Intelligent swarm-based typhoon autonomous tracking and observation system

Abstract

The application provides a typhoon autonomous tracking and observation system based on intelligent groups, comprising a multi-agent cluster, a trajectory tracking and prediction subsystem and a task autonomous planning subsystem; the multi-agent cluster is used for performing navigation, positioning and data acquisition tasks; the trajectory tracking and prediction subsystem is in communication connection with the multi-agent cluster, is used for receiving external typhoon prediction data, and inversely calculates and corrects a typhoon center position based on real-time feedback of on-site observation data of the multi-agent cluster to generate a tracking target point at a next time; the task autonomous planning subsystem abstracts a collaborative observation task of the multi-agent into a group game model and generates a task instruction for the multi-agent cluster. The application limits the use of external typhoon data, and combines multi-agent on-site measured data to track and observe typhoon, which lays a foundation for the multi-agent to continuously and actively track and observe typhoon under real conditions, and simultaneously provides a possibility for efficient on-site data collection of typhoon.

Description

Technical Field

[0001] This invention belongs to the field of multi-agent collaborative technology and relates to a typhoon autonomous tracking and observation system based on intelligent swarms. Background Technology

[0002] As one of the most destructive natural disasters, the accuracy of typhoon intensity and path forecasts directly impacts the effectiveness of disaster prevention and mitigation. Currently, improving forecasting capabilities faces a fundamental constraint: insufficient understanding of the fine internal structure of typhoons and their energy exchange processes with the lower ocean. This bottleneck stems from the difficulty in obtaining direct, three-dimensional, and high-resolution on-site observation data from the typhoon's core region.

[0003] Existing main observation methods, such as satellite remote sensing and limited shore-based buoy and ship observations, have inherent limitations in spatiotemporal resolution and data dimensionality. They are unable to effectively penetrate and monitor key physical parameters of the air-sea interface and upper ocean under the influence of typhoons, resulting in an inability to accurately characterize the coupling mechanism between ocean thermal structure response and typhoon intensity changes. Therefore, developing new observation technologies capable of penetrating deep into typhoon hazard areas and acquiring their three-dimensional fine structure data has become an urgent international need.

[0004] Utilizing clusters of unmanned intelligent devices for collaborative network observation is considered a revolutionary approach to overcome this predicament. Intelligent swarms composed of underwater gliders, autonomous underwater robots, and other similar devices possess enormous potential for autonomous, three-dimensional, and continuous tracking of typhoon paths, and are expected to fill critical data gaps.

[0005] However, the successful implementation of this technical approach heavily relies on the advanced autonomy and collaborative capabilities of multi-agent systems in complex and dynamic environments. Specifically, the observation swarm faces three core challenges: the dynamic variability of mission objectives, the significant heterogeneity of platform performance, and the strong uncertainty of the marine environment. Although theories such as game theory and multi-agent reinforcement learning offer possible directions for solving collaborative decision-making problems, existing methods often struggle to achieve an effective balance between individual autonomy and group cooperation, and between computational efficiency and policy optimality. Especially when dealing with sudden events such as abrupt typhoon path changes, communication disruptions, or equipment failures, how to achieve self-organization, self-adaptation, and continuous collaborative observation by the swarm remains an unresolved technical challenge.

[0006] Therefore, developing a comprehensive observation system that integrates high-precision environmental perception, intelligent trajectory prediction, and multi-agent autonomous collaborative decision-making is not only key to overcoming the challenge of typhoon intensity forecasting, but also a core driving force for promoting the development of next-generation intelligent ocean observation technology, and has significant scientific significance and application value. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing typhoon tracking and observation schemes, and to solve the problems of difficult underwater observation of typhoon processes and inaccurate typhoon intensity forecasts, by providing a typhoon autonomous tracking and observation system based on intelligent swarms. The two main objectives of this invention are: first, multi-agent dynamic autonomous typhoon tracking; and second, multi-agent autonomous observation task planning.

