Inland ship route coordinate prediction and optimization system based on AI algorithm

Abstract

The application discloses an inland ship route coordinate prediction and optimization system based on an AI algorithm, and relates to the technical field of intelligent transportation. The system comprises a dynamic passable grid construction unit which accesses and fuses static base data, quasi-real-time dynamic data and other multi-source heterogeneous data, generates and continuously updates a four-dimensional dynamic passable grid in space-time, quantifies the comprehensive passable state of each space-time grid, and executes integrated decision-making based on the dynamic passable grid through the closed-loop interaction of a high-level sequence optimization module and a bottom path search module, and outputs a globally optimal destination access sequence and a route coordinate sequence, while realizing the coordinated reservation of key resources such as routes and anchorage locks. The application can quickly respond to changes in waterways and deviations in navigation, improve the safety, executability and economy of routes, and effectively optimize the efficiency of multi-destination logistics transportation of inland ships.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to an inland waterway vessel route coordinate prediction and optimization system based on AI algorithms. Background Technology

[0002] As a crucial component of the integrated transportation system, the level of intelligence in inland waterway transportation is critical for improving transportation efficiency and ensuring navigation safety. With the development of the Internet of Things, big data, and artificial intelligence technologies, analyzing and supporting decision-making based on data from systems such as Automatic Identification Systems (AIS) has become a research hotspot in the field of intelligent waterway transportation. However, the inland waterway environment is more complex than that of the ocean, subject to multiple dynamic constraints such as bridges, dams, shoals, water level changes, and dense traffic flows. This makes simple trajectory prediction or static path planning insufficient to meet the comprehensive requirements of safety, economy, and timeliness in actual navigation.

[0003] In existing technologies, related research mostly focuses on optimizing single aspects. For example, invention patent CN117688366A provides a method and system for predicting ship trajectories. This scheme constructs a prediction model by integrating a multi-head spatiotemporal self-attention mechanism with a bidirectional gated recurrent neural network, and uses an artificial fish swarm algorithm to optimize hyperparameters, aiming to improve the prediction accuracy of future trajectory point time series. This technology focuses on learning and extrapolating ship behavior patterns from historical data, belonging to a descriptive model. However, it does not consider the fundamental impact of channel dynamic physical constraints on the future route of ships, and therefore is difficult to use for proactive and executable navigation planning.

[0004] Another invention patent, CN119624031B, relates to an intelligent berthing method for inland waterway anchorages. This method optimizes resource utilization at the critical anchorage by predicting vessel arrival times and intelligently allocating berths. However, this solution primarily addresses berth scheduling after vessels arrive at the anchorage; its operation presupposes a predetermined vessel arrival time. It does not address how vessels navigate the waterway to achieve that predicted time, nor does it incorporate the dynamic difficulties of waterway passage into the overall resource allocation considerations. Essentially, it is an independent scheduling model.

[0005] In summary, a current problem facing the intelligent application of inland waterway vessels is the disconnect between cognition and decision-making. Specifically, trajectory prediction technology, represented by CN117688366A, can learn how a vessel might travel, but lacks a deep understanding of the environmental constraints that dictate its course. Resource scheduling technology, represented by CN119624031B, can optimize what to do after a vessel arrives, but its decision-making premises are disconnected from the actual navigation process within dynamically constrained waterways. These two technological paradigms have developed independently without synergy. This results in the inability, in practical applications, to generate an optimal route that comprehensively considers multi-destination mission sequences and dynamic waterway constraints, while also integrating the navigation path with spatiotemporal resources such as anchorages and locks for coordinated planning.

[0006] Therefore, there is an urgent need for a new technology that can integrate dynamic cognition of the waterway environment with multi-task navigation decision-making to overcome the limitations of the existing technologies. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-based inland waterway vessel route coordinate prediction and optimization system. It generates a spatiotemporal four-dimensional grid by integrating multi-source data through dynamic navigability grid construction units, and then optimizes the destination sequence and specific route through a multi-task integrated planner, thereby achieving a deep integration of dynamic channel constraint cognition and multi-task navigation decision-making.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: an inland waterway vessel route coordinate prediction and optimization system based on AI algorithms, comprising dynamic navigability grid building units and a multi-task integrated planner. The dynamic accessibility grid construction unit is used to access and fuse multi-source heterogeneous data to generate and continuously update a spatiotemporal four-dimensional dynamic accessibility grid covering the entire target waterway. Each spatiotemporal grid of the dynamic accessibility grid stores a quantized value that represents the comprehensive accessibility status of the location within a specific time period. The multi-task integrated planner is connected to the dynamic accessibility grid construction unit and is used to receive logistics task orders containing multiple destination information. Based on the latest dynamic accessibility grid, the multi-task integrated planner performs an integrated decision-making process and outputs the globally optimal destination access sequence and the specific route coordinate sequence connecting each pair of consecutive destinations in the sequence. The integrated decision-making process is used to perform joint optimization calculations on destination access order decision and continuous spatial path search at the same time.

[0009] Furthermore, the multi-source heterogeneous data includes static base data, near real-time dynamic data, planned event data, real-time traffic flow data, and ship-specific data; The static base data includes at least the fixed channel boundaries, fixed obstacle locations, static bridge clearance, and wharf locations recorded on the electronic navigation map. The near real-time dynamic data includes at least real-time water level data, wind direction and speed data, and temporary waterway control notices. The planned event data includes at least a lock opening and closing schedule and a timeline for the reservation and occupancy of anchorage berths. The real-time traffic flow data includes at least the real-time location, speed, and heading information of ships within the area obtained through the Automatic Identification System (AIS). The vessel's individual data includes at least the vessel's draft, height above the waterline, hull width, and current load status.

[0010] Furthermore, the dynamic mobility mesh building unit includes multiple parallel constraint quantizers and a weighted fusion unit; The constraint quantizer is used to convert the accessed multi-source heterogeneous data into corresponding spatiotemporal grid passage cost or risk probability values. The constraint quantizer includes at least a water depth constraint quantizer, a net height constraint quantizer, a time window constraint quantizer, a traffic conflict quantizer, and a resource status quantizer. The water depth constraint quantizer is used to calculate the effective water depth based on the static water depth and the real-time water level, and to generate a grounding risk probability value in combination with the ship's draft. The clearance constraint quantizer is used to calculate the actual clearance height under the bridge based on the bridge's static clearance height, real-time water level, and weather forecast data, and to generate a collision risk probability value by combining the ship's height above the waterline. The time window constraint quantizer is used to convert the lock opening and closing plan into a passable state Boolean value for the corresponding spatiotemporal grid. The traffic conflict quantizer is used to calculate the traffic density of each grid based on real-time traffic flow data using a kernel density estimation algorithm, and to predict potential conflict points to generate congestion delay cost values. The resource status quantizer is used to map the reservation occupancy timeline of anchorage berths to the resource occupancy cost value of the corresponding spatiotemporal grid. The weighted fusion unit receives the outputs of all constrained quantizers and calculates the comprehensive passage cost value of each spatiotemporal grid using a weighted summation model. The weighted summation model is expressed as:

[0011] in, Indicates spatial raster index, Indicates the time slice index. To constrain the total number of quantizers, For the first The weight coefficients corresponding to each constrained quantizer, and satisfying , Indicates the first The multi-source heterogeneous data processed by the constraint quantizer in the raster and time slices Data sets on Indicates the first A constraint quantizer maps input data to a function of cost or risk values.

