Big data-based passenger-roll transport demand prediction and ship intelligent scheduling method and system

Abstract

The invention relates to a big data-based passenger-roll transport demand prediction and ship intelligent scheduling method and system. The method comprises the steps of obtaining multi-source shipping data for a target area; inputting the multi-source shipping data into the spatial-temporal feature mining model, and predicting passenger rolling transportation demand information of the target area; acquiring ship real-time position, passenger carrying capacity, energy consumption data and port real-time operation state in the target area, and dynamically generating an optimal scheduling scheme by adopting a shipping scheduling model in combination with the predicted passenger transport demand information; the shipping scheduling model is obtained by interacting a decision scheduling model with an intelligent agent corresponding to the passenger roller transportation system and performing iterative training by adopting a reinforcement learning algorithm; and converting the optimal scheduling scheme into visual information, pushing the visual information to operation terminals of the ship and port workers so as to start corresponding shipping scheduling operation, and monitoring the execution effect of the shipping scheduling operation in real time.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent shipping, and particularly relates to a passenger and vehicle roll-on / roll-off (passenger and vehicle roll-on / roll-off, P&V RORO) transportation demand prediction and ship intelligent scheduling method and system based on big data. BACKGROUND

[0002] Passenger and vehicle roll-on / roll-off (P&V RORO) refers to a transportation mode in which a ship can simultaneously transport passengers and vehicles (such as cars, trucks, etc.), and is commonly used for short-distance sea transportation in straits, bays, etc.

[0003] In the related art, transportation scheduling schemes are mainly formulated by the coordinated cooperation of port staff and ship staff, and historical transportation schemes. The traditional mode relies on manual collection of port operation logs, ship reporting information and ticket statistics data, and the data is summarized and analyzed through Excel tables or localized systems, and the data update cycle is generally more than 4-6 hours. The timeliness is poor, the scheduling efficiency is low, and it is difficult to respond to sudden situations such as transportation demand surge, transportation peak, abnormal weather, etc. When a sudden typhoon, holiday passenger surge, etc. occurs, the dispatch personnel need to manually retrieve historical similar scene schemes, and after consultation by multiple departments, the number of trips can be adjusted. From demand perception to scheme landing, the average time consumption is 8-12 hours. For example, a large-scale passenger flow arrives in a strait area, and part of the flights are temporarily suspended due to heavy fog weather. The traditional scheduling system cannot integrate ticket booking, weather warning and port berth scheduling and other multi-source information in real time, and cannot adjust the ship trips and transportation capacity in time, resulting in that the port throughput cannot be fully released, and congestion is formed. Or, the scheduling adjustment caused by single-ship mechanical failure often leads to chain changes in the scheduling plans of multiple ships due to the lack of global resource visualization (such as the cargo carrying capacity of other ships and the status of the backup berth in the port), and the average recovery time is long.

[0004] Therefore, it is urgent to design a P&V RORO transportation demand prediction and ship intelligent scheduling scheme based on big data to solve at least one of the above technical problems. SUMMARY

[0005] The application provides a P&V RORO transportation demand prediction and ship intelligent scheduling method and system based on big data to realize automatic prediction of P&V RORO transportation demand and intelligent scheduling of ships.

[0006] In a first aspect, the application provides a P&V RORO transportation demand prediction and ship intelligent scheduling method based on big data, which comprises: acquiring multi-source shipping data for a target area; the multi-source shipping data at least includes historical ticket data, weather information, holiday traffic, and port operation data; The multi-source shipping data is input into a spatiotemporal feature mining model to predict passenger-rolling transportation demand information of a target area; wherein model parameters of the spatiotemporal feature mining model are obtained based on a regional function structure, a regional carrying capacity and a safe carrying capacity boundary setting in the target area; Real-time positions, passenger-carrying capacities, energy consumption data and port real-time operation states of ships in the target area are obtained, and the predicted passenger-rolling transportation demand information is combined to dynamically generate an optimal scheduling scheme by using a shipping scheduling model; wherein the shipping scheduling model is obtained by iteratively training a decision scheduling model and an agent corresponding to a passenger-rolling transportation system by using a reinforcement learning algorithm; The optimal scheduling scheme is converted into visual information and pushed to operation terminals of ship and port staff to start corresponding shipping scheduling operations, and the execution effect of the shipping scheduling operations is monitored in real time.

[0007] In a second aspect, an embodiment of the present application provides a passenger-rolling transportation demand prediction and ship intelligent scheduling system based on big data, which comprises: A collection unit configured to obtain multi-source shipping data for a target area; the multi-source shipping data at least includes historical ticketing data, weather information, holiday traffic, and port operation data; A prediction unit configured to input the multi-source shipping data into a spatiotemporal feature mining model to predict passenger-rolling transportation demand information of the target area; wherein model parameters of the spatiotemporal feature mining model are obtained based on a regional function structure, a regional carrying capacity and a safe carrying capacity boundary setting in the target area; An optimization unit configured to obtain real-time positions, passenger-carrying capacities, energy consumption data and port real-time operation states of ships in the target area, combine the predicted passenger-rolling transportation demand information, and dynamically generate an optimal scheduling scheme by using a shipping scheduling model; wherein the shipping scheduling model is obtained by iteratively training a decision scheduling model and an agent corresponding to a passenger-rolling transportation system by using a reinforcement learning algorithm; A scheduling unit configured to convert the optimal scheduling scheme into visual information, push it to operation terminals of ship and port staff to start corresponding shipping scheduling operations, and monitor the execution effect of the shipping scheduling operations in real time.

[0008] In a third aspect, an embodiment of the present application provides an electronic device, which comprises: At least one processor, a memory and an input-output unit; The memory is configured to store a computer program, and the processor is configured to call the computer program stored in the memory to execute the passenger-rolling transportation demand prediction and ship intelligent scheduling method based on big data of the first aspect.

[0009] In a fourth aspect, a computer-readable storage medium is provided, which includes instructions that, when executed on a computer, cause the computer to perform the big data-based passenger-rolling transportation demand prediction and ship intelligent scheduling method of the first aspect.

[0010] The technical scheme provided by the embodiments of the present application firstly acquires multi-source shipping data for a target area; the multi-source shipping data at least includes historical ticket data, weather information, holiday traffic, and port operation data. Then, the multi-source shipping data is input into a spatio-temporal feature mining model to predict passenger-rolling transportation demand information of the target area; wherein the model parameters of the spatio-temporal feature mining model are obtained based on the regional function structure, the regional carrying capacity, and the safety transport capacity boundary setting in the target area. Next, the real-time position of the ship in the target area, the passenger carrying capacity, the energy consumption data, and the real-time operation state of the port are acquired, and the optimal scheduling scheme is dynamically generated by using a shipping scheduling model in combination with the predicted passenger-rolling transportation demand information; wherein the shipping scheduling model is obtained by iteratively training the reinforcement learning algorithm through the interaction between the decision scheduling model and the agent corresponding to the passenger-rolling transportation system. Finally, the optimal scheduling scheme is converted into visual information and pushed to the operation terminal of the ship and port staff to start the corresponding shipping scheduling operation, and the execution effect of the shipping scheduling operation is monitored in real time. The technical scheme of the present application realizes the automatic prediction of passenger-rolling transportation demand and the intelligent scheduling of ships, optimizes the ship scheduling efficiency, improves the port operation quality monitoring efficiency and the port operation quality, and guarantees the port operation service quality and reliability. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings, which are included to provide a further understanding of the present application, constitute a part of the present application, illustrate the illustrative embodiments of the present application and their description serve to explain the present application, and do not constitute improper limitations on the present application. In the drawings: Figure 1 is a flowchart of a big data-based passenger-rolling transportation demand prediction and ship intelligent scheduling method according to an embodiment of the present application; Figure 2 is a structural diagram of a big data-based passenger-rolling transportation demand prediction and ship intelligent scheduling system according to an embodiment of the present application; Figure 3 is a structural diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0012] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terminology used in the description herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application.

[0014] The embodiments of the present application provide a big data-based passenger and roll transportation demand prediction and ship intelligent scheduling method and system.