[0008] The technical solution adopted in this invention is as follows: A typhoon autonomous tracking and observation system based on intelligent swarms includes a multi-agent swarm, a trajectory tracking and prediction subsystem, and a mission autonomous planning subsystem; The multi-agent cluster consists of multiple heterogeneous or homogeneous intelligent devices with different performance characteristics, used to perform navigation, positioning and data acquisition tasks; The trajectory tracking and prediction subsystem is configured to communicate with the multi-agent cluster to receive external typhoon forecast data and, based on the real-time field observation data transmitted back by the multi-agent cluster, invert and correct the typhoon center position to generate the tracking target point for the next moment. The task autonomous planning subsystem is configured to: abstract the multi-agent collaborative observation task into a group game model, solve the model using a double cyclic equilibrium solution algorithm, output a set of agent strategies, the game situation of the agent strategies approximates Pareto Nash equilibrium, and generate task instructions for the multi-agent cluster according to the strategies. The trajectory tracking and prediction subsystem, the task autonomous planning subsystem, and the multi-agent cluster together constitute a closed-loop autonomous observation system.

[0009] Furthermore, the intelligent devices include underwater gliders, remotely controlled underwater robots, autonomous underwater robots, or unmanned surface vessels capable of diving; before the arrival of a typhoon, the intelligent devices are randomly deployed by the mother ship to sea areas where future typhoons may occur.

[0010] Furthermore, the trajectory tracking and prediction subsystem includes a typhoon external forecast module, a typhoon center calculation module, and a trajectory correction module; The external typhoon forecast module is used to acquire and preprocess typhoon trajectory forecast data from external meteorological stations; The typhoon center calculation module is configured to use multi-point positioning method to invert the coordinates of the center position of the typhoon at the air-sea interface based on the ocean current field data observed in real time by the multi-agent cluster. The trajectory correction module is configured to use the center position coordinates to correct the predicted center position coordinates in the forecast data through a dynamic weighted correction algorithm, and output the corrected center as the target of the multi-agent cluster's action at the next moment.

[0011] Furthermore, the multi-point positioning method specifically involves: based on ocean current data observed by multiple agents at different locations in the upper flow field induced by the typhoon, combined with a three-dimensional model of the typhoon-induced flow field, calculating the center coordinates of the typhoon at the air-sea interface using the principle of geometric positioning.

[0012] Furthermore, the dynamic weighted correction algorithm is configured as follows: Based on the historical trajectory prediction error sequence, a dynamic weight is calculated for each error value in the sequence. The dynamic weight value is determined by the absolute value of the error and a preset threshold bias parameter through an S-shaped function, such that the dynamic weight value monotonically increases from 0 to 1 as the absolute value of the historical trajectory prediction error increases. Then, a weighted average is calculated on the historical trajectory prediction error sequence based on the dynamic weight to obtain the prediction error value used to correct the typhoon forecast center.

[0013] Furthermore, the abstraction of the multi-agent collaborative observation task into a group game model specifically includes: Each agent is mapped to a player in a game, and a vector payoff function is defined for each agent to quantify the agent’s performance on at least two of the objectives of coverage effectiveness, operational security and energy efficiency. Establish a mapping relationship between the game situation and the state of the Markov decision process, and establish a correspondence between the policy and action space for each agent.

[0014] Furthermore, the establishment of a correspondence between policy and action space for each agent is achieved through an actor network in an actor-critic architecture, where the actor network is a parameterized policy function used to output actions or action probability distributions.

[0015] Furthermore, the state value of the Markov decision process is evaluated through a critic network in an actor-critic architecture, where the critic network is a parameterized value function.

[0016] Furthermore, the dual-loop equilibrium solution algorithm is specifically configured as a multi-layer semi-distributed training process, which is implemented by alternating the execution of the outer loop and the inner loop, wherein: The outer loop updates the group's game situation through iterative learning cycles, and determines whether the group strategy has reached global convergence by calculating the similarity of the group's game situation in consecutive learning cycles. The inner loop operates under a fixed game situation set by the outer loop. Each agent learns and updates its strategy based on the actor-commentator architecture, and determines whether the agent's strategy has reached local convergence by calculating the state similarity of the continuous decision-making stages. The outer and inner loops alternate until the group's joint strategy reaches global convergence, outputting a final strategy that approximates Pareto Nash equilibrium.

[0017] Furthermore, the update mechanism of the actor-critic architecture is as follows: in each decision step of the inner loop, the multi-objective payoffs output by the vector payoff function are weighted and fused according to a configurable target preference weight vector to generate an immediate reward, and the immediate reward is used as the training signal of the critic network to drive the agent's strategy update under the current game situation.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The typhoon tracking and correction scheme of this invention explores the distribution characteristics and correlations of typhoon external trajectory forecast data, and uses on-site observation data to predict typhoon trajectories based on extremely limited data. This lays the foundation for multi-agent continuous typhoon tracking under realistic conditions, and also makes efficient data collection possible in field operations.

