Intelligent berth scheduling method and system based on port multi-source data collaboration

Abstract

The embodiment of the application relates to the field of intelligent berth scheduling, and provides an intelligent berth scheduling method and system for port multi-source data cooperation. First, through a machine learning model, multi-source real-time data is converted into accurate prediction of future events, and a deviation vector between the prediction and a benchmark plan is actively calculated, thereby solving the core problems of passive response and lack of foresight in the prior art. Secondly, not all deviation vectors are processed, but the severity of each deviation vector is judged by quantitatively calculating the chain influence score of each deviation vector, and a minimum influence domain is dynamically defined based on a resource dependency graph. When the influence score exceeds a preset threshold, only the rapid re-optimization of the local influence domain is triggered. Through such accurate adjustment, the dilemma of response lag and plan instability caused by the blind global recalculation of the traditional technology is completely changed, and the real-time performance, robustness and low cost of the scheduling decision in a highly dynamic environment are ensured.

Description

Technical Field

[0001] This invention relates to the field of intelligent berth scheduling technology, and in particular to an intelligent berth scheduling method and system for multi-source data collaboration in ports. Background Technology

[0002] As a crucial hub in the global logistics chain, the operational efficiency of ports directly impacts the cost and timeliness of international trade. Berth scheduling, a core component of port operations, aims to scientifically and efficiently allocate berth resources and quay cranes to arriving vessels, minimizing vessel dwell time and maximizing berth utilization. However, ports are highly dynamic and complex systems fraught with uncertainty. Key parameters such as actual vessel arrival times and loading / unloading operation durations are frequently affected by various random disturbances, including weather changes, channel congestion, and temporary equipment malfunctions. These disturbances cause pre-defined static scheduling plans to frequently fail in practice, leading to a chain reaction of prolonged vessel anchorage waits, idle berth resources, and disruptions to subsequent cargo handling plans, severely hindering the improvement of port operational efficiency. Therefore, how to integrate multi-source real-time data from both inside and outside the port to construct an intelligent berth scheduling solution that proactively adapts to dynamic disturbances has become a pressing technical challenge in the field of smart port construction.

[0003] Currently, while some port scheduling systems have begun to integrate data sources such as Automatic Identification Systems (AIS) and terminal operating systems, they still have significant shortcomings in responding to dynamic disturbances. These systems often adopt a relatively passive response mode, that is, only after a disturbance event occurs and causes significant delays, can dispatchers intervene manually or trigger a global recalculation. This approach has two major problems: First, the response is delayed, lacking the ability to predict and quantify disturbances in advance, and failing to eliminate potential problems in their infancy; second, the adjustments are crude, as global rescheduling not only involves huge computational loads and excessive time consumption, leading to the system potentially facing new disturbances during the calculation period and falling into a state of continuous calculation and never converging, but also ignores the operational and coordination costs brought about by changing the scheduling plan itself, often sacrificing the stability and executability of the plan in pursuit of a theoretically optimal solution. Summary of the Invention

[0004] The present invention aims to solve at least one of the problems existing in the prior art, and to provide a method and system for intelligent berth scheduling based on multi-source data collaboration in ports.

[0005] One aspect of the present invention provides an intelligent berth scheduling method based on multi-source data collaboration in ports, comprising: The port's original AIS data, original meteorological data, original TOS data, and original quay crane data are synchronized digitally to obtain a unified port status. Based on the baseline scheduling plan, machine learning-based disturbance prediction and baseline deviation calculation are performed on the unified port status to obtain the deviation vector; Based on the deviation vector, the triggering of disturbance events and the evaluation of the impact domain are performed to obtain the triggered rescheduling events; Based on the unified port status and baseline scheduling plan, the triggered rescheduling events are subjected to local re-optimization based on the influence domain to obtain scheduling change proposals.

[0006] Optionally, based on the baseline scheduling plan, machine learning-based disturbance prediction and baseline deviation calculation are performed on the unified port status to obtain a deviation vector, including: Extract the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th vessel from the unified port status. The dynamic data of the i-th vessel in transit, the static data of the i-th vessel in transit, the target route and berth data of the i-th vessel in transit, and the environmental data along the predicted path are input into the pre-trained ETA prediction model to obtain the predicted ETA list. Extract cargo operation data of the i-th vessel en route from the baseline scheduling plan, and allocate quay crane data and quay crane performance data to the i-th vessel en route; Based on the cargo operation data of the i-th vessel in transit, the quay crane data and quay crane efficiency data allocated to the i-th vessel in transit, the berth operation time is predicted based on equipment efficiency and operational constraints, resulting in a list of predicted operation times. The predicted job duration list and the predicted ETA list are vectorized to obtain the prediction-planning deviation vector.