[0015] Specifically, in the big data-based passenger and roll transportation demand prediction and ship intelligent scheduling scheme, the above-mentioned big data-based passenger and roll transportation demand prediction and ship intelligent scheduling method brings significant effects through a series of technical means from data processing, demand prediction, scheduling decision to scheme execution, and the specific effects are as follows: Multi-source shipping data is acquired, covering historical ticketing, weather, holiday traffic, port operation and other information. By integrating data of different dimensions, a more comprehensive data foundation is constructed. Compared with the traditional manual collection and Excel analysis, the data updating cycle is more than 4-6 hours. This method uses digital means to realize real-time data acquisition and processing, and the data updating delay is controlled within 500 ms, greatly improving the timeliness of the data. Multi-source data fusion also avoids the limitations of single data, reduces data errors through cross-validation and analysis, and provides more accurate basis for subsequent prediction and decision-making.

[0016] The multi-source shipping data is input into a spatio-temporal feature mining model. The model parameters are set based on the regional function structure, carrying capacity and safe transport capacity boundary. Combined with technologies such as multi-modal graph spatio-temporal network MSTGN, the periodic and trend changes of the data are captured, and the features are mined from multiple spatio-temporal dimensions. Compared with the traditional prediction error based on three-year average calculation and holiday passenger flow peak, the prediction error is reduced, and the RMSE is reduced from 35 people / ship to 15 people / ship. Precise prediction enables shipping enterprises to plan transport capacity in advance, reasonably arrange ship schedules and deploy port resources, and avoid resource waste or deficiency, such as reducing ship empty rate to within 10%, improving berth utilization rate to 78%, and improving resource utilization efficiency.

[0017] This method combines real-time vessel location, passenger capacity, energy consumption data, port real-time operation status, and demand forecast information to generate optimal scheduling schemes using a reinforcement learning-based shipping scheduling model. Traditional scheduling relies on manual experience and suffers from insufficient real-time vessel data collection. In the face of emergencies, such as typhoon warnings, manually notifying vessels to adjust their routes takes an average of 30 minutes and lacks a global perspective. This new method, through a multi-agent PPO algorithm, can complete global state encoding within 15 seconds and generate a scheduling scheme within 30 seconds. It can rapidly adjust scheduling strategies in response to changes in demand or emergencies; for example, during a typhoon, it can adjust the routes of all en route vessels within 5 minutes. The path deviation between the planned result and the recommended shelter anchorage is controlled within 500 meters, improving the system's ability to handle complex situations and ensuring safe and smooth shipping.

[0018] The optimal scheduling plan is converted into visual information and pushed to the operation terminal, making it easier for staff to intuitively understand the scheduling arrangements, reducing communication costs and misunderstandings, and improving execution efficiency. Real-time monitoring of shipping scheduling operations allows for simulation of scheduling effects and evaluation of plans through cloud-based data twins, identifying potential problems in advance. Real-time monitoring of the execution process enables timely local re-optimization and the push of compensation plans when deviations occur, such as automatically pushing free rebooking and delay compensation calculations for delayed vessels, improving service quality and customer satisfaction. Traditional scheduling lacks visualization and real-time monitoring, resulting in long recovery times for abnormal events, such as single-vessel malfunctions leading to cascading scheduling changes, with an average recovery time of up to 2 hours. This method significantly improves this situation.

[0019] The big data-based passenger and vehicle ferry demand forecasting and intelligent ship scheduling scheme provided in this application embodiment can also be executed by electronic devices, such as servers, server clusters, or cloud servers. These electronic devices can also be terminal devices such as mobile phones, computers, tablets, wearable devices, or dedicated devices (such as dedicated terminal devices with a big data-based passenger and vehicle ferry demand forecasting and intelligent ship scheduling system). These electronic devices can also carry the chips described in the above embodiments. Alternatively, these electronic devices can also install service programs for executing the big data-based passenger and vehicle ferry demand forecasting and intelligent ship scheduling scheme.

[0020] Figure 1 A schematic diagram illustrating a big data-based method for predicting passenger and roll-on / roll-off (Ro-Ro) transport demand and intelligent ship scheduling, as provided in this application embodiment, is shown below. Figure 1 The method includes the following steps: 101. Obtain multi-source shipping data for the target area.

[0021] In this embodiment of the application, the multi-source shipping data includes at least: historical ticketing data, meteorological information, holiday traffic, and port operation data.

[0022] In the ro-ro transportation scenario, multi-source shipping data collection in the target area is the basis for intelligent scheduling. Historical ticket data covers ticket sales, passenger types (individuals, groups, etc.), ticket purchase time, and refund records on different dates and routes. These data can intuitively reflect the demand size and trend of ro-ro transportation in the past period, for example, by analyzing historical ticket data, we can find the demand peak rules of fixed dates every week and specific time periods every month.

[0023] Weather information includes real-time and historical wind speed, wind direction, temperature, precipitation, wave height, and visibility, which has a crucial impact on ro-ro transportation. For example, strong winds, high waves, and low visibility can directly limit the speed and safety of ship navigation, even leading to suspension. Therefore, accurate weather information not only helps to predict fluctuations in transportation demand caused by weather changes, but also provides key basis for ship navigation safety and scheduling decisions.

[0024] Holiday traffic data includes holiday arrangements, passenger flow, vehicle transportation, and source and destination distribution on various routes during holidays. Since holidays are often accompanied by significant growth in travel demand, and travel patterns differ greatly from daily routines, this data can help predict transportation demand size and structure during holidays, so as to make advance preparations for transportation capacity allocation and service support.

[0025] Port operation data involves berth usage (berth occupancy rate, idle time), cargo handling efficiency, port operation time, port throughput, equipment operation status (usage frequency and failure rate of cranes, forklifts, etc.), and personnel scheduling. These data reflect the port's operation capacity and efficiency, and are an important reference for assessing port carrying capacity, optimizing ship docking plans, and allocating port resources. For example, based on berth usage and handling efficiency, ship arrival times can be reasonably arranged to avoid port congestion or long waiting times for ships.

[0026] Through comprehensive collection of these multi-source shipping data, rich and accurate information can be provided to support subsequent demand prediction and intelligent scheduling.

[0027] 102, input the multi-source shipping data into the spatio-temporal feature mining model to predict the ro-ro transportation demand information in the target area.

[0028] In the embodiments of the present application, the model parameters of the spatio-temporal feature mining model are obtained based on the regional function structure, regional carrying capacity, and safety transport capacity boundary settings in the target area.

[0029] The spatio-temporal feature mining model fuses prior constraints such as regional functional structure (such as port distribution, route network, passenger and cargo transport distribution point layout), regional carrying capacity (maximum port throughput, maximum ship carrying capacity) and safe carrying capacity boundary (minimum safety distance limited by maritime safety regulations, maximum load threshold) to construct a model parameter space with physical meaning. The core principle is to map multi-source data such as historical ticket data and weather information into a spatio-temporal graph structure based on a multi-modal graph spatio-temporal network (MSTGN). Taking ports and ships as spatial nodes and route connectivity and passenger flow tidal rules as dynamic edges, the algorithm captures periodic features (such as holiday traffic surges and weekend travel peaks) and trend features (such as seasonal tourist passenger flow changes) at the year / month / day / hour scales through a time sliding window.

[0030] The spatio-temporal feature mining model uses a hierarchical cross-attention mechanism to fuse port-level, route-level and ship-level multi-scale features, combines multi-task learning to simultaneously predict passenger transport demand, cargo transport demand, ship scheduling demand and abnormal delay events, and improves prediction linkage by sharing underlying features. The algorithm integrates dynamic graph learning mechanism (self-adaptive adjustment of node connection weight), spatio-temporal joint coding (capture of spatial proximity and time series dependence) and constraint embedding technology (marine fluid equation, safe carrying capacity boundary as loss function regular term) to realize spatio-temporal coupling feature analysis of passenger and roll-on / roll-off transport demand, and finally output fine-grained prediction results including passenger flow, vehicle flow, time period distribution and port pressure, providing forward-looking data support for scheduling decisions.

