLU activation function. The fusion layer contains 128 neurons and is used to fuse the posture and risk features, and then the hidden correlations between parameters are mined through ReLU activation function. For example, in the state of "high risk + rapid rolling", a more aggressive counterweight strategy needs to be matched; the output layer contains 81 neurons corresponding to the action space, and outputs the action value Q s b

[0115] In addition, the network loss function of the embodiment scheme introduces a weighting factor of risk level to improve the training weight of high-risk samples:

[0116]

[0117] wherein, y t is the target Q value, is the risk weight, represents the Q value corresponding to the action in the current state; represents the model parameters of the neural network.

[0118] Experience replay and exploration strategy:

[0119] Partitioned storage and sampling of experience pool: the experience pool is divided into D 1, D 2, D 3 three sub-pools according to the risk level, respectively storing the interaction data in the corresponding scene s t , b t , r t , s t+1 . When sampling, it is not randomly extracted, but weighted according to the probability of risk occurrence in actual navigation (high-risk samples account for ≥40%), to ensure that the algorithm has enough learning samples for dangerous scenes and avoid the failure of strategies in high-risk scenes due to the high proportion of low-risk samples.

[0120] To balance "using known optimal actions" and "exploring new actions", the value is dynamically adjusted according to the risk level: under the three risk levels, the values are 0.2, 0.4 and 0.6 respectively. When the risk is low, =0.2 (mainly for utilization), and when the risk is high, =0.6 (increasing exploration to cope with complex situations). This strategy avoids missing the optimal solution due to excessive reliance on old experience in emergency scenarios.

[0121] ​​​​In the embodiment, according to the platoon global state s and the risk level R c , the process of querying the preset regulation strategy experience set to obtain an optimal regulation strategy is as follows: first, a plurality of regulation strategy experience sets with the same risk level as the current prediction structure are selected from the regulation strategy experience set; then, the s values of each regulation strategy are matched with the s values of the current ship; finally, the regulation strategy with the highest matching degree is selected as the optimal regulation strategy B* .

[0122] After obtaining the optimal regulation strategy B* , the embodiment executes the regulation strategy in the following manner:

[0123] When the yaw angle or the pitch angle of the ship exceeds the limit and triggers an imbalance alarm, the control instruction set B* is generated according to the optimal regulation strategy ; wherein the propulsion system thrust adjustment instruction u propulsion The propeller speed n and the steering angle are determined through fuzzy inference rules, and the specific relationship can be expressed as:

[0124]

[0125] The control instruction set U is sent to each ship with sub-millisecond delay through the time-sensitive network (TSN) between ships. After the u propulsion is received, the propeller speed is adjusted according to n , and the steering is changed according to ; the balance fin changes the attack angle according to the u fit in the fin instruction α fin to generate a righting moment; the counterweight moves quickly along the path planned according to the configuration instruction u weight at the speed of v * ( t ). During the regulation process, each ship forms a closed-loop control by feeding back the attitude data in real time through sensors; the regulation error is calculated according to the error function; and a new round of control instructions , and are adjusted based on the PID control algorithm:

[0126]

[0127] In the above formula, E 1、E 2 and E 3 are the control errors of the propulsion system, the balance fin and the ballast, respectively; K p1 , K p2 and K p3 are the proportional coefficients of the propulsion system, the balance fin and the ballast, respectively; K i1 , K i2 and K i3 are the integral coefficients of the propulsion system, the balance fin and the ballast, respectively; K d1 , K d2 and K d3 are the differential coefficients of the propulsion system, the balance fin and the ballast, respectively; each coefficient can be determined by off-line simulation optimization. By continuously adjusting the control instructions, the PID controller can ensure that and are met within t≤30s, restoring the balance state of the LNG ship formation and ensuring the safe and stable navigation of the LNG ship formation in complex sea conditions.

[0128] In addition, after each successful completion of the regulation in the embodiment, a new regulation strategy can be generated according to the regulation process and stored in the regulation strategy experience set.

[0129] Embodiment 2

[0130] The balance regulation method of the LNG ship formation combining virtual connection and distribution prediction provided in Embodiment 1 is essentially a data processing method. In order to better apply the scheme, the embodiment further provides a computer program product, a storage medium and a ship formation balance control device.