[0007] Optionally, the dynamic data of the i-th vessel en route, the static data of the i-th vessel en route, the target route and berth data of the i-th vessel en route, and the environmental data along the predicted path are input into the pre-trained ETA prediction model to obtain a list of predicted ETAs, including: Feature engineering is performed on the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th vessel in transit to obtain the feature vector of the i-th vessel in transit. Input the feature vector of the i-th ship en route into the pre-trained ETA prediction model to obtain the predicted time required for the i-th ship en route to complete the remaining voyage. Add the predicted time required for the i-th vessel to complete the remaining voyage to the current timestamp to obtain the predicted ETA for the i-th vessel.

[0008] Optionally, the feature vector of the i-th vessel en route includes the following: current ground speed, current ground heading, average speed over a preset time period, standard deviation of speed over a preset time period, instantaneous acceleration, turning rate, remaining distance, total planned distance, number of remaining waypoints, whether it is in the channel, distance to the pilot station, vessel type code, total length, beam, current draft, total tonnage, average wind speed of segment k, average wind direction of segment k, average relative wind angle of segment k, average ocean current speed of segment k, average ocean current direction of segment k, average relative current angle of segment k, average wave height of segment k, average visibility of segment k, vessel punctuality score, number of vessels at the current anchorage, and occupancy rate of the wharf where the target berth is located.

[0009] Optionally, based on the deviation vector, the triggering of disturbance events and the assessment of their impact domain are performed to obtain the triggered rescheduling events, including: Based on the dynamic threshold configuration library, the deviation vector is filtered for sensitive deviations based on multi-dimensional dynamic thresholds to obtain a list of sensitive deviations. The list of sensitive deviations is quantified and scored based on the cascading effects of port resource dependence graphs to obtain a list of scored events; Based on the rescheduling trigger score threshold, the triggering decision and influence domain breadth priority determination are performed on the scoring event list to obtain the triggered rescheduling events.

[0010] Optionally, the list of sensitive deviations is quantified and scored based on the cascading effects of the port resource dependency graph to obtain a list of scored events, including: Extract the first sensitive deviation from the list of sensitive deviations; In the port resource dependency graph, find the source task node corresponding to the first sensitive deviation, and find the set of all downstream task nodes directly reachable from the source task node. Calculate the time difference between the planned start time of each downstream task node in the set of downstream task nodes and the planned end time of the source task node to obtain a set of safety buffer times; The total impact score of the first sensitive deviation is calculated based on the set of safety buffer times.

[0011] Another aspect of the present invention provides an intelligent berth scheduling system for port multi-source data collaboration, comprising: The digital twin status synchronization module is used to synchronize the port's original AIS data, original meteorological data, original TOS data and original quay crane data with digital twins to obtain a unified port status. The deviation vector calculation module is used to perform machine learning-based disturbance prediction and baseline deviation calculation on the unified port status based on the baseline scheduling plan, and obtain the deviation vector. The rescheduling event triggering module is used to trigger disturbance events and evaluate the impact domain based on the deviation vector, so as to obtain the triggered rescheduling events. The local re-optimization module is used to perform local re-optimization based on the influence domain of triggered rescheduling events, based on the unified port status and baseline scheduling plan, to obtain scheduling change proposals.

[0012] Optionally, the deviation vector calculation module is specifically used for: Extract the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th vessel from the unified port status. The dynamic data of the i-th vessel in transit, the static data of the i-th vessel in transit, the target route and berth data of the i-th vessel in transit, and the environmental data along the predicted path are input into the pre-trained ETA prediction model to obtain the predicted ETA list. Extract cargo operation data of the i-th vessel en route from the baseline scheduling plan, and allocate quay crane data and quay crane efficiency data to the i-th vessel en route. Based on the cargo operation data of the i-th vessel in transit, the quay crane data and quay crane efficiency data are allocated to the i-th vessel in transit. Based on equipment efficiency and operational constraints, the berth operation time is predicted, and a list of predicted operation times is obtained. The predicted job duration list and the predicted ETA list are vectorized to obtain the prediction-planning deviation vector.