[0031] As an optional embodiment, before inputting the multi-source shipping data into the spatio-temporal feature mining model in 102 to predict the passenger and roll-on / roll-off transport demand information of the target region, the method further includes: Using a multi-modal graph spatio-temporal network (MSTGN), a spatial graph structure is constructed by integrating a multi-source shipping data training set; and a time series is divided into time slices, and a spatial node connection relationship in the time dimension is constructed through a time sliding window to capture periodic and trend changes in the multi-source shipping data training set to obtain a time graph structure. The spatial graph structure and the time graph structure are spatio-temporally coded, and a hierarchical cross-attention mechanism is used to fuse the spatio-temporally coded results to obtain multi-scale spatio-temporal joint features; wherein the spatial scale includes port level, route level and ship level, and the time scale includes year, month, day and hour. Using the multi-scale spatio-temporal joint features, a multi-task learning mechanism is used to predict passenger transport demand, cargo transport demand, ship scheduling demand and abnormal delay events, and the underlying shared prediction results are used to train the prediction process of the passenger and roll-on / roll-off transport demand information to obtain the spatio-temporal feature mining model.

[0032] In the preprocessing procedure of constructing the spatio-temporal feature mining model by the multi-modal graph network (MSTGN), the core principle is to model the multi-source shipping data from the spatial and temporal dimensions. First, taking the port, ship, passenger and cargo transport area as the spatial node, the spatial adjacency matrix is constructed by combining the route connectivity and functional interaction relationship, and the multi-modal data such as topological relationship, ticket time sequence and weather grid are integrated through the dynamic graph learning mechanism to generate a spatial graph structure that can adaptively adjust the node relationship and modal contribution. At the same time, the time series is divided into time slices, and the sliding window is used to capture the connection relationship of the spatial node under different time granularity (year / month / day / hour) to construct a time graph structure reflecting periodicity (such as holiday traffic rules) and trend (such as seasonal demand changes). Then, the spatial graph and the time graph are fused by spatio-temporal joint coding, and the hierarchical cross-attention mechanism is used to interact features among multiple spatial scales (port level, route level, ship level, etc.) and multiple time scales, to extract multi-scale joint features containing spatio-temporal dependency. Finally, based on the multi-task learning mechanism, the passenger demand, freight demand, ship scheduling demand and abnormal delay events are predicted simultaneously, the cross-task information complementation is realized through the bottom layer shared prediction results, and the loss function is optimized combined with the regional function structure, safety capacity boundary and other constraint conditions, so that the model can accurately capture the spatio-temporal dynamic rules of passenger and roll transport demand.

[0033] The technical effects of the scheme are that the graph network modeling is used to solve the heterogeneity and spatial correlation of multi-source data, the dynamic attention mechanism is used to adaptively allocate data weights, and the representation ability of the model for complex shipping scenarios is improved; multi-scale feature fusion realizes cross-level demand prediction from macro regional bearing to micro ship scheduling, and enhances the granularity adaptability of the prediction results; the multi-task learning and constraint embedding mechanism ensures that the prediction results not only conform to the business logic (such as port throughput limit, fluid mechanics constraint), but also enhance the prediction robustness through cross-task association, so that the spatio-temporal feature mining model can accurately predict the passenger and roll transport demand and provide reliable input for intelligent scheduling.

[0034] It can be understood that the multi-modal graph spatio-temporal network (MSTGN) is a spatio-temporal feature modeling method for multi-source heterogeneous data, mainly used for integrating multi-dimensional data in the passenger-cargo transport scene and mining the spatio-temporal correlation law. The network first maps the multi-source shipping data training set such as historical ticket data, weather information, holiday traffic, and port operation data into a spatial graph structure, in which the nodes can represent ports, ships, routes, and other entities, and the edges represent the spatial correlation between entities (such as the connection between ports, ship scheduling paths, etc.). At the same time, by dividing the time series into hours, days, months, years, etc., the sliding window mechanism is used to construct the dynamic connection relationship of the nodes in the time dimension to form a time graph structure, so as to capture the periodic fluctuations (such as holiday traffic peaks) and trend changes (such as seasonal transportation rules) of passenger-cargo transport demand at different time granularities.

[0035] At the level of spatio-temporal feature fusion, MSTGN encodes the spatial graph structure and the time graph structure through a hierarchical cross-attention mechanism, extracts multi-modal features at spatial scales such as port level, route level, and ship level, and at time scales such as year, month, day, and hour, and dynamically allocates weights to features of different scales through an attention mechanism to achieve fine modeling of spatio-temporal interaction patterns. For example, it can identify the cascading impact of ship scheduling delays in extreme weather on regional route networks, or analyze the spatio-temporal matching relationship between high-frequency flights and port carrying capacity during holidays.

[0036] Based on multi-scale spatio-temporal joint features, MSTGN combines a multi-task learning mechanism to simultaneously predict passenger demand, cargo demand, ship scheduling demand, and abnormal delay events. The bottom layer shares feature extraction to improve model efficiency, and the upper layer uses task-specific output layers to achieve multi-target prediction. This design enables the model to capture the complex coupling relationship between multiple elements in the passenger-cargo transport system from a global perspective, such as by analyzing the association between passenger travel patterns in historical ticket data and weather data to predict passenger demand fluctuations in extreme weather and identify potential ship delay risks in advance, providing accurate demand prediction for subsequent intelligent scheduling, ultimately improving the response capability and prediction accuracy of the passenger-cargo transport system to spatio-temporal dynamic changes.

[0037] Further optionally, in the above steps, a multi-modal graph spatio-temporal network (MSTGN) is used to integrate a multi-source shipping data training set to construct a spatial graph structure, including: The ports, ships, passenger transport areas, cargo transport areas, and service areas in the target area are taken as spatial nodes, and the route connectivity, spatial functional relationship, and spatial functional interaction relationship are taken as connection edges to construct a spatial adjacency matrix. Based on the spatial adjacency matrix, the port topology relationship in the multi-source shipping data training set, the historical ticket time sequence relationship, the weather grid, the holiday passenger flow change trend, and the historical port operation relationship are integrated to establish an adaptive adjacency matrix based on a dynamic graph learning mechanism. The contribution of different modal data in the multi-source shipping data training set and the node relationship in the adaptive adjacency matrix are dynamically adjusted by the attention dynamic weight to obtain a spatial graph structure.

[0038] The core principle of constructing the spatial graph structure is to realize the heterogeneous integration and spatial correlation modeling of multi-modal data through dynamic graph modeling. First, the port, ship, and passenger and cargo transport area are taken as spatial nodes, the initial spatial adjacency matrix is constructed based on the physical connectivity of the route (such as the direct route between ports), the spatial functional relationship (such as the passenger flow distribution relationship between the passenger transport area and the port), and the spatial functional interaction relationship (such as the loading and unloading efficiency correlation between the cargo transport area and the ship), and the basic topological connection between entities is described. On this basis, a dynamic graph learning mechanism is introduced, and multi-modal data such as port topology relationship (static spatial structure), historical ticket time sequence relationship (passenger flow time sequence), weather grid data (spatial weather element distribution), holiday passenger flow trend (periodic behavior characteristics), and historical port operation data (such as berth turnover rate and ship berthing time) are embedded in the adjacency matrix. The contribution of different modal data to the node relationship is dynamically calculated by the attention mechanism. For example, in the extreme weather scenario, the weight of the weather grid data will significantly increase, and the model can accordingly strengthen the correlation between the port node and the weather element and dynamically adjust the scheduling priority of the ship node. During the holiday, the weights of the historical ticket time sequence and the passenger flow trend data increase, and the model focuses on capturing the high-frequency interaction relationship between the passenger transport area and the port node.

[0039] Through the dynamic adjustment of the adaptive adjacency matrix, the model can reflect the dynamic evolution of the spatial entity relationship in the passenger-rolling transport system in real time, such as updating the traffic state (closed / open) of the route between ports according to real-time weather data, or dynamically weakening the connection strength between the congested port and other nodes based on the port operation efficiency data. This dynamic modeling mechanism solves the problem that the traditional static graph model cannot adapt to the real-time changes of multi-source data in the shipping scenario, enabling the spatial graph structure to adaptively integrate multi-modal information and accurately capture the key associations between entities in different operation scenarios, providing a more semantically represented spatial input for subsequent spatio-temporal feature fusion, and ultimately improving the analysis accuracy and dynamic adaptability of passenger-rolling transport demand prediction for complex spatial interaction patterns.

[0040] The dynamic graph learning mechanism usually integrates graph neural networks (GNN), attention mechanisms (Attention), and dynamic parameter updating algorithms in multi-modal graph spatio-temporal networks (MSTGN), and the core goal is to enable the graph structure to dynamically adjust node connections and weights over time or data modalities.