[0019] 2. The autonomous planning scheme for typhoon observation tasks in this invention combines the advantages of game theory and reinforcement learning, giving the proposed algorithm powerful solution capabilities. It can handle continuous actions and state spaces, while also possessing strong transferability; that is, a game model can be established based on actual operational requirements, and the solution can be obtained based on multi-layer distributed reinforcement learning. This invention not only overcomes the environmental instability of fully distributed reinforcement learning algorithms but also overcomes the curse of dimensionality in centralized learning.

[0020] 3. The solution provided by this invention is characterized by formalization, modularity, and high scalability. Any well-performing specific algorithm can be incorporated into a specific module of this invention, enabling the overall solution to achieve better results. Furthermore, this invention can also be applied to any research and development involving the intelligence of group systems, such as autonomous driving, the Internet of Things, and financial transactions. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall scheme steps for intelligent agents to perform autonomous tracking and observation in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the composition of the typhoon tracking and trajectory prediction module in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram illustrating the use of situation iteration to derive the current equilibrium strategy when multiple agents perform task planning in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram illustrating the simulation results of tracking and observation conducted in MATLAB, using Typhoon Hyunlanno (No. 11 of 2022) as an example. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] A typhoon autonomous tracking and observation system based on intelligent swarms consists of three core components: a multi-agent cluster, a trajectory tracking and prediction subsystem, and a task autonomous planning subsystem. These three components are interconnected to form a complete closed-loop system from environmental perception and intelligent decision-making to task execution.

[0027] The multi-agent cluster consists of multiple intelligent devices with varying performance characteristics, including but not limited to underwater gliders, remotely operated underwater vehicles (ROVs), autonomous underwater vehicles (AUVs), or submersible unmanned surface vessels. In practice, these devices are randomly deployed by the mother ship to areas where future typhoons are likely to occur before the typhoon arrives, in order to acquire measured data on the ocean current field beneath the desired typhoon, including data on temperature, salinity, observation depth, and ocean current velocity.

[0028] The trajectory tracking and prediction subsystem is configured to communicate with the multi-agent cluster to receive external typhoon forecast data and, based on the real-time field observation data transmitted back by the multi-agent cluster, invert and correct the typhoon center position to generate the tracking target point for the next moment.

[0029] The internal logic and data flow of the trajectory tracking and prediction subsystem are shown in the attached figure. Figure 2 As shown. In its implementation, it is achieved through a Forecast Path Correction Model (FPCM). This model defines and calculates key points (center points) of tropical cyclone paths and uses these key points to correct the path in real time. Its core function is to improve the autonomous navigation capabilities of agents, ensuring they can accurately and quickly track tropical cyclones. The model defines a total of four types of key points: Forecast Center: The location of the typhoon center predicted from external meteorological forecast data has a certain degree of error; Computation Center: The center at the air-sea interface of the typhoon, calculated from ocean and meteorological data collected by the agent in the field, and used to correct the error of the prediction center; Correction Center: The expected position of the typhoon center at the next moment, obtained by correcting the prediction center using the computation center, is the target point that the multi-agent cluster actually needs to reach at the next moment. Actual center: The true center location of a typhoon, which cannot be accurately measured in real time, and is usually the final benchmark for target tracking.

[0030] The specific functions of the trajectory tracking and prediction subsystem are accomplished collaboratively by the following three modules: Typhoon External Forecast Module: Used to acquire typhoon trajectory forecast data from external meteorological stations during communication, and to perform data preprocessing for subsequent calculations.

[0031] Typhoon Center Calculation Module: Based on real-time ocean flow field data observed by a multi-agent cluster, the module inverts the coordinates of the typhoon's center position at the air-sea interface.

[0032] Trajectory correction module: Using the center position coordinates, the predicted center position coordinates in the forecast data are corrected through a dynamic weighted correction algorithm, and the optimal path of the typhoon center under the current conditions is output. At the same time, the correction center can be determined as the target of the multi-agent cluster's action at the next moment.

[0033] The typhoon center calculation module involves calculating the center of the typhoon at the air-sea interface. In this embodiment, the invention assumes multi-agent observations at a depth of several meters in the upper-level flow field generated by the typhoon. A three-dimensional model of the typhoon-generated flow field under static and stable conditions is then used; this model can be replaced by a more realistic theoretical model. A multi-point positioning method can be used to obtain the typhoon center at the air-sea interface. This method can also be replaced by a more realistic theoretical model.