[0012] Furthermore, the multi-task integrated planner includes a high-level sequence optimization module and a low-level path search module, which together form a closed-loop interaction structure. The high-level sequence optimization module is implemented using a deep reinforcement learning framework based on an attention mechanism. Its state space is defined as the set of currently visited destinations, the ship's current position and time, and the attribute set of all unvisited destinations. Its action space is defined as selecting the next destination to be visited. Its reward value is calculated from the actual sailing cost and time cost fed back by the low-level path search module. The underlying path search module is implemented using a time-varying A* search algorithm. When the higher-level sequence optimization module outputs an action, i.e., selects the next destination, this module is activated. The underlying path search module uses the ship's current position as the search starting point and the selected destination as the search ending point to search on the dynamic navigability grid, looking for the path with the minimum sum of comprehensive travel costs, and outputs the precise coordinate sequence of the path, the estimated travel time, and the total cost. The total cost is fed back to the higher-level sequence optimization module in real time for reward calculation.

[0013] Furthermore, the evaluation function of the time-varying A* search algorithm used in the underlying path search module... for:

[0014] in, This represents the current node in the search process, corresponding to a spatiotemporal grid in the dynamic décor grid. This indicates the distance from the starting point of the search to the current node. The actual cost already incurred, its value being from the starting point to... The total passage cost of all spatiotemporal grids traversed The sum of the sums, Indicates starting from the current node The estimated remaining cost to the search endpoint is calculated based on the distance-first principle of the basic navigation algorithm, using Euclidean geometric distance as a heuristic estimate to accelerate search convergence.

[0015] Furthermore, the high-level sequence optimization module incorporates a deep reinforcement learning agent, and the reward function of the deep reinforcement learning agent... Designed as follows:

[0016] in, This represents the estimated total travel time to complete the entire multi-destination mission, obtained by summing the estimated travel times for each segment of the route. This indicates the number of destinations in the task list that have hard time window constraints. This indicates that according to the current plan, the destination will be reached at the designated location. Estimated travel time for a destination with time window requirements. Indicates the first The latest arrival time required for each destination This represents the total travel cost of the segment corresponding to this action, as reported by the underlying path search module. , , These are the total time penalty coefficient, the time window violation penalty coefficient, and the navigation cost penalty coefficient, all of which are adjustable hyperparameters with a value greater than zero.

[0017] Furthermore, the system also includes a pre-training module for a large model of flight route data; The route data big model pre-training module is used to pre-train the high-level sequence optimization module in the multi-task integrated planner in an offline state using historical ship route big data, so as to learn empirical destination access order patterns and channel selection preferences, and provide initial decision strategies for the high-level sequence optimization module.

[0018] Furthermore, the system also includes a rolling replanning trigger module; The rolling replanning trigger module is connected to the dynamic navigability grid construction unit and the multi-task integrated planner, and is used to monitor the actual navigation status of the ship and the changes in key areas of the dynamic navigability grid; When a ship deviates from its original planned route by more than a predetermined threshold, or when the comprehensive passage cost of the grid within a predetermined range ahead of the ship in the dynamic passability grid changes by more than a preset threshold, the rolling replanning trigger module automatically starts the multi-task integrated planner. Starting from the ship's latest position and with the remaining unfinished tasks as the objective, it executes a new round of integrated decision-making process to generate an updated route coordinate sequence.

[0019] Furthermore, the dynamic accessibility grid building unit is also used to encode and integrate the spatiotemporal occupancy status of anchorage berths and locks as part of the planned event data into the dynamic accessibility grid; When making integrated decisions, the multi-task integrated planner can generate route coordinate sequences based on the resource occupancy costs encoded in the grid, and simultaneously plan out unoccupied spatiotemporal resource slots for ships at corresponding anchorages or locks that meet the expected arrival time, thereby achieving the collaborative generation and reservation of routes and key resources.

[0020] Furthermore, the closed-loop interaction process between the high-level sequence optimization module and the low-level path search module is specifically as follows: The high-level sequence optimization module outputs a probability distribution based on the current policy network and selects the next destination action accordingly. The underlying path search module is then invoked to calculate the optimal path to the destination, along with its cost and time, based on the current dynamic drivability grid. The calculated cost and time are encapsulated into a reward signal and returned to the high-level sequence optimization module; The high-level sequence optimization module uses the reward signal to update its policy network parameters and enters the next state. This process is repeated until all destinations have been visited, and finally outputs the destination visit sequence and each fine route segment that maximizes the cumulative reward.

[0021] Compared with existing technologies, this AI-based inland waterway vessel route coordinate prediction and optimization system has the following advantages: I. This invention utilizes dynamic accessibility grid construction units to access and integrate multi-source heterogeneous data, generating a spatiotemporal four-dimensional dynamic accessibility grid covering the entire target waterway, accurately quantifying the dynamic constraints and resource occupancy status of the waterway. Furthermore, combined with the closed-loop interactive structure of a multi-task integrated planner, it jointly optimizes destination access order decisions, continuous spatial path search, and reservations for key resources such as anchorages and locks. This solves the problems of fragmented trajectory prediction and resource scheduling, and a lack of deep understanding of dynamic waterway constraints in existing technologies. It ensures that the generated route plan is adapted to both the individual characteristics of the vessel and the actual waterway environment, while also matching the needs for key resource usage, significantly improving the global optimality and practical feasibility of the route.

[0022] Second, this invention utilizes historical route big data to pre-train the high-level sequence optimization module through the route data big model pre-training module, providing empirical initial strategies for the decision-making process and shortening the convergence time of online decision-making; at the same time, it uses the rolling replanning triggering module to monitor the ship's navigation status and changes in key areas of the waterway in real time, and automatically starts replanning when navigation deviations or sudden environmental changes occur, so that the route can dynamically adapt to changes.

[0023] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0025] Figure 1 This is a schematic diagram of the overall system architecture and data flow of the present invention; Figure 2 This is a schematic diagram of the dynamic accessibility mesh construction process of the present invention; Figure 3 This is a schematic diagram of the closed-loop interaction of the multi-task integrated planner of the present invention. Detailed Implementation

[0026] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0027] Example 1 like Figure 1 and Figure 3 As shown in the figure, this embodiment discloses a specific implementation of an AI-based inland waterway vessel route coordinate prediction and optimization system. This embodiment is applicable to multi-destination logistics transportation scenarios in the middle and lower reaches of the Yangtze River. This scenario features a dense waterway network, numerous bridges and dams, significant water level fluctuations due to seasonality and rainfall, heavy traffic, and vessels needing to complete cargo loading and unloading tasks at multiple docks. Simultaneously, navigation safety, timeliness, and cost must be considered. The following provides a detailed description of the system's components and working process.