[0131] The computer program product provided in the embodiment includes a computer program which, when executed by a processor, implements the balance regulation method of the LNG ship formation combining virtual connection and distribution prediction as in Embodiment 1; and further predicts the imbalance risk of the whole formation according to the real-time state of each ship in the formation, and generates adaptive regulation actions to maintain the balance of the ships.

[0132] The storage medium provided in the embodiment has a computer program stored therein. The computer program, when executed by a processor, implements the balance regulation method of the LNG ship formation combining virtual connection and distribution prediction as in Embodiment 1; and further predicts the imbalance risk of the whole formation according to the real-time state of each ship in the formation, and generates adaptive regulation actions to maintain the balance of the ships.

[0133] The ship formation balance control device provided by the embodiment comprises a memory, a processor and a computer program stored in the memory and running on the processor. When the processor executes the computer program, the balance adjustment method of the LNG ship formation with fusion virtual connection and distribution prediction is realized as in Embodiment 1; and then the imbalance risk of the whole formation is predicted according to the real-time state of each ship in the formation, and adaptive regulation actions are generated to maintain the balance of the ships.

[0134] The ship formation balance control device provided by the embodiment. In essence, it is a computer device. In actual application, the computer device can be an embedded device, or an independent computer device, such as a notebook computer, a tablet computer, a desktop computer, or a medium / large computer device capable of executing a computer program, such as a rack server, a blade server, a tower server, or a cabinet server (including a standalone server or a server cluster composed of multiple servers). Then the MRI output from the front end is optimized in the back end.

[0135] The computer device of the embodiment at least includes but is not limited to a memory and a processor which can be connected to each other through a system bus. In the embodiment, the memory (i.e. a readable storage medium) includes a flash memory, a hard disk, a multimedia card, a card-type memory (e.g. an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as a hard disk or a memory of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. of the computer device. Of course, the memory can include both the internal storage unit and the external storage device of the computer device. In the embodiment, the memory is usually used to store the operating system and various application software installed on the computer device, etc. In addition, the memory can also be used to temporarily store various data that have been output or will be output.

[0136] The processor in some embodiments can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor is usually used to control the overall operation of the computer device.

[0137] Performance test

[0138] To verify the effectiveness of the LNG ship formation balanced adjustment method provided by the fusion of virtual connection and distribution prediction, the technical personnel simulate the related scheme, design multiple groups of comparative experiments in the experimental process, construct an LNG ship formation model conforming to the actual navigation scene through the simulation platform, and quantitatively verify the risk prediction accuracy, balanced regulation effect and energy consumption optimization capability of the scheme.

[0139] I. Experimental content

[0140] 1.1, Simulation parameter setting

[0141] In this experiment, a simulation platform based on MATLAB and Python is used as a joint simulation environment to construct a formation model containing three LNG ships (the ship parameters refer to the design of a 170,000 cubic meter LNG transport ship), which integrates rigid body dynamics module, virtual connection force field module, LNG liquid sloshing module and reinforcement learning regulation module. Some simulation parameters are as follows:

[0142] (1) Ship basic parameters: single ship mass is 50000 kg, roll inertia moment is 10 8 kgm 2 , pitch inertia moment is 2x10 8 kgm 2 , LNG density in the tank is 425 kg / m 3 , tank volume is 1000 m 3 .

[0143] (2) Virtual connection parameters: elastic coefficient is 10 4 N / m, damping coefficient is 5x10 3 Ns / m, torsional elastic coefficient is 5x10 3 Nm / rad, torsional damping coefficient is 2x10 3 Nms / rad, effective action distance threshold is 150 m.

[0144] (3) Environmental load parameters: wind and wave load is simulated by using the Morison formula, wind is 15 m / s, effective wave height is 3 m, wave period is 8 s, and current velocity is 2 m / s.

[0145] (4) Risk prediction model parameters: the WXS model uses 3-layer wavelet decomposition, the dynamic threshold parameter is 0.8, the basic learning rate is 0.01, and the number of categories of cross-entropy loss function is 3 (corresponding to three risk levels).

[0146] (5) Reinforcement learning parameters: DQN algorithm experience pool capacity is 10 5 , target network update interval is 100 steps, discount factor is 0.95, and exploration rate is dynamically adjusted according to risk level.