[0034] The dynamic weighted correction algorithm in the trajectory correction module is as follows: Based on the historical trajectory prediction error sequence, a dynamic weight is calculated for each error value in the sequence. The dynamic weight value is determined by the absolute value of the error and a preset threshold bias parameter using a sigmoid function, such that the dynamic weight value monotonically increases from 0 to 1 as the absolute value of the historical trajectory prediction error increases. Then, a weighted average is calculated based on the dynamic weights to obtain the prediction error value used to correct the typhoon forecast center. The pseudocode for the dynamic weighted correction algorithm is summarized below:

[0035] The task autonomous planning subsystem is configured to: abstract the multi-agent collaborative observation task into a group game model, solve the model using a double cyclic equilibrium solution algorithm, output a set of agent policies, the game situation of the agent policies approximates Pareto Nash equilibrium, and generate task instructions for the multi-agent cluster according to the policies.

[0036] Specifically, abstracting the multi-agent collaborative observation task into a group game model includes: Each agent is mapped to a game player, and a reward function is defined for each individual agent and the collective agent to quantify the agent’s performance on at least two of the objectives of coverage effectiveness, operational security and energy efficiency. A mapping relationship between the game situation and the state of the Markov decision process is established based on an actor-commentator architecture, and a correspondence between the policy and action space is established for each agent. The correspondence between the agent's policy and action space is implemented through an actor network; the state value of the Markov decision process is evaluated through a commentator network. The actor and commentator networks adopt a multilayer perceptron structure. The actor network takes the agent's local observed state as input and outputs actions or action probability distributions; the commentator network takes the system's global state and joint actions as input and outputs state value evaluations. Network parameters are updated using gradient descent, and the learning rate is adjustable within the range of 0.001-0.0001.

[0037] The dual-loop equilibrium solution algorithm is specifically configured as a multi-layer semi-distributed training process, which is implemented by alternating the execution of the outer loop and the inner loop, wherein: The outer loop updates the group's game situation through iterative learning cycles, and determines whether the group strategy has reached global convergence by calculating the similarity of the group's game situation in consecutive learning cycles. The inner loop operates under a fixed game situation set by the outer loop. Each agent learns and updates its strategy based on the actor-commentator architecture, and determines whether the agent's strategy has reached local convergence by calculating the state similarity of the continuous decision-making stages. The outer and inner loops alternate until the group's joint strategy reaches global convergence, outputting a final strategy that approximates Pareto Nash equilibrium.

[0038] The update mechanism of the actor-critic architecture is as follows: In each decision step of the inner loop, the multi-objective payoffs output by the vector payoff function are weighted and fused according to a configurable target preference weight vector to generate an immediate reward. The immediate reward is then used as the training signal for the critic network to drive the agent's strategy update under the current game situation.

[0039] The pseudocode example of the dual-loop equilibrium solution algorithm is as follows:

[0040] In practice, the observation tasks involved in the autonomous task planning subsystem are not fixed. The specific observation content is given by the observation team before the multi-agents execute the task. This invention only abstracts the observation task into a task planning problem.

[0041] The trajectory tracking and prediction subsystem and the autonomous task planning subsystem of this invention can be housed within the intelligent agent to perform autonomous tracking and observation (e.g. Figure 1 (As shown). Before the typhoon arrives, when the multi-agent cluster has the ability to communicate with the outside world, it acquires as many real-time and forecast typhoon tracks as possible from surrounding weather stations. When the conditions for starting the task are met, the multi-agent cluster begins typhoon tracking and generates the target location for the next moment based on the typhoon tracking and trajectory prediction module. Furthermore, based on the relative distance between the multi-agent cluster and the typhoon, it can autonomously determine whether to start the predetermined typhoon observation task and autonomously plan the observation task according to the status of each agent, seeking a balanced strategy that saves public resources.

[0042] In one specific embodiment of the system of this invention, four autonomous underwater vehicles (AUVs) in a circular formation conducted a simulation experiment to track and observe Typhoon Hyunlanno, the 11th typhoon of 2022. This embodiment uses the actual typhoon path as the actual center and randomly generates positional errors based on this actual center. This method, rather than using actual forecast data from a specific institution at a specific lead time, promotes the versatility of the approach. The open-source typhoon dataset used in this embodiment was provided by the Kitamoto Laboratory (http: / / agora.ex.nii.ac.jp / digital-TC / summary / wnp / k / 202211.html.en). This comprehensive dataset contains various aspects of the typhoon, including its geographic coordinates, time information, and radius data. This data is directly used as the actual center of the typhoon, and cubic spline interpolation technology is used to observe the positional information of the AUVs in real time.