[0028] This system mainly includes a dynamic accessibility grid building unit, a multi-task integrated planner, a route data large-scale model pre-training module, and a rolling replanning triggering module. These modules interact efficiently through data interfaces. The dynamic accessibility grid building unit provides the multi-task integrated planner with real-time quantitative data on the waterway environment. The route data large-scale model pre-training module provides the multi-task integrated planner with initial decision-making strategies. The rolling replanning triggering module dynamically initiates the replanning process based on navigation status and environmental changes, ensuring the real-time optimality of the route.

[0029] In this embodiment, the core function of the dynamic accessibility grid building unit is to fuse multi-source heterogeneous data to generate and update a spatiotemporal four-dimensional dynamic accessibility grid. The specific implementation process is as follows.

[0030] Access and preprocessing of multi-source heterogeneous data: The access of multi-source heterogeneous data adopts a distributed data acquisition architecture, and realizes real-time acquisition and format unification of various types of data through standardized interfaces. The specific data types and acquisition methods are as follows: Static base data: obtained from the electronic nautical chart database provided by the maritime authorities, including the latitude and longitude coordinates of fixed channel boundaries, the location coordinates of fixed obstacles, the static clearance of bridges, and the specific location coordinates and number of berths of wharves. The data format is uniformly converted to vector data in the WGS84 coordinate system, and after deduplication and coordinate calibration, it is stored in the static database.

[0031] Near real-time dynamic data: Real-time water level data is acquired through IoT sensors at hydrological monitoring stations along the waterway, with a sampling frequency of once every 10 minutes; wind direction and speed data are acquired through the meteorological department's public API interface, with an update cycle of every 30 minutes; temporary waterway control notices are acquired simultaneously through the maritime safety administration's VHF communication channel and electronic bulletin system, and key information such as the control area, control period, and control reasons are extracted using natural language processing technology.

[0032] Planned event data: The lock opening and closing schedule is obtained through the lock management department's scheduling system interface, including information such as the daily opening and closing times and single-pass capacity of each lock; the anchorage berth reservation timeline is obtained through the anchorage management system, recording the reservation vessel information, expected start and end times of each berth.

[0033] Real-time traffic flow data: The AIS signals of all ships in the area are received by the Automatic Identification System (AIS) base stations deployed at key nodes of the waterway. The real-time position, speed, and heading of the ships are obtained by parsing the data. The data is updated every 30 seconds. After data cleaning to remove abnormal drift data, the data is stored in the real-time traffic flow database.

[0034] Ship-specific data: This data is proactively reported by the ship's onboard terminal and includes the ship's design draft, actual draft, height above the waterline, hull width, and deadweight tonnage corresponding to the current cargo load. The data is reported once per hour, and if the load status changes, it is reported in real time.

[0035] The specific implementation of the constrained quantizer: The constrained quantizer adopts a parallel computing architecture. Each quantizer independently processes the corresponding type of data and outputs a standardized cost or risk value. The specific implementation is as follows: Water depth constraint quantizer: First, the effective water depth is calculated based on the static water depth and real-time water level data from the static base data. The effective water depth equals the difference between the real-time water level and the static water depth. Then, combined with the actual draft from the vessel's individual data, a logistic regression model is used to construct a grounding risk probability calculation function. ,in For grid In time slice The dataset includes data on effective water depth and ship draft. When the effective water depth is 1.2 times or more than the ship's draft, the probability of grounding is 0; when the effective water depth is between 1.0 and 1.2 times the ship's draft, the probability of grounding increases linearly to 0.8 as the effective water depth decreases; when the effective water depth is less than 1.0 times the ship's draft, the probability of grounding is 1.0.

[0036] The clearance constraint quantifier calculates the actual clearance height under the bridge based on the bridge's static clearance height, real-time water level, and wind speed data from weather forecasts. The actual clearance height equals the bridge's static clearance height minus the real-time water level rise. A correction value for ship turbulence caused by wind speed is then considered, determined by the product coefficient of wind speed and the ship's height above the waterline. Combining this with the ship's height above the waterline, an exponential function is used to construct a collision risk probability calculation function. in For grid In time slice The data set includes the actual clearance height under the bridge and the height of the vessel above the waterline. When the actual clearance height under the bridge is 1.1 times or more than the height of the vessel above the waterline, the probability of collision is 0; when the actual clearance height under the bridge is less than the height of the vessel above the waterline, the probability of collision is 1.0; in other cases, the probability of collision increases exponentially as the actual clearance height under the bridge decreases.

[0037] Time window constraint quantizer: Converts the lock opening and closing time schedule into passable Boolean values ​​of a spatiotemporal raster, and constructs a function. in For grid In time slice The corresponding lock opening and closing schedule data. If time slice t falls within the lock's open period, the passability Boolean value is 1, indicating that the grid is passable during this period; if it falls within the lock's closed period, the Boolean value is 0, indicating that it is not passable.

[0038] Traffic Conflict Quantifier: Based on real-time traffic flow data, a kernel density estimation algorithm is used to calculate the traffic density of each grid cell. A Gaussian kernel function is selected, and the bandwidth is determined to be 1.0 using cross-validation. Potential conflict points are then predicted based on traffic density. When the traffic density of a grid cell exceeds a preset threshold, it is identified as a potential conflict area. A congestion delay cost calculation function is then constructed based on the statistical relationship between traffic density and congestion delay time. in For grid In time slice This is a collection of real-time traffic flow data, including ship positions, speeds, and headings. Congestion delay costs are linearly positively correlated with traffic density; when traffic density is 0, the delay cost is 0; when traffic density reaches a threshold, the delay cost reaches its maximum value.

[0039] Resource Status Quantizer: Maps the timeline of anchorage berth reservations to resource occupancy costs, and constructs functions. in For grid In time slice The corresponding anchorage berth reservation occupancy data. (If time slice) If the berth corresponding to the grid has been reserved, the resource occupancy cost is the preset maximum value; if it has not been reserved, the cost value is calculated based on the remaining capacity of the berth. The more remaining capacity, the lower the cost value, showing a linear negative correlation.

[0040] Specific implementation of the weighted fusion unit: The weighted fusion unit receives the outputs of the five constrained quantizers mentioned above and calculates the comprehensive passage cost value of each spatiotemporal grid through a weighted summation model. The model formula is as follows:

[0041] The parameters are explained below: Spatial grid index uses a two-dimensional grid encoding method, which divides the entire target channel area into several spatial grids according to a preset grid size. Each grid is assigned a unique index value to locate the specific spatial position of the grid.

[0042] Time slice index: Divide a 24-hour day into several time slices according to a preset time interval. Each time slice is assigned a unique index value to identify the specific time period.