[0147] 1.2, Control group setting

[0148] The experiment takes the scheme of the application as an experimental group, and sets two control groups, so as to verify the advantages of the scheme of the application through comparative experiments. The experimental group and the control groups are as follows:

[0149] ① Experimental group (ship formation 1): using the method of the application (fusion of virtual connection + WXS risk prediction + reinforcement learning regulation)

[0150] ② Control group (ship formation 2): traditional single-ship independent regulation method (without virtual connection, only based on single-ship attitude feedback regulation)

[0151] ③ Control group (ship formation 3): formation regulation method without risk prediction (retaining virtual connection, only passive regulation after the attitude angle exceeds the threshold value)

[0152] Each group of experiments lasts for 300 seconds, the sampling frequency is 10 Hz, and the average value is taken for 5 repeated experiments.

[0153] II. Experimental results and analysis

[0154] 2.1, Verification of wind and wave coupling effect

[0155] Firstly, the time sequence curves of the LNG liquid sloshing additional force in the X direction of different ships in the formation under the wind and wave load are tested, and the experimental results are as shown in Figure 4 .

[0156] According to the data in the figure, the additional force X component of ships 1, 2 and 3 all show obvious periodic fluctuation characteristics, and the fluctuation period is basically the same as the coupling period of sea wave excitation and ship attitude response, reflecting the strong correlation between liquid sloshing and environmental load, ship movement. The amplitude of the additional force X component of ship 3 is higher than that of ship 1 and ship 2 as a whole, and the average amplitude is close to 2000 N, which is due to the fact that ship 3 is in the rear position in the formation, and is affected by the superposition of the flow field disturbance of the front ship and the attitude coupling effect of itself. The indirect effect of the following ship on the ship 1 as the formation leader is relatively weak, and the amplitude of the additional force X component is relatively small, about 1000 N on average.

[0157] Then the time sequence curves of the LNG liquid sloshing additional force in the Y direction of different ships in the formation under the wind and wave load are tested, and the experimental results are as shown in Figure 5 .

[0158] According to the data in the figure, the additional force Y component of the three ships has both synchronization and difference in the time dimension. The synchronization reflects the correlation of the movement response of the ships in the formation, and the difference is caused by the different space positions, initial roll attitude and liquid sloshing phase of each ship in the formation. The distribution characteristics of the liquid sloshing additional force Y component bring challenges to the lateral balance regulation of the ship formation, and further verify the necessity of the virtual connection force field and the distributed prediction method proposed in the application in coordinating the roll direction sloshing response of multiple ships.

[0159] Next, the experiment continues to test the LNG liquid sloshing additional moment of different ships in the formation under the wind and wave load in the X direction, and the experimental results are shown in Figure 6

[0160] According to the data in the figure, the additional moment X component of ship 3 has the largest amplitude, with a positive peak of about 120000 Nm and a negative peak of about 100000 Nm. The amplitude of ship 2 is smaller, and the amplitude of ship 1 is the smallest. The change trend of the additional moment X component of the three ships has synchronization but phase difference, which is caused by the difference in space position, initial roll state and liquid sloshing phase of each ship.

[0161] The experiment also tests the LNG liquid sloshing additional moment of different ships in the formation under the wind and wave load in the Y direction, and the experimental results are shown in Figure 7

[0162] According to the data in the figure, it can be seen that the additional moment Y component of ship 2 has a relatively larger amplitude, with a positive peak of about 42000 Nm, and the amplitudes of ship 1 and ship 3 are relatively small. This distribution increases the complexity of the roll balance regulation of the formation, and further illustrates the importance and effectiveness of the method of the application in coordinating the roll response of multiple ships.

[0163] Finally, the experiment draws the time sequence change curve of the liquid sloshing additional force and additional moment, and the experimental results are shown in Figure 8

[0164] According to the data in the figure, it can be seen that the additional moment of ship 3 is larger as a whole (after 10 times scaling), with a peak of about 12000 Nm. As for the additional force, the additional force of ship 2 is relatively more prominent, with a peak of about 4000 N. The change trend of the additional force and additional moment of each ship has both certain cooperation and difference due to the position of the ship in the formation, liquid sloshing characteristics and other factors, which also reflects the application value of the method of the application in the coordinated regulation of the liquid sloshing response of multiple ships.