[0041] By dynamically calibrating the influence of multi-modal data on spatial connection through attention mechanism, and combining with gate or convolution operation to realize adaptive evolution of graph structure, the model can capture the dynamic coupling relationship between entities such as "weather-port-ship" in passenger and roll transport in real time (such as when a port is temporarily closed due to heavy rain, the model automatically weakens the connection between the port and other nodes, and enhances the connection weight of the surrounding alternative port), thereby improving the representation ability of spatial graph structure for complex scenarios, and providing more accurate input basis for subsequent spatio-temporal joint feature extraction.

[0042] Further optionally, after the above steps, the multi-scale spatio-temporal joint feature is used to predict passenger transport demand, freight transport demand, ship scheduling demand, and abnormal delay event in combination with a multi-task learning mechanism, and after the bottom layer shared prediction result is fused to obtain the passenger and roll transport demand information, the method further comprises the following steps: Discretize the ocean fluid equation Navier-Stokes into difference form, and couple with meteorological data to generate a ship navigation resistance coefficient matrix Combine the fluid mechanics mechanism to embed the ocean fluid equation Navier-Stokes into the reward function of the spatio-temporal feature mining model, so as to improve the robustness of the spatio-temporal feature mining model in bad sea conditions; Wherein, the ship navigation resistance coefficient matrix is used to calculate the square modulus of the residual of the Navier-Stokes equation, and the calculation process of the ship navigation resistance coefficient matrix is represented as: is a three-dimensional sea area flow velocity vector field, is used to simulate the navigation resistance, represents the convection term describing the transport effect of the fluid due to its own speed, is the static pressure field of the fluid, is related to the depth and density of the ocean current, represents the pressure gradient term, is the sea water viscosity coefficient, represents the viscosity term reflecting the influence degree of fluid viscosity on movement; Encode the regional functional structure as a discrete variable, take the regional carrying capacity and the safety transport capacity boundary as the linear inequality constraint condition, and take the port throughput limit as the hard constraint, and embed it into the loss function; the calculation process of the loss function is represented as the following formula: Wherein, , , are all adaptively adjusted according to the violation degree of the corresponding constraint loss term, so as to ensure that the prediction result meets the physical rules and safety boundary; is a basic prediction loss used to represent the error between the model prediction value and the true value, ​​a physical constraint loss representing the ship navigation resistance correlation, a load over-limit penalty, is the actual load containing passenger flow and cargo volume, is the regional maximum carrying capacity, a ship over-limit penalty, is the actual number of ships in the target area, is the safe transport capacity boundary containing the anchorage capacity upper limit.

[0043] After the spatio-temporal feature mining model completes multi-scale spatio-temporal joint feature fusion and multi-task prediction, the embedding of physical mechanism and safety constraint improves the reliability of the model in complex scenarios. The specific principle is as follows: First, the Navier-Stokes equation in ocean fluid mechanics is discretized into a difference form and coupled with meteorological data (such as wind speed, sea wave height) to construct a ship navigation resistance coefficient matrix. This matrix quantifies the influence of fluid motion on ship navigation by calculating the square modulus of the equation residual, where the convection term reflects fluid transport effects, the pressure gradient term is related to the static pressure field, and the viscosity term embodies the effect of seawater viscosity. Embedding this matrix into the reward function of the model enables the model to learn the prediction rules in adverse sea conditions (such as high resistance scenarios) during training, and guides the model output to conform to the principles of fluid mechanics, enhancing robustness.

[0044] Secondly, for regional operation safety, the functional structure of the target area (such as port type, route level) is coded as a discrete variable, and the regional carrying capacity (passenger flow / cargo volume upper limit) and the safe transport capacity boundary (anchorage capacity, ship density limit) are used as linear inequality constraints, and the port throughput limit is used as a hard constraint embedded in the loss function. The loss function consists of a basic prediction loss, a physical constraint loss, a load over-limit penalty, and a ship over-limit penalty, where the adaptive weight factor is dynamically adjusted according to the constraint violation degree. For example, when the predicted port throughput exceeds the physical limit, the load over-limit penalty term automatically increases, forcing the model to correct the prediction value; if the number of ships breaks the safe transport capacity boundary, the ship over-limit penalty term triggers, ensuring that the prediction result does not break the safety threshold.

[0045] In this way, through the physical embedding of the Navier-Stokes equation, the model can capture the dynamic influence of natural factors such as sea waves and ocean currents on the roll-on / roll-off transportation, especially in adverse weather such as typhoons and heavy rain. The model can adjust the demand prediction based on the principle of fluid mechanics (such as reducing the scheduling of ships on high-resistance routes) to avoid the prediction bias caused by the lack of samples in pure data-driven models. The introduction of safety constraints ensures that the prediction results meet the actual carrying capacity from the perspective of operation management, preventing problems such as "over-scheduling" or "overload" that are theoretically feasible but practically inoperable, and improving the engineering practicality and safety of the prediction results. The combination of the two realizes the fusion of data-driven and physical mechanism, making the model have both data fitting ability and meet physical rules and operation specifications in complex environments, ultimately improving the accuracy and reliability of roll-on / roll-off transportation demand prediction.

[0046] As an optional embodiment, before inputting the multi-source shipping data in 102 into the spatio-temporal feature mining model to predict the roll-on / roll-off transportation demand information in the target area, it further includes: From the target area, identify a newly built port area with a historical data volume less than a set sparsity threshold; use a meta-learning algorithm to construct a meta-training data set containing multi-source data of mature port areas and corresponding demand prediction results; use a meta-learner to learn a general feature representation across ports using a gradient update strategy, so that the model parameters of the spatio-temporal feature mining model have the ability to quickly adapt to newly built port areas; for newly built port areas, obtain the first month's ticket data and port basic layout in the initial data, and generate model parameters adapted to the spatio-temporal feature mining model through the meta-learner to realize rapid generalization of demand prediction for newly built port areas and reduce prediction errors in the cold start stage.

[0047] Among them, the specific principle of the optional embodiment is to solve the cold start problem of insufficient data in newly built ports through the Meta-Learning mechanism.

[0048] Firstly, identify new ports with historical data volume below the sparsity threshold in the target area, and construct a meta-training dataset containing multi-source data of mature ports (historical ticketing, weather, port operation, etc.) and corresponding demand prediction results. Then design a meta-learner using the Model-Agnostic Meta-Learning (MAML) framework, which learns the general feature representation across ports (such as holiday passenger flow fluctuation patterns, weather influence rules, etc.) in the meta-training phase through the "gradient update strategy", enabling the initial parameters of the spatio-temporal feature mining model to quickly adapt to new environments. For new ports, only a small amount of initial data (such as the first month's ticketing data and basic layout information) is required, and the meta-learner can quickly fine-tune the parameters through internal loops to generate model weights adapted to the region, achieving the transfer from "known port knowledge" to "unknown port prediction". This method transforms the traditional "zero training" into "rapid adaptation based on existing knowledge", and the essence is to learn "how to learn" to significantly reduce the dependence on large amounts of historical data.

[0049] Compared to traditional methods that require months of data accumulation to achieve stable prediction accuracy in new port scenarios, this scheme uses meta-learning to transfer mature port knowledge, and only needs a single month of initial data to control the prediction error (RMSE) within 15% (the first month error of traditional methods usually exceeds 40%), improving the cold start efficiency. By learning common features across regions, the model can still capture demand change patterns under data scarcity conditions. For example, the passenger transport demand prediction error of a new port during typhoon weather decreases, and the ship scheduling demand prediction accuracy improves, effectively avoiding resource waste (such as reducing empty running rate) caused by insufficient data. The parameter initialization point generated by the meta-learner is closer to the optimal solution, making the model more robust when facing data distribution shifts. For example, when a large-scale event suddenly occurs around a new port, causing a surge in passenger flow, the model can quickly adjust to the new distribution, with a prediction error increase of only 5%-8% (the traditional model error may double). Without waiting for a long time for data accumulation, new ports can obtain reliable demand prediction at the initial stage of opening, optimizing resource allocation in advance. For example, a new port uses this scheme to plan berth usage in advance, reducing congestion incidents and improving loading and unloading efficiency in the first quarter.

[0050] This meta-learning-based rapid generalization mechanism enables the spatio-temporal feature mining model to maintain high-precision prediction capability for mature ports while breaking the data scarcity limitation for new regions, providing key support for the dynamic expansion of passenger-rolling transport systems.