[0043] The total number of constraint quantizers, in this embodiment This corresponds to five constraint quantizers.

[0044] : No. The weight coefficients corresponding to each constrained quantizer satisfy the following conditions: The weighting coefficients are determined using the analytic hierarchy process (AHP), taking into account the safety priority principle for inland waterway navigation, and the weights of the water depth constraint quantizer. Net height constraint quantizer weights Traffic Conflict Quantizer Weights Time window constrained quantizer weights Resource status quantizer weights .

[0045] : No. The multi-source heterogeneous data processed by the constraint quantizer in the raster and time slices The dataset consists of preprocessed, standardized data.

[0046] : No. Each constraint quantizer maps input data to a function of cost or risk value; that is, the output value range of the specific functions defined in each constraint quantizer is normalized to the range [0,1].

[0047] The comprehensive passage cost value of a specific spatiotemporal grid. Identify cost attributes, It limits the time and space range, accurately reflecting the difficulty and risk of passage through the corresponding grid in the corresponding time period.

[0048] : It is an abbreviation used when the context clearly defines the spatiotemporal scope, and the physical meaning remains consistent, simplifying the expression and avoiding redundancy.

[0049] The weighted fusion unit performs the above formula calculation on each spatiotemporal grid through parallel computing to generate a spatiotemporal four-dimensional dynamic navigability grid covering the entire target waterway, and updates it in real time according to the time slice index to ensure that the grid data is consistent with the actual environmental state of the waterway.

[0050] In this embodiment, the multi-task integrated planner includes a high-level sequence optimization module and a low-level path search module. The two form a closed-loop interactive structure to achieve joint optimization of destination access sequence and specific route coordinates. The specific implementation is as follows.

[0051] Implementation details of the high-level sequence optimization module: The high-level sequence optimization module is implemented using a deep reinforcement learning framework based on an attention mechanism. The core is to train a deep reinforcement learning agent capable of outputting the optimal destination access sequence. The specific implementation process is as follows: Network Structure Design: The network structure of the deep reinforcement learning agent includes an input layer, an attention layer, hidden layers, and an output layer. The input layer receives state space data, and its dimension is determined according to the number of features in the state space. The attention layer uses a multi-head attention mechanism with 8 heads to capture the correlation between different destinations and the correlation between the destination and the current ship state. The hidden layer uses a two-layer fully connected network with 256 neurons in each layer, and the ReLU activation function is used. The output layer uses the Softmax activation function to output the probability distribution of the next destination.

[0052] State space definition: The state space consists of three parts: First, the set of currently visited destinations, represented by one-hot encoding, where each destination corresponds to a bit; if a destination has been visited, the bit is 1, and if it has not been visited, it is 0. Second, the ship's current position and time, where the ship's current position is its latitude and longitude coordinates, and the time is its current time slice index. Third, the set of attributes for all unvisited destinations, including the destination's latitude and longitude coordinates, the time required for cargo loading and unloading, the latest arrival time, and other attributes.

[0053] Action space definition: The action space is the selection of the next destination to be visited. Each action corresponds to an unvisited destination. When all destinations have been visited, the action space is empty.

[0054] Reward function calculation: Reward function The following formula is used for calculation:

[0055] The parameters are explained below: The estimated total travel time to complete the entire multi-destination mission is obtained by adding up the estimated travel times of each segment of the route, and is expressed in hours.

[0056] The number of destinations with hard time window constraints in the task order is directly specified by the logistics task order.

[0057] According to the current plan, we will reach the [number]th [location]. The estimated travel time for a destination with a time window requirement is calculated by the underlying path search module in hours, starting from 00:00 on the current day.

[0058] : No. The latest arrival time required for each destination is specified by the logistics task order, in hours, calculated from 00:00 on the same day.

[0059] The total travel cost of the segment corresponding to this action, fed back by the underlying path search module, is the comprehensive travel cost value of all spatiotemporal grids traversed by this segment. The sum of these values ​​has been normalized to the range [0,1].

[0060] Total time penalty coefficient, used to adjust the weight of total travel time in the reward function, with a value of 0.4, which can be adjusted according to actual transportation needs.

[0061] The time window penalty coefficient is used to penalize situations where the latest arrival time is exceeded. It has a value of 0.5 to ensure that ships strictly comply with the time window constraints.

[0062] The voyage cost penalty coefficient is used to adjust the weight of voyage costs in the reward function. It has a value of 0.1 and can be adjusted according to the sensitivity of ship operating costs.

[0063] To find the maximum value function, This represents the default threshold within the time window. When the estimated arrival time is later than the latest arrival time, the function outputs the default duration as a penalty; otherwise, it outputs... To avoid negative interference in non-default scenarios and ensure that the reward function guides the process in accordance with time constraints.

[0064] The reward function is calculated as follows: the longer the total travel time, the longer the time window default duration, and the higher the travel cost, the smaller the reward value, and vice versa. This guides the agent to learn the optimal destination access sequence.

[0065] Training process: In the offline training phase, the network parameters of the agent are initialized using the initial decision-making strategy provided by the large model pre-training module of flight route data. In the online training phase, the agent outputs the probability distribution of the next destination based on the current state, selects the destination with the highest probability as the action, and calculates the reward value after the underlying path search module provides feedback on the cost and time. The network parameters are updated using the stochastic gradient descent method, with the learning rate set to 0.001, the number of training iterations being 10,000, and the batch size for each iteration being 32.

[0066] The underlying path search module is implemented using a time-varying A* search algorithm. It is activated after the higher-level sequence optimization module selects the next destination, searching for the optimal path from the ship's current position to the selected destination. The specific implementation is as follows: Algorithm initialization: Set the search starting point to the spatiotemporal grid corresponding to the ship's current position and the search ending point to the spatiotemporal grid corresponding to the selected destination. Initialize the open list and the closed list. The open list is used to store the grids to be searched, and the closed list is used to store the grids that have been searched.

[0067] Evaluation function calculation: Evaluation function of the time-varying A* search algorithm The formula used to measure the estimated total cost from the current node to the destination is as follows:

[0068] The parameters are explained below: The current node in the search process corresponds to a spatiotemporal raster in the dynamic accessibility grid, and contains a spatial raster index. and time slice index .

[0069] From the starting point of the search to the current node The actual cost already incurred, its value being from the starting point to... The total passage cost of all spatiotemporal grids traversed The sum of these values ​​reflects the actual cost of the route already traveled.

[0070] From the current node The estimated remaining cost to the search endpoint is calculated using the Euclidean geometric distance formula, based on the distance-first principle.

[0071]

[0072] in , For the current node latitude and longitude coordinates , The latitude and longitude coordinates of the search endpoint. Distance-cost conversion factor, used to convert geographical distance to cost-effectiveness. The cost of the same order of magnitude is 0.001.

[0073] Evaluation function The smaller the value, the closer the path corresponding to the current node is to the optimal path, and the algorithm prioritizes searching for it. Nodes with smaller values ​​are selected to accelerate search convergence.