[0165] 2.2, risk prediction accuracy

[0166] ​​​The experiment further verifies the time sequence comparison of local risk and global risk of ship formation in different schemes, and draws the corresponding change curve as shown in Figure 9 The comparison of risk prediction performance of the present application and the second control group scheme is shown in chart 1:

[0167] Table 1: Comparison of risk prediction performance indicators

[0168]

[0169] Analysis of the risk prediction indicators in Table 1 shows that:

[0170] The global risk dynamically changes in the interval [0.2, 0.8], with R1 accounting for 32%, R2 accounting for 45%, and R3 accounting for 23%. The coincidence degree with the actual attitude threshold value is 92%, proving that the WXS model can accurately map the relationship between ship state and risk level. In addition, the risk prediction accuracy of the experimental group is 90.5%, which is 22.3 percentage points higher than that of the control group 2 (without risk prediction) of 68.2%. The false positive rate (no risk misjudgment as risk) is reduced to 5.3%, and the false negative rate (high risk misjudgment as low risk) is reduced to 4.2%. At the same time, the global risk fusion weight based on the entropy weight method (ship 1: 0.32, ship 2: 0.35, ship 3: 0.33) matches the stress weight of the ship in the formation, avoiding the deviation of single ship risk dominating the global judgment.

[0171] 2.3, balance control effect verification

[0172] The experiment further tests the time sequence changes of the yaw angle and pitch angle of the ship formation in different schemes, and the obtained change curve is as shown in Figure 10 The balance control effect under the same state is shown in Table 2.

[0173] Table 2: Comparison of attitude balance control effect indicators

[0174]

[0175] Analysis of the control effect indicators in Table 2 shows that:

[0176] The average amplitude of yaw angle of the experimental group is 2.8°, and the average amplitude of pitch angle is 2.5°, both of which are controlled within the safety threshold of 3°; the average amplitude of yaw angle of the control group 1 (single ship regulation) reaches 4.5°, and the pitch angle reaches 3.8°, and the proportion of time exceeding the threshold is 18.3 %; the proportion of time exceeding the threshold of the control group 2 (without risk prediction) is 12.7 %, which proves that the invention can effectively suppress the attitude over-limit through the "prediction-regulation" closed loop; at the same time, when a short-time strong load causes the attitude to approach the threshold, the experimental group can restore the attitude to within 1° within 15-20s, which is about 50% faster than the control group 1 (35-40s) and about 30% faster than the control group 2 (25-30s), which reflects the rapid response advantage of the reinforcement learning regulation strategy; the attitude standard deviation of the experimental group 3 (yaw angle: 0.52°, pitch angle: 0.48°) is reduced by 58%-55% compared with the control group 1 (yaw angle: 1.23°, pitch angle: 1.05°), which proves that the virtual connection can realize the cooperation of multiple ship attitudes and avoid the problem of "single ship stability and formation shaking".

[0177] 2.4, verification of energy consumption optimization effect

[0178] The energy consumption results of different schemes are further analyzed in this experiment, as shown in Table 3.

[0179] Table 3 Comparison of energy consumption and comprehensive performance index

[0180]

[0181] According to the energy consumption index in Table 3, the total energy consumption of the experimental group balance weight system is 2850kJ, which is reduced by 30.8% compared with the control group 1 (4120kJ) and by 22.5% compared with the control group 2 (3680kJ), because the energy consumption weight in the reinforcement learning reward function is 0.3, which moderately relaxes the energy consumption limit in high-risk situations and prioritizes energy saving in low-risk situations, achieving a coordinated optimization of balance and energy consumption. At the same time, the experimental group is superior to the control groups in four core indicators: risk prediction accuracy (90.5%), balance recovery speed (15-20s), attitude control precision (maximum amplitude 2.8°), and energy consumption efficiency (2850kJ). The radar chart comprehensive score is 42% higher than the control group 1 and 27% higher than the control group 2, which proves the systematic advantage of the invention method.

[0182] 2.5, comparison of comprehensive performance of different schemes

[0183] According to the above experimental data analysis, this experiment draws a multi-dimensional performance radar comparison chart of the ship formation of the invention and the traditional ship formation regulation scheme, to intuitively compare the performance of the research method and the traditional method in the four core indicators of risk prediction accuracy, balance recovery speed, attitude control precision, and energy consumption efficiency. The results are shown in Figure 11 .