[0051] 103, Obtain real-time ship location, passenger capacity, energy consumption data, and port real-time operation status within the target area, combine with predicted passenger-rolling transport demand information, and use a shipping scheduling model to dynamically generate an optimal scheduling scheme.

[0052] In the embodiments of the present application, the shipping scheduling model is obtained by interacting with the agents corresponding to the roll-on / roll-off transportation system through the decision scheduling model and iteratively training using a reinforcement learning algorithm.

[0053] The shipping scheduling model is constructed based on a reinforcement learning framework. The core principle is to interact with the agents (such as ships and ports) in the roll-on / roll-off transportation system in real time through the decision scheduling model, forming a closed loop of "state perception - decision generation - environmental feedback - model optimization". Specifically, the model first integrates dynamic information such as real-time ship position, passenger carrying capacity, energy consumption data, and real-time port operating status with predicted roll-on / roll-off transportation demand information (such as passenger volume, freight volume, and ship scheduling demand) into a multi-dimensional state space. The agents generate scheduling actions (such as ship route adjustment, port arrival sequence arrangement, and transport capacity allocation) in this state space through a policy network. After the actions are executed, the system environment (such as port congestion status, ship energy consumption changes, and demand satisfaction) generates feedback reward signals (such as on-time rate improvement rewards and energy consumption penalties). The model iteratively optimizes the policy network parameters through a reinforcement learning algorithm, enabling the agents to learn scheduling strategies that maximize cumulative rewards in the long term, thereby achieving the goal of dynamically generating optimal scheduling solutions. This process simulates the mechanism of human dispatchers optimizing decisions through experience accumulation in complex dynamic environments, but with higher real-time performance and global optimization capabilities.

[0054] Through the combination of data-driven and intelligent optimization, the model significantly improves the operational efficiency and stability of the roll-on / roll-off transportation system. In addition, the model learns from historical scheduling data through continuous learning, adapting to different seasonal and holiday demand fluctuation patterns, which improves the on-time rate of passenger transportation and reduces the delay rate of freight transportation during peak seasons, significantly enhancing the system's robustness in complex scenarios. In the long term, the application of the model can reduce the operating costs of roll-on / roll-off transportation enterprises while improving customer satisfaction and market competitiveness.

[0055] Further optionally, in the above steps, the step of interacting with the agents corresponding to the roll-on / roll-off transportation system through the decision scheduling model and iteratively training using a reinforcement learning algorithm to obtain the shipping scheduling model further comprises: An agent in the roll-on / roll-off transportation system is created for each ship and each port region in the target area. The ship agent state space includes real-time position, remaining passenger capacity, cargo carrying capacity, current energy consumption rate, estimated arrival time, and route planning path. The port agent state space includes berth occupancy, availability of loading and unloading equipment, cargo backlog, and current period demand forecast value. An action space is constructed, wherein the executable actions of the ship agent include adjusting the sailing speed, applying for a berth change, and requesting priority loading and unloading, and the executable actions of the port agent include assigning a berth to a specified ship, scheduling loading and unloading equipment, and issuing a port congestion warning. The positive reward function, the negative reward function and the global reward function are set; the positive reward function is constructed with the goal of improving ship utilization, reducing empty load rate, reducing port congestion index and reducing energy consumption below expectation, the negative reward function is constructed with the goal of ship waiting time exceeding threshold, berth allocation conflict and safety boundary, and the global reward function introduces a cooperative reward item based on the communication mechanism between agents, and the cooperative reward item at least reduces the ship sailing of adjacent port agents, and a score is set for each reward; An initialization model of the shipping scheduling model is constructed, wherein the initialization model includes a passenger carrying capacity model, a ship energy consumption model and a port loading and unloading model. Multi-agent proximal policy optimization (PPO) is adopted, each agent in the ro-ro passenger transport system is interacted with the initialization model to obtain real-time state data matched by each agent; in combination with an action space, a positive reward function, a negative reward function and a global reward function, an initialization strategy network composed of a passenger carrying capacity model, a ship energy consumption model and a port loading and unloading model is used to predict the action state of each agent in the next period, and the action state is added to a shared experience pool; a reinforcement learning training mechanism is used to iterate the action state in the shared experience pool and the prediction process until a preset convergence condition is met to output a final shipping scheduling model. The clip strategy is used to update the strategy network, and the degree of violation of the safety capacity boundary by the shipping scheduling model is evaluated to correct abnormal logic in the shipping scheduling model.

[0056] In the optional embodiment, the shipping scheduling model is trained by constructing a multi-agent system and combining a reinforcement learning algorithm. First, an agent is created for each ship and each port region in the target region. The state space of the ship agent includes real-time position, remaining passenger and cargo carrying capacity, energy consumption rate, expected arrival time and route planning path, etc. The state space of the port agent includes berth occupancy, loading and unloading equipment availability, cargo backlog and demand forecast, etc. to fully depict the dynamic characteristics of the system entities. In the action space design, the ship agent can perform actions such as adjusting speed, applying for berth change, and requesting priority loading and unloading, and the port agent can perform operations such as berth allocation, equipment scheduling, and congestion warning, to realize the active regulation of system resources by the agent. The reward function system is divided into positive reward, negative reward and global reward. The positive reward focuses on improving ship utilization, reducing empty load rate, relieving port congestion and controlling energy consumption, etc. The negative reward is aimed at ship waiting time exceeding threshold, berth allocation conflict, safety boundary breakthrough, etc. The global reward introduces a cooperative reward item (such as a reward of a certain score for adjacent port cooperation to reduce ship sailing) through the communication mechanism between agents to encourage the cooperation between agents and balance individual optimization and global efficiency.

[0057] The initialization model integrates the passenger carrying capacity model, the ship energy consumption model, and the port loading and unloading model to form a basic strategy network. In the training process, the multi-agent proximal policy optimization (PPO) algorithm is used. Each agent interacts with the initialization model to obtain real-time state data. The strategy network predicts the next time period action state and stores it in the shared experience pool. Based on the reinforcement learning mechanism, the action state and the prediction process in the experience pool are iteratively optimized until the preset convergence condition (such as stable reward value or error below threshold) is met, and the final model is output. At the same time, the clip strategy is used to update the strategy network. The model's violation of the safe capacity boundary is evaluated to correct abnormal logic and ensure that the scheduling scheme meets the physical constraints and safety rules.

[0058] The multi-agent modeling realizes the fine characterization of the dynamic state of ships and ports. The design of the action space and the reward function closely matches the shipping business needs, enabling the model to learn optimization strategies that improve resource utilization, reduce energy consumption, and reduce congestion. The collaborative reward term in the global reward promotes collaboration among agents, reducing inefficient behaviors such as ship detours and improving the overall efficiency of the regional shipping system. The PPO algorithm combined with the shared experience pool training mechanism enhances the model's adaptability and generalization ability in complex dynamic environments. The clip strategy and safety boundary evaluation ensure the safety and feasibility of the scheduling scheme, and the final optimal scheduling scheme can balance efficiency, energy consumption, and safety in real-time dynamic environments, improving the intelligent scheduling level of the ro-ro passenger transport system.

[0059] Further optionally, in 103, the real-time positions of ships in the target area, the passenger carrying capacity, the energy consumption data, and the real-time operating status of the port are obtained. Combined with the predicted ro-ro passenger transport demand information, a shipping scheduling model is used to dynamically generate an optimal scheduling scheme, including: The real-time positions of ships in the target area, the passenger carrying capacity, the energy consumption data, the real-time operating status of the port, and the ro-ro passenger transport demand information are obtained to construct the current global state in the target area; Each agent generates its corresponding action state through the strategy network based on the corresponding local real-time state in the current global state and the global collaborative state with other agents; The action states of all agents are integrated to verify whether the safe capacity boundary and the port throughput limit in the target area are met, and an optimal scheme including ship schedule adjustment, route optimization, and port resource allocation is generated.

[0060] For example, if the global collaborative information includes the utilization rate of the berths of adjacent ports, the predicted action state includes: a ship agent applying for early berthing at an idle berth of a port, or a ship agent applying for delayed berthing at an idle berth of a port.