[0074] Path search process: Select from the open list The node with the smallest value is selected as the current node. It is then checked whether this node is the endpoint. If it is the endpoint, the search ends, and the optimal path is obtained by backtracking. If it is not the endpoint, the neighboring grids of the current node are expanded, and the value of each neighboring grid is calculated. , and If an adjacent raster value is not in the open or closed list, add it to the open list; if it is already in the open list, compare the most recently calculated value. Value and original If the new value is smaller, then update the node's... , Value and parent node; ignore the node if it is already in the closed list. Repeat the above process until the endpoint is found or the open list is empty.

[0075] Output: After the path search is completed, the precise coordinate sequence of the path is output, i.e., the latitude and longitude coordinates and corresponding time slices of each spatiotemporal grid passed by the path. Simultaneously, the estimated travel time and total cost are calculated. The estimated travel time is calculated based on the path length and the ship's average speed, and the total cost is the sum of the costs of all spatiotemporal grids passed by the path. The summation of the estimated travel time and total cost is fed back to the high-level sequence optimization module.

[0076] The specific implementation of the closed-loop interaction process: The closed-loop interaction process between the high-level sequence optimization module and the low-level path search module is as follows: The high-level sequence optimization module, based on the set of currently visited destinations, the ship's current position and time, and the set of attributes of unvisited destinations, outputs the probability distribution of the next destination through the policy network and selects the destination with the highest probability as the current action.

[0077] The underlying path search module is invoked, starting from the ship's current position and ending at the selected destination. It executes a time-varying A* search algorithm on the dynamic navigability grid to calculate the optimal path and outputs the precise coordinate sequence of the path, the estimated travel time, and the total cost.

[0078] The underlying path search module encapsulates the estimated travel time and total cost into a reward signal and feeds it back to the higher-level sequence optimization module.

[0079] The high-level sequence optimization module calculates the reward value based on the reward signal, updates the policy network parameters using the stochastic gradient descent method, and updates the state space, that is, marks the current destination as visited and updates the ship's current position and time.

[0080] Repeat the above steps until all destinations have been visited. The high-level sequence optimization module outputs the destination visit sequence that maximizes the cumulative reward, and the low-level path search module outputs the precise coordinate sequence of each path segment. These are combined to form the globally optimal route plan.

[0081] In this embodiment, the route data large model pre-training module is used to pre-train the high-level sequence optimization module in an offline state, and the specific implementation is as follows: Historical data collection and preprocessing: Collect historical vessel route data for the target waterway over the past three years, including vessel AIS trajectory data, logistics task order data, and waterway environmental data. Clean the historical data, remove outliers, fill in missing data, and standardize it according to the same format as real-time data to construct a historical dataset.

[0082] Pre-training process: The deep reinforcement learning agent for the high-level sequence optimization module is pre-trained using historical datasets. The goal of pre-training is to enable the agent to learn destination access order patterns and route selection preferences from historical data. The reward function in the pre-training stage is consistent with that in the online training stage, with 5000 training iterations, a learning rate of 0.002, and a batch size of 64.

[0083] Initial decision strategy output: After pre-training, the agent's network parameters are stored as the initial decision strategy in the multi-task integrated planner, providing an initial basis for online decision-making, shortening the convergence time of online training, and improving the rationality of the initial decision.

[0084] In this embodiment, the rolling replanning triggering module is used to monitor the changes in key areas of the ship's actual navigation status and the dynamic navigability grid, and trigger rolling replanning. The specific implementation is as follows: Monitoring content and methods: Real-time monitoring of the deviation between the actual navigation position of the vessel and the original planned route. The actual position of the vessel is obtained through the shipborne GPS equipment and compared with the coordinate sequence of the original planned route to calculate the deviation distance. At the same time, the changes in the comprehensive passage cost value of the grid within a predetermined range in front of the vessel in the dynamic passability grid are monitored, and the current grid data is compared with the grid data of the previous update cycle in real time.

[0085] Triggering conditions: When the deviation between the ship's actual navigation position and the original planned route exceeds a predetermined threshold, or when the change in the grid comprehensive passage cost value within a predetermined range ahead of the ship exceeds a preset threshold, it is determined that rolling replanning is required.

[0086] Rolling Replanning Execution: After rolling replanning is triggered, the rolling replanning triggering module automatically starts the multi-task integrated planner. Starting from the ship's current actual position and with the remaining unfinished destination tasks as the target, it performs an integrated decision-making process based on the latest dynamic navigability grid, generates updated destination access sequences and route coordinate sequences, and sends them to the shipboard navigation system.

[0087] In this embodiment, the dynamic accessibility grid construction unit encodes and integrates the spatiotemporal occupancy status of anchorage berths and locks into the dynamic accessibility grid. The multi-task integrated planner generates route coordinate sequences while simultaneously achieving coordinated reservation of routes and key resources. The specific implementation is as follows: Spatiotemporal resource coding: The timeline for the reservation and occupancy of anchorage berths and the lock opening and closing schedule are coded according to the index of the spatiotemporal grid. The spatiotemporal occupancy status of anchorage berths is coded as a resource occupancy cost value, and the spatiotemporal occupancy status of locks is coded as a passable Boolean value. These values ​​are then integrated with the outputs of other constraint quantizers to form a comprehensive passage cost value. The calculation.

[0088] Resource reservation planning: When making integrated decisions, the multi-task integrated planner, based on the resource occupancy cost encoded in the grid and combined with the estimated arrival time of each path segment, plans unoccupied anchorage berths or lock space-time resource slots that meet the estimated arrival time for the vessel. For example, when a vessel is expected to arrive at an anchorage in a certain time slot, the system automatically queries the resource occupancy status of that anchorage in that time slot, selects an unoccupied berth as the reservation target, and submits a reservation application through the anchorage management system. For locks, the system plans the optimal time window for the vessel to pass through the lock based on the estimated arrival time of the vessel and the lock opening and closing time schedule, ensuring that the vessel can pass through in a timely manner after arrival.

[0089] In this embodiment, the complete workflow of the system is as follows: Data access and grid construction: The dynamic accessibility grid construction unit accesses multi-source heterogeneous data, which is processed by the constraint quantizer and weighted fusion to generate and continuously update the spatiotemporal four-dimensional dynamic accessibility grid.

[0090] Pre-training initialization: The large-scale model pre-training module for flight route data uses historical data to pre-train the high-level sequence optimization module, providing an initial decision-making strategy.

[0091] Integrated decision-making: The multi-task integrated planner receives logistics task orders, and the high-level sequence optimization module and the low-level path search module interact in a closed loop to output the globally optimal destination access sequence and route coordinate sequence.

[0092] Navigation monitoring and replanning: The rolling replanning trigger module monitors the ship's navigation status and grid changes in real time. When the triggering conditions are met, rolling replanning is initiated to generate an updated route plan.

[0093] Resource Coordination and Reservation: During the route planning process, the time and space resources of anchorage berths and locks are reserved simultaneously to ensure the feasibility of the route.