[0184] According to the data in the figure, the performance of the scheme (blue area) is better than that of the traditional method (red area) in each index, and the advantages in risk prediction accuracy and balance recovery speed are remarkable, which fully reflects the comprehensive performance advantages of the method in the balance adjustment of the ship formation.

[0185] The above merely describes preferred embodiments of the present application and is not intended to limit the present application, and any modification, equivalent replacement, and improvement within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for balanced regulation of LNG ship formation with fusion of virtual connection and distribution prediction, characterized in that, It comprises: A multi-body system representing LNG ship formation is constructed based on rigid body dynamics theory; The multi-body system determines whether there is a virtual connection relationship between any two ships according to the distance between them, and generates the corresponding interaction force and interaction torque; acquiring each ship i a real-time sailing status, generating a corresponding risk status vector according to the sailing status by using the multi-body system X risk ; collecting a sway angular velocity i , a liquid level height , a liquid level height h i and a planar acceleration a xi , a yi of each ship, X risk,i and a risk state vector j of other ships X risk,j with a virtual connection relationship as a risk prediction vector D i of the current ship; i =1… n , n represents the number of ships in the formation; input the risk prediction vector of each ship into a trained risk prediction model WXS based on time series data, to generate the corresponding risk prediction result and the fusion weight determined based on the entropy weight method w i According to generate a global risk ; When the preset safety threshold is exceeded, it is determined that there is a risk; at this time, according to the platoon position (v x , y ), the platoon speed v and the risk prediction vector D of each ship i , the platoon global state s and the risk level R c , a preset control strategy experience set is queried to obtain an optimal control strategy B* , B* { b i}, i =1… n ; and when the yaw angle and / or the pitch angle of any vessel exceeds a safety threshold, according to B* issuing a balancing control instruction to its vessels; wherein, b i denotes the control action of the i th ship, ; Δ v i and denote the velocity and displacement vectors of the ballast in the ship, respectively.

2. The method of claim 1, wherein, Let the generalized coordinate vector of each ship be defined as ; the moments of inertia in three dimensions are respectively: I xi , I yi , I zi ; in,( x i , y i Corresponding ships i Plane coordinates; , , For ships i Yaw angle, pitch angle and heading angle: The translational dynamics equation of any i-th ship in the multi-body system is: i The translational dynamics equation of any i-th ship in the multi-body system is: ； in, j Indicates with ships i Other vessels with virtual connection relationships; m i Indicates a ship i quality; F ij For ships j For ships i The interaction force vector; F li F is the vector of the additional force generated by the sloshing of LNG liquid. wi r is the environmental load force vector that includes wind, waves, and current. i For ships i The position vector, ; Any of the i The rotational dynamics equation for any of the ships is: ； wherein is the vector of the interaction forces and moments acting on the ship j is the vector of the interaction forces and moments acting on the ship i is the vector of the interaction forces and moments acting on the ship is the vector of the additional moments due to the LNG liquid sloshing is the vector of the environmental load moments i is the vector of the environmental load moments i is the vector of the environmental load moments i = diag( I xi , I yi , I zi ); is the vector of the angular velocity of the ship i is the vector of the angular velocity of the ship is the vector of the angular velocity of the ship 3. The method of balanced regulation of LNG fleet formation fusing virtual connection with distributed prediction according to claim 2, characterized in that: Interaction force vector F between any two vessels i and j Interaction force vector F between any two vessels ij and interaction torque vector Based on the virtual connection relationship and relative position, attitude information determination, and through the following virtual force field model calculation: ； ； In the above formulae, k ij and c ij are the virtual connection's elastic and damping coefficients, respectively; k tij and c tij are the virtual connection's torsional elastic and damping coefficients, respectively; r i and r j are the position vectors of the ship i and j , respectively; and are the yaw angles of the ship i and j , respectively; d ij is the actual distance between the two ships; d th is the preset effective action distance threshold.

4. The method of balanced regulation of LNG fleet formation fusing virtual connection with distributed prediction according to claim 3, characterized in that, The risk state vector of any nth ship generated using the multi-body system i X risk,i The data format is:​ ； In the above formula, E li represents the sway energy of the ship, ; u l is the kinematic viscosity of the liquid.