[0043] When conditions permit, the multi-agent cluster should begin typhoon tracking tasks. For example... Figure 4 As shown, this example selects the segment where the typhoon's movement speed is relatively slow but its direction changes significantly. In this task segment, four AUVs start from a predefined point and move towards the vicinity of the initial target point (𝐶0). The pink-filled circles represent AUVs deployed around the target point with a fixed radius, the size of which is controlled by the parameter FN. The AUVs will continuously track the typhoon until a specific time step is reached or the distance to the target point exceeds a certain threshold.

[0044] Assume the angular velocity of the typhoon wind field satisfies:

[0045] In the formula, r and r0 are the distance from the center of the typhoon and the radius of the maximum wind speed, respectively, and a and b are constants.

[0046] The expression for the typhoon calculation center obtained by multiple AUVs using the multi-point positioning method during the tracking process is as follows:

[0047] In the formula, , , , , where A is an (n-1)×2 matrix, B is an (n-1)-dimensional column vector, n represents the total number of AUVs with observed ocean current data, n> 3.

[0048] Based on this, the trajectory tracking and prediction subsystem can track and predict and correct the trajectory of typhoons.

[0049] When the distance between the multiple AUVs and the typhoon is suitable, the planning of the planned typhoon observation mission can begin. In this example, mission planning focuses on the selection of target points for multi-agent observation. The specific observation content can be chosen by the implementers, assuming at least ocean current velocity can be observed. The mission includes the following specific objectives: minimizing mission execution time by calculating the arrival time of each agent at the target point and determining the shortest mission completion time; minimizing energy consumption by calculating the energy consumption to reach the target point based on the average power of the agents and minimizing overall energy consumption; and minimizing mission redundancy by avoiding multiple agents selecting the same target point as much as possible, thereby maximizing information collection efficiency.

[0050] Based on the above objectives, the task autonomous planning subsystem abstracts the multi-agent collaborative observation task into a group game model: Map each AUV to a game player. Randomly initialize the finite pure strategy space for each player. Assume that players choose strategies with a certain probability. The pure strategy, the player's mixed strategy space is represented as Each player selects a hybrid strategy from their hybrid strategy space. Forming a game situation .

[0051] Construct a mapping between the game model and the transitional form of the Markov decision process, and establish the game situation. With state There is a one-to-one correspondence between strategies and actions, and the situation-state mapping relationship is set as follows: ; make ,in, , , Action-state value function: , making Design value response mapping: , making ,in Representing players Benefit function and equivalence. and These correspond to the policy network and value evaluation network of the Actor-Critic learning structure, respectively.

[0052] After establishing the above mapping relationship, a double-loop equilibrium solution algorithm is used, such as... Figure 3 As shown, the algorithm solves the decision-making problem by progressively approaching the equilibrium state of the group game. Guided by the Double Delay Deep Deterministic Policy Gradient (TD3) algorithm, the algorithm iteratively adjusts the agent's policy. In simulations, the iterative situation algorithm demonstrates its ability to gradually approach the equilibrium state, validating its effectiveness in solving group game problems.

[0053] like Figure 4 As shown, the typhoon autonomous tracking and observation system based on intelligent swarms of this invention significantly improves the accuracy of typhoon path prediction. The corrected path effectively reduces errors in traditional prediction methods, enabling AUVs to track typhoons more quickly and accurately. When approaching a typhoon, the AUV can effectively select target points through autonomous decision-making and achieve globally balanced task execution through interaction.