[0094] This embodiment, through the above technical solution, realizes intelligent planning and optimization of multi-destination routes for inland waterway vessels, and has the following beneficial effects compared with the prior art: Taking into account both the dynamic constraints of the waterway environment and the individual characteristics of the vessels, the generated route plan is safer and effectively reduces the risks of grounding, collision, and congestion.

[0095] It achieves joint optimization of destination access sequence and specific route, while taking into account both flight timeliness and cost, thereby improving the efficiency of logistics transportation and reducing operating costs.

[0096] The rolling replanning mechanism enables rapid response to changes in the waterway environment and deviations in ship navigation, ensuring the real-time optimality and feasibility of the route.

[0097] It enables coordinated planning and reservation of key resources such as shipping routes, anchorages, and locks, improving the utilization rate of waterway resources and reducing vessel waiting time.

[0098] Example 2 like Figure 1 and Figure 3 As shown in Embodiment 2, this embodiment discloses a detailed implementation of the workflow of an inland waterway vessel route coordinate prediction and optimization system based on AI algorithms. This embodiment is applicable to multi-destination logistics transportation scenarios in the Jingjiang section of the middle reaches of the Yangtze River. This waterway section is characterized by its winding and narrow course, dense distribution of shoals, and numerous locks and bridges. The water level fluctuates significantly due to seasonal rainfall, and cargo ships frequently travel through the area. Ships need to sequentially complete cargo loading and unloading tasks at multiple ports, placing extremely high demands on balancing navigation safety, transportation timeliness, and operating costs. This embodiment further clarifies the collaborative operation mode of each component unit by fully presenting the entire workflow mechanism of the system from initialization to task completion.

[0099] The core components of this system include a dynamic navigability grid construction unit, a multi-task integrated planner, a route data large-scale model pre-training module, and a rolling replanning triggering module. Each unit achieves data exchange and command transmission through standardized data interfaces. The dynamic navigability grid construction unit provides the multi-task integrated planner with real-time updated quantitative data support for the waterway environment. The route data large-scale model pre-training module provides the initial optimization foundation for the multi-task integrated planner's decision-making logic. The rolling replanning triggering module dynamically adjusts the planning strategy based on changes in navigation status and environment. These four components work together to form a complete route planning and optimization system.

[0100] The complete workflow of this system is described in detail below: I. System Initialization and Pre-training Preparation: After system startup, the initialization phase begins, completing hardware debugging, software module self-testing, and data interface connectivity verification to ensure that all units can work together normally. During this phase, the route data large model pre-training module simultaneously starts the offline pre-training process, providing initial decision strategies for the multi-task integrated planner.

[0101] Specifically, the route data big data pre-training module collects historical vessel navigation data for the Jingjiang River section of the target waterway over the past three years, including vessel AIS trajectory data, logistics task record records, waterway environmental monitoring data, and anchorage lock usage data. The collected historical data is preprocessed to remove data deviating from the trajectory due to vessel malfunctions or equipment abnormalities, as well as invalid data lacking key information. Historical data in different formats are then uniformly converted into a standardized data format to construct a structured historical dataset.

[0102] Based on this historical dataset, the route data large model pre-training module pre-trains the high-level sequence optimization module in the multi-task integrated planner. During training, the module learns from historical data the destination access order patterns of ships in similar task scenarios, path selection preferences under different waterway environments, and the correlation characteristics between ship navigation costs and time, forming an initial decision strategy. After pre-training, this initial decision strategy is stored in the storage unit of the multi-task integrated planner, providing a basis for subsequent online decision-making, shortening the strategy convergence time in the online decision-making process, and improving the rationality of the initial planning scheme.

[0103] II. Multi-source heterogeneous data access and dynamic accessibility grid generation: After the system initialization is completed, the dynamic accessibility grid construction unit continuously accesses multi-source heterogeneous data and generates a spatiotemporal four-dimensional dynamic accessibility grid through data fusion processing, providing real-time environmental support for route planning.

[0104] Specifically, the dynamic accessibility grid construction unit adopts a distributed data acquisition architecture, which realizes real-time acquisition and format unification of various types of data through a pre-set standardized interface. Among them, the static base data is obtained from the electronic navigable map database provided by the maritime administration department, which includes fixed geographical information such as fixed channel boundaries, fixed obstacle locations, static bridge clearance, and wharf locations. After coordinate calibration and deduplication, the data is stored in the static database as the basic geographical framework for grid construction.

[0105] Near real-time dynamic data is acquired synchronously through multiple channels. Real-time water level data is collected by IoT sensors at hydrological monitoring stations deployed along the waterway. Wind direction and speed data are obtained through the meteorological department's public data interface. Temporary waterway control notices are received synchronously through the maritime safety administration's VHF communication channel and electronic bulletin system. Key information such as control areas and control periods are extracted using natural language processing technology to ensure the timeliness and accuracy of the data.

[0106] Planned event data is obtained directly from the relevant management system interface. The lock opening and closing time schedule comes from the lock management department's scheduling system, which includes information such as the opening time and passage capacity of each lock. The anchorage berth reservation and occupancy timeline comes from the anchorage management system, which records the reservation and occupancy time of each berth and the information of the occupant vessels.

[0107] Real-time traffic flow data is received by AIS base stations deployed at key nodes of the waterway. The data is parsed to obtain information such as the real-time position, speed, and heading of all ships in the area. After the data is cleaned and drifting abnormal data is removed, it is stored in the real-time traffic flow database in a high-frequency update manner.

[0108] The ship's individual data is actively reported by the ship's onboard terminal. It includes the ship's own attribute information such as draft, height above the waterline, hull width, and current load status. The data is reported in real time when the load status changes to ensure that the data is consistent with the actual situation of the ship.

[0109] After data access is completed, multiple constraint quantizers in the dynamic accessibility grid construction unit work in parallel to transform various multi-source heterogeneous data into corresponding spatiotemporal grid access costs or risk probability values. The water depth constraint quantizer calculates the effective water depth by combining static water depth and real-time water level, and then determines the grounding risk probability of each grid based on the ship's draft; the clearance constraint quantizer calculates the actual clearance height under the bridge based on the bridge's static clearance height, real-time water level, and meteorological data, and generates the collision risk probability by combining the ship's height above the waterline; the time window constraint quantizer transforms the lock opening and closing plan into grid accessibility status indicators; the traffic conflict quantizer calculates the traffic density of each grid based on real-time traffic flow data, predicts potential conflict points, and generates congestion delay costs; and the resource status quantizer maps the anchorage berth reservation occupancy timeline into resource occupancy costs.

[0110] The outputs of each constraint quantizer are synchronously transmitted to the weighted fusion unit. The weighted fusion unit uses a preset weighted summation logic to perform fusion calculations on various cost and risk values, obtaining the comprehensive passage cost value for each spatiotemporal grid. This generates a spatiotemporal four-dimensional dynamic passability grid covering the entire target waterway. This grid is continuously updated at preset time intervals to ensure that the grid data remains synchronized with the actual environmental conditions of the waterway in real time, providing accurate environmental quantification data for subsequent route planning.