[0061] In optional embodiments, the specific principle is to construct a hierarchical decision-making mechanism to generate a scheduling scheme through the collaborative optimization of global state perception and local decision-making of agents. First, the system integrates real-time ship position, passenger and cargo carrying capacity, energy consumption data, port operation status and demand prediction information to construct a global state matrix containing multi-dimensional information such as space, time and physical constraints. Based on the global state, each agent (ship / port) extracts its own local real-time state (such as the current load of the ship, the berth occupancy rate of the port), and obtains the global collaborative state (such as the resource usage of adjacent ports) through the message passing mechanism, forming a complete understanding of the environment.

[0062] In the strategy execution phase, the ship agent generates action decisions (such as adjusting speed, applying for berth change) through a strategy network (such as LSTM or Transformer), and the port agent synchronously generates resource allocation strategies (such as berth scheduling, device priority). These decisions are not generated independently, but are weighted by the global collaborative state through an attention mechanism. For example, when a ship applies for a berth, it will focus on the congestion index of adjacent ports, and when a port allocates resources, it will consider the energy consumption state of the ship and the urgency of demand. The system then integrates all agent actions, ensures that the scheme meets the safety and capacity boundaries (such as ship density threshold and port throughput limit) through a constraint verification module, and finally generates an optimal scheme including schedule adjustment, route optimization and resource allocation.

[0063] In this way, in terms of sudden response capability, the model can respond to sudden changes (such as ship mechanical failure, passenger flow surge) through real-time data updates of the global state, with improved response capability compared to traditional scheduling systems; in terms of global optimization capability, the collaborative decision-making of adjacent port agents improves berth utilization and shortens the average waiting time of ships; in terms of safety compliance, the constraint verification mechanism ensures that the scheme meets the safety boundaries, such as automatically following the maximum throughput limit of the port to avoid disruptions due to overloading. For example, in a strait passenger-rolling transportation scenario, this scheme improves the accuracy of ship scheduling during typhoons and fog, reduces the single scheduling decision-making time from 40 minutes to 3 minutes, and does not have any safety violations, demonstrating strong complex scenario response capability and engineering practicality.

[0064] 104, convert the optimal scheduling scheme into visual information, and push it to the operation terminal of the ship and port staff to start the corresponding shipping scheduling operation, and monitor the execution effect of the shipping scheduling operation in real time.

[0065] As an optional embodiment, in 104, the optimal scheduling scheme is converted into visual information and pushed to the operation terminal of the ship and port staff to start the corresponding shipping scheduling operation, including: Based on the target area of the port three-dimensional model, route geographic information, ship dynamics parameters, cloud-based digital twin matching with the target area; Real-time synchronization of port operation data, ship sensor data, and weather data in the target area to the digital twin; the port operation data at least includes berth status, loading and unloading progress, the ship sensor data at least includes ship position, ship speed in and out of the port, detention time, ship energy consumption, ship supply condition, ship passenger capacity, and ship cargo capacity, and the weather data at least includes wind speed and wave height; After each start of the optimization scheduling, the optimal scheduling scheme is mapped to the data twin, triggering virtual ship navigation simulation and port operation process, monitoring ship arrival time error, berth idle time change, and energy consumption prediction deviation in the simulation scheduling process, and using the monitoring results to evaluate the optimal scheduling scheme; if the monitoring results exceed the safety boundary or the port congestion is aggravated, the optimal scheduling scheme is regenerated until it passes the evaluation; For the optimal scheduling scheme that passes the evaluation, according to the mapping information of each scheduling node in the data twin, visual instructions are generated according to the format requirements of the ship and port operation terminal and are issued to the corresponding operation terminal.

[0066] Among them, the specific principle of the optional embodiment is to realize the visual verification and accurate execution of the scheduling scheme through digital twin technology. First, based on the port three-dimensional model, route geographic information, and ship dynamics parameters (such as load-speed curve), a cloud digital twin is constructed to form a virtual mirror of the target area passenger-roller transportation system. Through real-time data interface, the dynamic data of port berth status, loading and unloading progress, ship position, energy consumption, and weather wind speed and wave height are synchronized to ensure that the virtual environment is consistent with the physical world state. After generating the optimal scheduling scheme each time, the system injects the scheme parameters (such as ship expected arrival time, berth allocation instruction) into the digital twin, triggering virtual ship navigation simulation and port operation process, and real-time monitoring of key indicators in the simulation process: ship arrival time error (within ±10 minutes), berth idle time change (more than 10% reduction after optimization), and energy consumption prediction deviation (not more than 10%). If the simulation finds that the safety boundary is broken (such as ship load exceeding limit) or the port congestion is aggravated, it automatically reverts to the scheduling decision module to regenerate the scheme, forming a "simulation-evaluation-correction" closed loop.

[0067] Through the evaluation scheme, the mapping relationship of the scheduling node in the digital twin (such as the correspondence between the ship ID and the navigation system, and the correspondence between the port berth and the scheduling screen) is used to generate visual instructions adapted to different terminals: the ship terminal receives dynamic navigation information containing route adjustment path (electronic chart annotation) and speed suggestion (dashboard prompt); the port terminal obtains berth allocation Gantt chart and unloading equipment scheduling list (with priority marking). At the same time, the system synchronously deploys the monitoring scheme, collects the execution data through the ship sensors (such as GPS and energy consumption meter) and port IoT devices (such as RFID card reader and infrared sensor), compares it with the expected state in the digital twin, and forms a "execution-feedback-tuning" control loop.

[0068] Therefore, pre-validation reduces decision-making risk, and through digital twin simulation, it discovers scheme defects in advance, improves the first-pass rate of the scheduling scheme, and avoids resource waste caused by traditional "trial and error" scheduling. Visualization improves execution efficiency, and the graphical instructions on the operation terminal shorten the response time of the staff by 50% and reduce the berth allocation misoperation rate. Dynamic closed loop ensures system stability, real-time monitoring and deviation feedback mechanism with the twin, so that the scheduling correction time under emergency conditions is controlled within 5 minutes, the alignment degree of the actual ship track and the planned route is improved, and the standardization degree of the port operation process is improved. For example, after a certain port applies this scheme, the simulation evaluation time of the ship's risk avoidance scheme during a typhoon is compressed from 30 minutes to 5 minutes, the post-disaster resumption scheduling efficiency is improved, and no safety accidents caused by scheduling errors occur, fully reflecting the value of digital twin technology in risk prevention and efficiency optimization in passenger-rolling transportation intelligent scheduling.

[0069] As an optional embodiment, the execution effect of the shipping scheduling operation is monitored in real time in 104, including: generating a monitoring scheme for the shipping scheduling operation for the ship and port operation terminals, and synchronously issuing deployment information of the monitoring scheme to the ship and port operation terminals; in the deployment information, deploying real-time position tracking and energy consumption monitoring on the ship terminal, deploying infrared sensor real-time feedback berth occupation monitoring on the port terminal, and accessing the port terminal through the RFID cargo tracking system data to monitor the unloading efficiency in real time; when the actual execution effect deviates from the scheduling scheme by more than a threshold value, automatically triggering local re-optimization, and generating a correction action through the digital twin to the decision-making scheduling model.

[0070] In the optional embodiment, the specific principles and technical effects are as follows: by constructing a dynamic closed-loop mechanism of “monitoring-feedback-correction”, precise control of the whole process of shipping scheduling is realized. The core principle is as follows: first, a monitoring scheme is customized for the ship and the port operation terminal. Real-time position (update frequency 2 minutes / second) and energy consumption data (such as main engine power and fuel consumption rate) are collected at the ship end by using GPS and energy consumption sensors. A high-density data collection network is constructed at the port end by using infrared sensors (berth occupation state is updated every 2 seconds) and RFID cargo tracking systems (real-time recording of cargo loading and unloading time and quantity). All monitoring data are synchronized to the cloud digital twin through edge computing nodes in real time, forming a dynamic mapping of the physical system and the virtual model. When the deviation between the actual execution data (such as the ship's sailing track deviating from the planned route by more than 2 nautical miles, and the port's loading and unloading efficiency being lower than expected by 30%) and the scheduling scheme exceeds the preset threshold, the system automatically triggers local re-optimization logic: the digital twin updates the current global state based on real-time deviation data, and the decision-making scheduling model generates correction actions within seconds by using multi-agent reinforcement learning algorithm (such as PPO). For example, when a ship is late and causes a conflict in the occupation of subsequent berths, the model automatically adjusts the berthing time of adjacent ships and reallocates idle berths. The correction instructions are pushed to the corresponding terminal for execution after being simulated and verified by the digital twin.