5. The method of balanced regulation of LNG fleet formation fusing virtual connection with distributed prediction according to claim 4, characterized in that, Additional force vector F li And additional moment vector The calculation formula is as follows: ； In the above formula, is the LNG liquid density, V l is the LNG liquid volume in the ship tank; v l is the liquid micro-element velocity vector, r cl is the liquid micro-element position vector relative to the ship center of mass; denotes the material derivative; Liquid cell velocity vector v l By solving the Navier-Stokes equations considering sloshing of the free surface: ； wherein p l is the liquid pressure; u l is the liquid dynamic viscosity, g is the gravitational acceleration vector; denotes a gradient operator.

6. The method of balanced regulation of LNG fleet formation fusing virtual connection with distributed prediction according to claim 5, characterized in that: The risk prediction model adopts a cross-entropy comprehensive prediction model based on wavelet analysis, which predicts the current risk prediction result of the ship according to the input risk prediction vector D i of the time sequence data x t ​​​ The wavelet decomposition formula used by the risk prediction model is as follows: ； In the above formula, t Indicates time; and These are the scaling function and the wavelet function, respectively. a j,k For the model number j The approximation coefficients of the layer at time k; d j,k For the model number j The layer in the first k The detail factor of each moment, k The time axis index represents the time series data; d j,k Processing is performed using a dynamic threshold function as follows to enhance feature signal-to-noise ratio: ； wherein, denotes the enhanced detail coefficient; denotes a dynamic threshold parameter, ; is a standard deviation estimate of the detail coefficients, N is a data length of the time series data; And / or, the risk prediction model employs a cross-entropy loss function in the training phase L CE : ； In the above formula, p c a true probability distribution representing the risk of imbalance of the ship; is a model-predicted class probability distribution; c represents an index of the risk class; C is the total number of classes of risk. And / or, the dynamic weight update formula used by the risk prediction model is: ； wherein, T denotes the number of iterations; and are the weights of the first T and T the first m base models included in the risk prediction model in the m th iteration; denotes the learning rate; and / or, Adopting adaptive learning rate related to the fluctuation rate of prediction error: ； In the above formula, is the base learning rate, Var Error represents the variance of the prediction error.​ 7. The method of balanced regulation of LNG fleet formation fusing virtual connection with distributed prediction according to claim 6, characterized in that: The regulation strategy experience set is a data set representing the global state of any formation generated by a reinforcement learning algorithm according to historical navigation data of the ship formation s and risk levels R c and the mapping relationship between the regulation strategy B* ​ and / or, in the reinforcement learning framework of building the set of regulation policy experiences, the definition of each global state in the state space s is: ； wherein , ; each action in the action space b i is: ； Reward function is: ； In the above formula, R balance and R energy respectively represent the balance reward and the energy consumption reward; α and β respectively represent the weights of the balance reward and the energy consumption reward; and respectively represent the ship yaw angle and the pitch angle after the execution of the control action; represents the global state after the execution of the control action; and respectively represent the penalty term determined according to the standard deviation of the yaw angle and the pitch angle; Δ E is the energy consumption increment generated by the control action; represents the risk adaptation coefficient determined according to the risk level.

8. The method of balanced regulation of LNG fleet formation fusing virtual connection with distributed prediction according to claim 7, characterized in that: Global risk predicted by the WXS model with risk level R c The mapping relationship is: When 0.5 ≥ > 0.4, R c = R1, indicating low risk; When 0.6 ≥ > 0.5, R c = R2, indicates medium risk; When > 0.6, R c = R3, indicates high risk; And / or, each action in the reinforcement learning framework adopts an adaptive action space according to the risk level, including: When R c = R1, ; When R c = R2, ; When R c = R3, .

9. A computer program product comprising a computer program, characterised in that, When the computer program is executed by the processor, the balance adjustment method of the LNG ship formation integrating virtual connection and distribution prediction is realized, and then the imbalance risk of the whole formation is predicted according to the real-time state of each ship in the formation, and adaptive control actions are generated to maintain the balance of the ships.

10. A vessel formation balancing control device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that: When the computer program is executed by the processor, the balance adjustment method of the LNG ship formation integrating virtual connection and distribution prediction is realized, and then the imbalance risk of the whole formation is predicted according to the real-time state of each ship in the formation, and adaptive control actions are generated to maintain the balance of the ships.

Images