[0111] III. Multi-task integrated planning and resource collaborative reservation: When the system receives a logistics task order containing information on multiple destinations, the multi-task integrated planner starts the integrated decision-making process. Combining dynamic accessibility grid data and initial decision-making strategies, it generates the globally optimal destination access sequence and route coordinate sequence, and at the same time completes the resource collaborative reservation of anchorage and lock.

[0112] Specifically, the high-level sequence optimization module and the low-level path search module in the multi-task integrated planner operate in a closed-loop interactive manner. The high-level sequence optimization module adopts a deep reinforcement learning framework based on an attention mechanism. Its state space includes the set of currently visited destinations, the ship's current position and time, and attribute information of all unvisited destinations. Based on the initial decision policy and the environmental state reflected by the current dynamic accessibility grid, the module outputs the selection probability of the next destination to be visited through the policy network, and determines the next target destination based on this probability.

[0113] After the high-level sequence optimization module outputs the destination selection results, the low-level path search module is activated. The low-level path search module employs a time-varying A* search algorithm, using the ship's current position as the search starting point and the selected destination as the search endpoint, to conduct a path search on a dynamic navigability grid. During the search, the module comprehensively considers the overall navigability cost of each spatiotemporal grid, seeking the path with the minimum sum of overall navigability costs. It then generates the precise coordinate sequence of this path and calculates the corresponding estimated travel time and total cost.

[0114] The bottom-level path search module feeds back the calculated estimated travel time and total cost to the top-level sequence optimization module. Based on this feedback, the top-level sequence optimization module calculates a reward value, taking into account factors such as total travel time, time window constraint satisfaction, and travel cost. The reward value is then used to adjust the strategy network parameters and optimize the destination selection logic. Subsequently, the top-level sequence optimization module enters the next state, selecting a new unvisited destination. The bottom-level path search module then performs path search and cost calculation again, repeating this process until all destinations are included in the visit sequence.

[0115] During the integrated decision-making process, the multi-task integrated planner simultaneously achieves coordinated reservation of routes and key resources. The spatiotemporal occupancy status of anchorage berths and locks is encoded in the dynamic accessibility grid. Based on this encoded information and the estimated arrival time of each route segment, the multi-task integrated planner plans passage periods for ships that meet time requirements and are not currently occupied at anchorage berths or locks. For anchorages, the system queries berth occupancy status within the estimated arrival time of the ship, selects an available berth, and submits a reservation application through the anchorage management system. For locks, the system combines the lock opening and closing schedule with the estimated arrival time of the ship to plan the optimal passage time window, ensuring timely passage after arrival and achieving coordinated matching between route planning and resource utilization.

[0116] Once all destinations have completed their route planning, the multi-task integrated planner outputs the globally optimal destination visit sequence, as well as the specific route coordinate sequence connecting each consecutive destination. It also outputs the reservation information for anchorage berths and locks, forming a complete navigation plan.

[0117] IV. Route Issuance and Navigation Execution Monitoring: The navigation plan generated by the multi-task integrated planner is issued to the ship's onboard navigation system via a wireless communication network. The onboard navigation system provides real-time navigation guidance to the crew based on the route coordinate sequence, and the ship begins navigation according to the planned destination visit sequence and route.

[0118] During navigation, the rolling replanning trigger module continuously monitors the ship's actual navigation status and changes in key areas of the dynamic navigability grid. The module acquires the ship's real-time position information via onboard GPS equipment, compares it in real-time with the coordinate sequence of the planned route, and calculates the deviation distance between the actual navigation position and the planned route. Simultaneously, the module monitors changes in the overall passage cost of the grid within a predetermined range ahead of the ship in the dynamic navigability grid, comparing the current grid data with the grid data from the previous update cycle.

[0119] V. Rolling Replanning Execution: When the rolling replanning trigger module detects that the deviation between the ship's actual navigation position and the planned route exceeds a preset threshold, or the change in the grid comprehensive passage cost within a predetermined range ahead of the ship exceeds a preset threshold, the rolling replanning process is automatically initiated.

[0120] The rolling replanning trigger module sends a replanning instruction to the multi-task integrated planner. Upon receiving the instruction, the multi-task integrated planner uses the ship's current actual position as the new planning starting point and the remaining unfinished destination tasks as the planning objectives. It then calls upon the latest dynamic maneuverability grid data and re-executes the integrated decision-making process. Through closed-loop interaction between the high-level sequence optimization module and the low-level path search module, updated destination access sequences and route coordinate sequences are generated. Simultaneously, anchorage lock resource reservation information is re-optimized to ensure that the new planning scheme can adapt to changes in navigation conditions and the environment.

[0121] The updated navigation plan is pushed to the crew via the ship's navigation system. Once the crew confirms it, they can continue sailing according to the new route, ensuring that the ship always maintains the optimal navigation status in a complex and dynamic waterway environment.

[0122] VI. Task Completion and Data Archiving: After a vessel completes cargo loading and unloading tasks at all destinations according to the planned scheme, the system automatically records complete data for this voyage, including the actual navigation trajectory, actual navigation time for each segment of the route, passage costs, anchorage and lock resource utilization, and key change data of the dynamic accessibility grid. This data is then archived into the system's historical database. On the one hand, this provides new training data for the route data large-scale model pre-training module, enabling subsequent model iteration and optimization, and improving the accuracy of initial decision-making strategies. On the other hand, it provides data references for waterway management and logistics scheduling, contributing to the continuous improvement of the intelligent level of inland waterway transportation.

[0123] This second embodiment presents the complete workflow of the system, clearly explaining the implementation methods, collaborative relationships, and core functions of each component, which can effectively ensure that ships can achieve safe, efficient, and economical navigation in multi-destination transportation scenarios.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An inland waterway vessel route coordinate prediction and optimization system based on AI algorithms, characterized in that, Includes dynamic accessibility grid building blocks and a multi-task integrated planner: The dynamic accessibility grid construction unit is used to access and fuse multi-source heterogeneous data to generate and continuously update a spatiotemporal four-dimensional dynamic accessibility grid covering the entire target waterway. Each spatiotemporal grid of the dynamic accessibility grid stores a quantized value that represents the comprehensive accessibility status of the location within a specific time period. The multi-task integrated planner is connected to the dynamic accessibility grid construction unit and is used to receive logistics task orders containing multiple destination information. Based on the latest dynamic accessibility grid, the multi-task integrated planner performs an integrated decision-making process and outputs the globally optimal destination access sequence and the specific route coordinate sequence connecting each pair of consecutive destinations in the sequence. The integrated decision-making process is used to perform joint optimization calculations on destination access order decision and continuous spatial path search at the same time.

2. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to claim 1, characterized in that, The multi-source heterogeneous data includes static base data, near real-time dynamic data, planned event data, real-time traffic flow data, and ship-specific data; The static base data includes at least the fixed channel boundaries, fixed obstacle locations, static bridge clearance, and wharf locations recorded on the electronic navigation map. The near real-time dynamic data includes at least real-time water level data, wind direction and speed data, and temporary waterway control notices. The planned event data includes at least a lock opening and closing schedule and a timeline for the reservation and occupancy of anchorage berths. The real-time traffic flow data includes at least the real-time location, speed, and heading information of ships within the area obtained through the Automatic Identification System (AIS). The vessel's individual data includes at least the vessel's draft, height above the waterline, hull width, and current load status.

3. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to any one of claims 1 or 2, characterized in that, The dynamic mobility mesh building unit includes multiple parallel constraint quantizers and a weighted fusion unit; The constraint quantizer is used to convert the accessed multi-source heterogeneous data into corresponding spatiotemporal grid passage cost or risk probability values. The constraint quantizer includes at least a water depth constraint quantizer, a net height constraint quantizer, a time window constraint quantizer, a traffic conflict quantizer, and a resource status quantizer. The water depth constraint quantizer is used to calculate the effective water depth based on the static water depth and the real-time water level, and to generate a grounding risk probability value in combination with the ship's draft. The clearance constraint quantizer is used to calculate the actual clearance height under the bridge based on the bridge's static clearance height, real-time water level, and weather forecast data, and to generate a collision risk probability value by combining the ship's height above the waterline. The time window constraint quantizer is used to convert the lock opening and closing plan into a passable state Boolean value for the corresponding spatiotemporal grid. The traffic conflict quantizer is used to calculate the traffic density of each grid based on real-time traffic flow data using a kernel density estimation algorithm, and to predict potential conflict points to generate congestion delay cost values. The resource status quantizer is used to map the reservation occupancy timeline of anchorage berths to the resource occupancy cost value of the corresponding spatiotemporal grid. The weighted fusion unit receives the outputs of all constrained quantizers and calculates the comprehensive passage cost value of each spatiotemporal grid using a weighted summation model. The weighted summation model is expressed as: , in, Indicates spatial raster index, Indicates the time slice index. To constrain the total number of quantizers, For the first The weight coefficients corresponding to each constrained quantizer, and satisfying , Indicates the first The multi-source heterogeneous data processed by the constraint quantizer in the raster and time slices Data sets on Indicates the first A constraint quantizer maps input data to a function of cost or risk values.

4. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to claim 1, characterized in that, The multi-task integrated planner includes a high-level sequence optimization module and a low-level path search module, which together form a closed-loop interactive structure. The high-level sequence optimization module is implemented using a deep reinforcement learning framework based on an attention mechanism. Its state space is defined as the set of currently visited destinations, the ship's current position and time, and the attribute set of all unvisited destinations. Its action space is defined as selecting the next destination to be visited. Its reward value is calculated from the actual sailing cost and time cost fed back by the low-level path search module. The underlying path search module is implemented using a time-varying A* search algorithm. When the higher-level sequence optimization module outputs an action, i.e., selects the next destination, this module is activated. The underlying path search module uses the ship's current position as the search starting point and the selected destination as the search ending point to search on the dynamic navigability grid, looking for the path with the minimum sum of comprehensive travel costs, and outputs the precise coordinate sequence of the path, the estimated travel time, and the total cost. The total cost is fed back to the higher-level sequence optimization module in real time for reward calculation.

5. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to claim 4, characterized in that, The evaluation function of the time-varying A* search algorithm used in the underlying path search module. for: , in, This represents the current node in the search process, corresponding to a spatiotemporal grid in the dynamic décor grid. This indicates the distance from the starting point of the search to the current node. The actual cost already incurred, its value being from the starting point to... The total passage cost of all spatiotemporal grids traversed The sum of the sums, Indicates starting from the current node The estimated remaining cost to the search endpoint is calculated based on the distance-first principle of the basic navigation algorithm, using Euclidean geometric distance as a heuristic estimate to accelerate search convergence.

6. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to claim 4, characterized in that, The high-level sequence optimization module incorporates a deep reinforcement learning agent, whose reward function is... Designed as follows: ,in, This represents the estimated total travel time to complete the entire multi-destination mission, obtained by summing the estimated travel times for each segment of the route. This indicates the number of destinations in the task list that have hard time window constraints. This indicates that according to the current plan, the destination will be reached at the designated location. Estimated travel time for a destination with time window requirements. Indicates the first The latest arrival time required for each destination This represents the total travel cost of the segment corresponding to this action, as reported by the underlying path search module. , , These are the total time penalty coefficient, the time window violation penalty coefficient, and the navigation cost penalty coefficient, all of which are adjustable hyperparameters with a value greater than zero.

7. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to claim 1, characterized in that, The system also includes a pre-training module for a large model of flight route data; The route data big model pre-training module is used to pre-train the high-level sequence optimization module in the multi-task integrated planner in an offline state using historical ship route big data, so as to learn empirical destination access order patterns and channel selection preferences, and provide initial decision strategies for the high-level sequence optimization module.

8. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to claim 1, characterized in that, The system also includes a rolling replanning trigger module; The rolling replanning trigger module is connected to the dynamic navigability grid construction unit and the multi-task integrated planner, and is used to monitor the actual navigation status of the ship and the changes in key areas of the dynamic navigability grid; When a ship deviates from its original planned route by more than a predetermined threshold, or when the comprehensive passage cost of the grid within a predetermined range ahead of the ship in the dynamic passability grid changes by more than a preset threshold, the rolling replanning trigger module automatically starts the multi-task integrated planner. Starting from the ship's latest position and with the remaining unfinished tasks as the objective, it executes a new round of integrated decision-making process to generate an updated route coordinate sequence.

9. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to claim 1, characterized in that, The dynamic accessibility grid construction unit is also used to encode and integrate the spatiotemporal occupancy status of anchorage berths and locks as part of the planned event data into the dynamic accessibility grid; When making integrated decisions, the multi-task integrated planner can generate route coordinate sequences based on the resource occupancy costs encoded in the grid, and simultaneously plan out unoccupied spatiotemporal resource slots for ships at corresponding anchorages or locks that meet the expected arrival time, thereby achieving the collaborative generation and reservation of routes and key resources.

10. The inland waterway vessel route coordinate prediction and optimization system based on AI algorithm according to claim 4, characterized in that, The closed-loop interaction process between the high-level sequence optimization module and the low-level path search module is as follows: The high-level sequence optimization module outputs a probability distribution based on the current policy network and selects the next destination action accordingly. The underlying path search module is then invoked to calculate the optimal path to the destination, along with its cost and time, based on the current dynamic drivability grid. The calculated cost and time are encapsulated into a reward signal and returned to the high-level sequence optimization module; The high-level sequence optimization module uses the reward signal to update its policy network parameters and enters the next state. This process is repeated until all destinations have been visited, and finally outputs the destination visit sequence and each fine route segment that maximizes the cumulative reward.

Images