[0071] Therefore, real-time defect repair, through high-density data collection of Internet of Things devices, the discovery delay of abnormal events is compressed from 30 minutes of traditional manual inspection to real-time response, and the average time consumption of local re-optimization is only 3 minutes, which is more than 90% faster than manual scheduling. The ship position tracking error is controlled within 5 meters, the energy consumption monitoring accuracy is ±2%, the port berth occupation state recognition accuracy is more than 99%, the state consistency of the virtual twin and the physical system is ensured, and the effectiveness of the correction action is improved to 92%. Through the feedback of historical deviation data to the model training, the prediction accuracy of the system for similar abnormalities is improved from 65% to 88%. For example, in the scene of repeated failure of loading and unloading equipment, the model can give an early warning of 30 minutes and automatically adjust the equipment scheduling strategy, reducing the delay loss by 70%. After applying the mechanism to a cross-sea passenger and roll transportation scene, the actual execution deviation rate of the ship scheduling plan is reduced from 18% to 5%, the average recovery time of port congestion is shortened from 2 hours to 20 minutes, and no safety incidents caused by scheduling errors occur throughout the year, fully verifying the significant value of the scheme in improving the reliability and operation efficiency of the system.

[0072] In the embodiments of the present application, the automatic prediction of passenger and roll transportation demand and the intelligent scheduling of ships are realized, the efficiency of ship scheduling is optimized, the efficiency of port operation quality monitoring is improved, and the port operation quality is improved, thereby ensuring the quality and reliability of port operation services.

[0073] In still another embodiment of the present application, a big data-based passenger and roll transportation demand prediction and ship intelligent scheduling system is also provided, which is described with reference to Figure 2 The big data-based passenger and roll transportation demand prediction and ship intelligent scheduling system comprises the following units: The acquisition unit is configured to obtain multi-source shipping data for a target area; the multi-source shipping data at least comprises historical ticket data, weather information, holiday traffic, and port operation data; The prediction unit is configured to input the multi-source shipping data into a spatiotemporal feature mining model to predict passenger and roll transportation demand information of the target area; wherein the model parameters of the spatiotemporal feature mining model are obtained based on the regional function structure, the regional carrying capacity, and the safety carrying capacity boundary setting in the target area; The optimization unit is configured to obtain real-time positions of ships in the target area, passenger carrying capacity, energy consumption data, and real-time operation states of ports, combine the predicted passenger and roll transportation demand information, and dynamically generate an optimal scheduling scheme by using a shipping scheduling model; wherein the shipping scheduling model is obtained by iteratively training the reinforcement learning algorithm through the interaction between the decision scheduling model and the agent corresponding to the passenger and roll transportation system; The scheduling unit is configured to convert the optimal scheduling scheme into visual information, push the visual information to operation terminals of ship and port staff to start corresponding shipping scheduling operations, and monitor the execution effect of the shipping scheduling operations in real time.

[0074] The functions of each unit in the above system embodiment can be referred to the specific description in the method embodiment, which will not be repeated here.

[0075] In still another embodiment of the present application, an electronic device is also provided, which comprises a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory complete mutual communication through the communication bus; The memory is used to store a computer program; The processor is used to execute the program stored on the memory to implement the big data-based passenger and roll transportation demand prediction and ship intelligent scheduling method described in the method embodiment.

[0076] The communication bus 1140 mentioned in the above electronic device can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc.

[0077] The embodiment of the application provides a big data-based cruise and roll transportation demand prediction and ship intelligent scheduling method for constructing a low-energy consumption operation device.

[0078] For ease of representation, Figure 3 Only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus. The communication interface 1120 is used for communication between the electronic device and other devices. The memory 1130 can include a random access memory (RAM) and can also include a non-volatile memory, such as at least one disk memory. Optionally, the memory can also be at least one storage device located away from the aforementioned processor. The aforementioned processor 1110 can be a general-purpose processor, including an artificial intelligence processor, a graphics processing unit (GPU), a machine learning unit (MLU), a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0079] Correspondingly, the embodiment of the application also provides a computer readable storage medium storing a computer program, and the computer program can implement each step that can be executed by the electronic device in the method embodiment when executed.

Claims

1. A big data-based passenger and roll transportation demand prediction and ship intelligent scheduling method, characterized in that, The method comprises: acquiring multi-source shipping data for a target area; the multi-source shipping data at least comprising: historical ticket data, weather information, holiday traffic, port operation data; inputting the multi-source shipping data into a spatio-temporal feature mining model to predict passenger-rolling transportation demand information in the target area; wherein model parameters of the spatio-temporal feature mining model are obtained based on regional function structure, regional carrying capacity and safety transport capacity boundary setting in the target area; acquiring real-time positions of ships, passenger-carrying capacity, energy consumption data and real-time operation states of ports within the target area, combining the predicted passenger-rolling transportation demand information, and using a shipping scheduling model to dynamically generate an optimal scheduling scheme; wherein the shipping scheduling model is obtained by iteratively training a decision scheduling model and an agent corresponding to a passenger-rolling transportation system using a reinforcement learning algorithm; converting the optimal scheduling scheme into visual information and pushing it to operation terminals of ship and port staff to start corresponding shipping scheduling operations and monitor the execution effect of the shipping scheduling operations in real time.

2. The big data-based passenger and roll transportation demand prediction and ship intelligent scheduling method according to claim 1, characterized in that, Before the inputting the multi-source shipping data into the spatio-temporal feature mining model to predict the passenger-rolling transportation demand information, the method further comprises: using a multi-modal graph spatio-temporal network (MSTGN) to integrate a multi-source shipping data training set to construct a spatial graph structure; and dividing a time sequence into time slices to construct a spatial node connection relationship in the time dimension through a time sliding window to capture periodic change rules and trend change rules in the multi-source shipping data training set to obtain a time graph structure; spatio-temporally jointly encoding the spatial graph structure and the time graph structure, using a hierarchical cross-attention mechanism to fuse the spatio-temporally jointly encoded results to obtain multi-scale spatio-temporal joint features; wherein the spatial scale comprises a port level, a route level and a ship level, and the time scale comprises a year, a month, a day and an hour; using the multi-scale spatio-temporal joint features to predict passenger transportation demand, freight transportation demand, ship scheduling demand and abnormal delay events in combination with a multi-task learning mechanism, sharing prediction results from a bottom layer to train a prediction process of the passenger-rolling transportation demand information to obtain the spatio-temporal feature mining model.

3. The big data based passenger and roll transportation demand prediction and ship intelligent scheduling method according to claim 2, characterized in that, The using the multi-modal graph spatio-temporal network (MSTGN) to integrate the multi-source shipping data training set to construct the spatial graph structure comprises: using ports, ships, passenger transportation areas, freight transportation areas and service areas in the target area as spatial nodes, using route connectivity, spatial function relationship and spatial function interaction relationship as connection edges to construct a spatial adjacency matrix; based on the spatial adjacency matrix, integrating port topological relationship, historical ticket time sequence relationship, weather grid, holiday flow change trend and historical port operation relationship in the multi-source shipping data training set, based on a dynamic graph learning mechanism to establish an adaptive adjacency matrix, dynamically adjusting node relationship in the adaptive adjacency matrix and contribution of different modal data in the multi-source shipping data training set through attention dynamic weight to obtain the spatial graph structure.