[0054] The above description is merely one specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A typhoon autonomous tracking and observation system based on intelligent swarms, characterized in that, It includes a multi-agent cluster, a trajectory tracking and prediction subsystem, and a task autonomous planning subsystem; The multi-agent cluster consists of multiple heterogeneous or homogeneous intelligent devices with different performance characteristics, used to perform navigation, positioning and data acquisition tasks; The trajectory tracking and prediction subsystem is configured to communicate with the multi-agent cluster to receive external typhoon forecast data and, based on the real-time field observation data transmitted back by the multi-agent cluster, invert and correct the typhoon center position to generate the tracking target point for the next moment. The task autonomous planning subsystem is configured to: abstract the collaborative observation task of the multi-agent cluster into a group game model, use a double-loop equilibrium solution algorithm to solve the group game model, output a set of equilibrium solutions of agent strategies, and generate task instructions for the multi-agent cluster according to the agent strategies. The trajectory tracking and prediction subsystem, the task autonomous planning subsystem, and the multi-agent cluster together constitute a closed-loop autonomous observation system. The trajectory tracking and prediction subsystem includes a typhoon external forecast module, a typhoon center calculation module, and a trajectory correction module. The typhoon external forecast module is configured to acquire and preprocess typhoon trajectory forecast data from external meteorological stations during communication. The typhoon center calculation module is configured to invert the center position coordinates of the typhoon at the air-sea interface based on the ocean current field data observed in real time by the multi-agent cluster. The trajectory correction module is configured to use the center position coordinates to correct the typhoon trajectory forecast data through a dynamic weighted correction algorithm, and output the corrected center as the target of the multi-agent cluster's next action. The abstraction of the collaborative observation task of a multi-agent cluster into a group game model specifically includes: mapping each agent to a player, defining a reward function for each individual agent and the cluster to quantify the agent's performance on at least two objectives: coverage effectiveness, operational security, and energy efficiency; establishing a mapping relationship between the game situation and the state of the Markov decision process based on the actor-commentator architecture, and establishing a correspondence between the policy and action space for each agent.

2. The typhoon autonomous tracking and observation system based on intelligent swarms as described in claim 1, characterized in that, The intelligent devices are underwater gliders, remotely controlled underwater robots, autonomous underwater robots, or unmanned surface vessels capable of diving; before the arrival of a typhoon, the intelligent devices are randomly deployed by the mother ship to sea areas where the typhoon may occur.

3. The typhoon autonomous tracking and observation system based on intelligent swarms as described in claim 1, characterized in that, The method for inverting the center position coordinates at the air-sea interface specifically includes: based on ocean current field data observed at different locations in the upper flow field induced by a multi-agent cluster, combined with a three-dimensional model of the upper flow field induced by the typhoon, the center position coordinates of the typhoon at the air-sea interface are calculated using a multi-point positioning method.

4. The typhoon autonomous tracking and observation system based on intelligent swarms as described in claim 1, characterized in that, The dynamic weighted correction algorithm is configured as follows: Based on the historical trajectory prediction error sequence, a dynamic weight is calculated for each error value in the historical trajectory prediction error sequence. The dynamic weight is determined by the absolute value of the error value and a preset threshold bias parameter through an S-shaped function, such that the dynamic weight increases monotonically from 0 to 1 as the absolute value of the historical trajectory prediction error value increases. Then, a weighted average is calculated on the historical trajectory prediction error sequence based on the dynamic weight to obtain the prediction error value used to correct the typhoon trajectory forecast data.

5. The typhoon autonomous tracking and observation system based on intelligent swarms according to claim 1, characterized in that, The process of establishing a correspondence between policy and action space for each agent is achieved through an actor network in an actor-critic architecture. The actor network is a parameterized policy function used to output actions or action probability distributions.

6. The typhoon autonomous tracking and observation system based on intelligent swarms as described in claim 1 or 5, characterized in that, The state value of the Markov decision process is evaluated through a critic network in an actor-critic architecture, where the critic network is a parameterized value function.

7. The typhoon autonomous tracking and observation system based on intelligent swarms according to claim 1, characterized in that, The dual-loop equilibrium solution algorithm is specifically configured as a multi-layer semi-distributed training process, which is implemented by alternating the execution of the outer loop and the inner loop, wherein: The outer loop updates the group game model through iterative learning cycles and determines whether the agent's strategy has reached global convergence. The inner loop operates under a group game model set by the outer loop. Each agent learns and updates its agent strategy based on the actor-critic architecture and determines whether the agent strategy has reached local convergence. The outer and inner loops alternate until global convergence is achieved, outputting a final agent policy that approximates Pareto Nash equilibrium.

8. The typhoon autonomous tracking and observation system based on intelligent swarms according to claim 7, characterized in that, The update mechanism of the actor-critic architecture is as follows: In each decision step of the inner loop, the multi-objective payoffs output by the payoff function of the target preference weight vector are weighted and fused according to a configurable target preference weight vector to generate an immediate reward. The immediate reward is then used as the training signal for the critic network to drive the agent's policy update under the current game situation.

Images