4. The big data based passenger and roll transportation demand prediction and ship intelligent scheduling method according to claim 2, characterized in that, The multi-scale space-time combined feature is used to predict passenger transport demand, freight transport demand, ship scheduling demand and abnormal delay events, share the prediction results from the bottom layer, and fuse the passenger and roll transport demand information. Discretize the marine fluid equation Navier-Stokes into difference form, and generate ship sailing resistance coefficient matrix coupled with meteorological data , the marine fluid equation Navier-Stokes is embedded into the reward function of the spatiotemporal feature mining model combined with the mechanism of fluid mechanics, so as to improve the robustness of the spatiotemporal feature mining model in severe sea conditions; Wherein, the ship sailing resistance coefficient matrix is used for calculating the square modulus of the Navier-Stokes equation residual, and the calculation process of the ship sailing resistance coefficient matrix is represented as: ; is a three-dimensional sea area flow velocity vector field, for simulating sailing resistance, represents a convection term describing the transport effect of the fluid due to its own speed, is a hydrostatic pressure field, related to the depth and density of the ocean current, represents a pressure gradient term, is a seawater viscosity coefficient, represents a viscosity term reflecting the influence degree of fluid viscosity on movement; The regional function structure is encoded as a discrete variable, the regional carrying capacity and the safe transport capacity boundary are taken as linear inequality constraints, and the port throughput limit is taken as a hard constraint and embedded into the loss function; the calculation process of the loss function is represented by the following formula: ； wherein, , , are all adaptively adjusted according to the violation degree of the corresponding constraint loss term to ensure that the prediction result conforms to the physical rules and safety boundaries; is a basic prediction loss for representing the error between the model prediction value and the true value, represents a physical constraint loss related to the ship navigation resistance, is a load overrun penalty, is an actual load containing passenger flow and cargo volume, is the maximum carrying capacity of the region, is a ship overrun penalty, is the actual number of ships in the target region, is a safe transport capacity boundary containing the upper limit of anchorage capacity.

5. The big data based passenger and roll transportation demand forecasting and ship intelligent scheduling method according to claim 1, characterized in that, Before the multi-source shipping data is input into the space-time feature mining model to predict the passenger and roll transport demand information of the target region, the following steps are further included: Identify new port regions with historical data volume less than a set sparsity threshold from the target region; An meta-learning algorithm is used to construct a meta-training dataset, which includes multi-source data of mature port regions and corresponding demand prediction results; Through the meta-learner, a gradient update strategy is used to learn the general feature representation across ports, so that the model parameters of the space-time feature mining model have the ability to quickly adapt to new port regions; For new port regions, the first month ticket data and port basic layout in the initial data are obtained, and the model parameters of the space-time feature mining model are generated by the meta-learner to realize rapid generalization of demand prediction for new port regions and reduce prediction errors in the cold start stage.

6. The big data based passenger and roll transportation demand forecasting and ship intelligent scheduling method according to claim 1, characterized in that, The steps of interacting with the agent corresponding to the passenger and roll transport system by the decision scheduling model, and iteratively training the shipping scheduling model by using the reinforcement learning algorithm, further include: An agent in the passenger and roll transport system is created for each ship and each port region in the target region; the ship agent state space includes real-time position, remaining passenger capacity, cargo carrying capacity, current energy consumption rate, expected arrival time, and route planning path; the port agent state space includes berth occupancy, loading and unloading equipment availability, cargo backlog, and current period demand prediction value; An action space is constructed, wherein the executable actions of the ship agent include adjusting the sailing speed, applying for a berth change, and requesting priority loading and unloading, and the executable actions of the port agent include assigning a berth to a specified ship, scheduling loading and unloading equipment, and issuing a port congestion warning; A positive reward function, a negative reward function, and a global reward function are set; the positive reward function is constructed based on the goals of improving ship utilization, reducing empty load rate, reducing port congestion index, and reducing energy consumption below expectation, the negative reward function is constructed based on the goals of ship waiting time exceeding a threshold, berth allocation conflict, and safety boundary, and the global reward function introduces a cooperative reward item based on an inter-agent communication mechanism, the cooperative reward item at least cooperates with adjacent port agents to reduce ship detours, and each reward is set to a score; An initialization model of the shipping scheduling model is constructed, wherein the initialization model includes a passenger transport carrying capacity model, a ship energy consumption model, and a port loading and unloading model; The multi-agent proximal policy optimization (PPO) is adopted to interact each agent in the passenger-rolling transportation system with an initialized model to obtain real-time state data matched with each agent; in combination with an action space, a positive reward function, a negative reward function, and a global reward function, an initialized strategy network composed of a passenger carrying capacity model, a ship energy consumption model, and a port loading and unloading model is used to predict the action state of each agent in the next period, and the action state is added to a shared experience pool; a reinforcement learning training mechanism is adopted to iterate the action state in the shared experience pool and the prediction process until a preset convergence condition is met to output a final shipping scheduling model; The clip strategy is used to update the strategy network to evaluate the degree of violation of the safety capacity boundary by the shipping scheduling model to correct abnormal logic in the shipping scheduling model.

7. The big data based passenger and roll transportation demand forecasting and ship intelligent scheduling method according to claim 6, characterized in that, The real-time position, passenger carrying capacity, energy consumption data, and port real-time operation state of the ship in the target area are obtained, and the predicted passenger-rolling transportation demand information is combined to dynamically generate an optimal scheduling scheme by using the shipping scheduling model, including: Real-time acquisition of the real-time position, passenger carrying capacity, energy consumption data, and port real-time operation state of the ship in the target area to construct the current global state in the target area; Based on the corresponding local real-time state in the current global state and the global collaborative state with other agents, each agent generates a corresponding action state through the strategy network; The action states of each agent are integrated to verify whether the safety capacity boundary and the port throughput limit in the target area are met to generate an optimal scheme including ship schedule adjustment, route optimization, and port resource allocation.

8. The big data based passenger and roll transportation demand forecasting and ship intelligent scheduling method according to claim 1, characterized in that, The optimal scheduling scheme is converted into visual information and pushed to the operation terminal of the ship and port staff to start the corresponding shipping scheduling operation, including: Based on the port three-dimensional model, route geographic information, and ship dynamics parameters of the target area, a digital twin matching the target area is established in the cloud; Real-time synchronization of the port operation data, ship sensor data, and weather data of the target area to the digital twin; the port operation data at least includes berth state and loading progress, the ship sensor data at least includes ship position, ship speed in and out of the port, detention time, ship energy consumption, ship supply condition, ship passenger capacity, and ship cargo capacity, and the weather data at least includes wind speed and wave height; After each optimization scheduling is started, the optimal scheduling scheme is mapped to the data twin to trigger virtual ship navigation simulation and port operation process, monitor the ship arrival time error, berth idle time change, and energy consumption prediction deviation in the simulation scheduling process, and evaluate the optimal scheduling scheme by using the monitoring results; if the monitoring results exceed the safety boundary or the port congestion is aggravated, the optimal scheduling scheme is regenerated until it passes the evaluation; For the optimal scheduling scheme that passes the evaluation, visual instructions are generated according to the mapping information of each scheduling node in the data twin and the format requirements of the ship and port operation terminal, and are sent to the corresponding operation terminal.

9. The big data based passenger and roll transportation demand forecasting and ship intelligent scheduling method according to claim 8, characterized in that, The execution effect of the shipping scheduling operation is monitored in real time, including: A monitoring scheme of shipping scheduling operation is generated for a ship and a port operation terminal, and deployment information of the monitoring scheme is synchronously issued to the ship and the port operation terminal; In the deployment information, real-time position tracking and energy consumption monitoring are deployed at the ship end, and real-time feedback berth occupation monitoring by an infrared sensor and data access to the port end through an RFID cargo tracking system are deployed at the port end, so as to monitor the loading and unloading efficiency in real time; When the actual execution effect deviates from the scheduling scheme by more than a threshold value, local re-optimization is automatically triggered, and a correction action is generated for the decision scheduling model through a digital twin. 10.A big data based demand forecasting and intelligent dispatching system for ro-ro shipping, characterized in that, The system comprises: The acquisition unit is configured to obtain multi-source shipping data for a target area; the multi-source shipping data at least includes historical ticket data, weather information, holiday traffic, and port operation data; The prediction unit is configured to input the multi-source shipping data into a spatiotemporal feature mining model to predict the cruise shipping demand information of the target area; wherein the model parameters of the spatiotemporal feature mining model are obtained based on the regional function structure, regional carrying capacity, and safety carrying capacity boundary setting in the target area; The optimization unit is configured to obtain the real-time position, passenger carrying capacity, energy consumption data of the ship in the target area, and the real-time operation state of the port, combine the predicted cruise shipping demand information, and dynamically generate an optimal scheduling scheme by using a shipping scheduling model; wherein the shipping scheduling model is obtained by interacting the decision scheduling model with the agent corresponding to the cruise shipping system, and is obtained by iterative training using a reinforcement learning algorithm; The scheduling unit is configured to convert the optimal scheduling scheme into visual information, and push the visual information to the operation terminal of the ship and the port staff to start the corresponding shipping scheduling operation, and monitor the execution effect of the shipping scheduling operation in real time.

Images