[0013] Compared with existing technologies, the present invention provides a port multi-source data collaborative intelligent berth scheduling method and system. First, it uses a machine learning model to transform multi-source real-time data into accurate predictions of future events (such as ship arrivals and operation durations), and actively calculates the deviation vector between the prediction and the baseline plan, thereby solving the core problems of passive response and lack of foresight in existing technologies. Second, it does not process all deviation vectors, but judges their severity by quantifying the chain effect score of each deviation vector, and dynamically defines the minimized influence domain based on the resource dependency graph. Only when the influence score exceeds a preset threshold is a rapid re-optimization triggered only for that local influence domain. Through this precise adjustment method, it completely changes the dilemma of response lag and unstable planning caused by blind global recalculation in traditional technologies, ensuring the real-time performance, robustness and low cost of scheduling decisions in highly dynamic environments. Attached Figure Description

[0014] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0015] Figure 1 A flowchart of a port multi-source data collaborative intelligent berth scheduling method according to an embodiment of the present invention; Figure 2 This is a data flow diagram of an intelligent berth scheduling method for port multi-source data collaboration according to an embodiment of the present invention; Figure 3 The flowchart illustrates the benchmark-based scheduling plan of the intelligent berth scheduling method for port multi-source data collaboration according to an embodiment of the present invention, which performs machine learning-based disturbance prediction and benchmark deviation calculation on a unified port status to obtain a deviation vector. Figure 4 The flowchart illustrates the process of inputting the dynamic data of the i-th vessel en route, the static data of the i-th vessel en route, the target route and berth data of the i-th vessel en route, and the environmental data along the predicted path into a pre-trained ETA prediction model to obtain a predicted ETA list, according to an embodiment of the present invention. Figure 5 The flowchart illustrates the intelligent berth scheduling method for port multi-source data collaboration according to an embodiment of the present invention, which uses deviation vectors to trigger disturbance events and evaluate the impact domain, resulting in a flowchart of the triggered rescheduling event. Figure 6 A flowchart is provided to illustrate the intelligent berth scheduling method for port multi-source data collaboration according to an embodiment of the present invention, which quantifies and scores the chain effects of a sensitive deviation list based on a port resource dependency graph, resulting in a scoring event list. Figure 7 This is a block diagram of a port multi-source data collaborative intelligent berth scheduling system according to an embodiment of the present invention. Detailed Implementation

[0016] As indicated in this invention and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0017] While this invention makes various references to certain modules in systems according to embodiments of the invention, any number of different modules can be used and run on user terminals and / or servers. The modules are merely illustrative, and different aspects of the systems and methods may use different modules.

[0018] This invention uses flowcharts to illustrate the operations performed by the system according to embodiments of the invention. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously, as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0019] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0020] To address the technical challenges of frequent failures in port static scheduling plans due to their inability to adapt to real-world disturbances such as changes in ship arrival times and sudden weather events, as well as the slow response and excessively wide adjustment range of existing mechanisms, this invention proposes an intelligent berth scheduling method and system based on multi-source data collaboration. First, it deeply integrates various isolated real-time data streams from the port, such as Automatic Identification System (AIS) data, Terminal Operation System (TOS) data, and meteorological data, to form a unified and synchronized port status view. Then, based on this port status view, it proactively utilizes machine learning models to transform the real-time status into predictions of key future parameters, such as the estimated time of arrival of ships. Accurate predictions of arrival times (ETAs) and actual berth operation times are made, and these predictions are compared with baseline plans to quantify potential scheduling conflicts and extract a warning deviation vector. Then, intelligent decision-making first uses dynamic thresholds to filter out tolerable minor deviations, retaining only potentially threatening sensitive deviations. Next, an innovative scoring model, combined with a port resource dependency graph, calculates the severity of the cascading impact of each sensitive deviation, intelligently determining whether intervention is necessary. Once intervention is decided, a breadth-first traversal is performed along the dependency graph to precisely delineate the smallest set of tasks directly or indirectly affected by the disturbance, i.e., the impact domain. Finally, a fast re-optimization algorithm is initiated only for this isolated local area to generate a scheduling change proposal that resolves the current conflict while maximizing the stability of the original plan. This approach ensures that the response to dynamic disturbances is both predictive and accurate, fundamentally improving the robustness and operational efficiency of port scheduling.

[0021] One embodiment of the present invention proposes an intelligent berth scheduling method based on multi-source data collaboration in ports. Figure 1 This is a flowchart of a port multi-source data collaborative intelligent berth scheduling method according to an embodiment of the present invention. Figure 2 This is a data flow diagram illustrating the intelligent berth scheduling method for multi-source data collaboration in ports according to an embodiment of the present invention. (In conjunction with...) Figure 1 and Figure 2 According to an embodiment of the present invention, a port multi-source data collaborative intelligent berth scheduling method includes the following steps: S100, performing digital twin state synchronization on the port's original AIS data, original meteorological data, original TOS data, and original quay crane data to obtain a unified port state; S200, based on a baseline scheduling plan, performing machine learning-based disturbance prediction and baseline deviation calculation on the unified port state to obtain a deviation vector; S300, based on the deviation vector, performing disturbance event triggering and impact domain assessment to obtain triggered rescheduling events; S400, based on the unified port state and the baseline scheduling plan, performing local re-optimization based on the impact domain on the triggered rescheduling events to obtain scheduling change proposals.

[0022] Specifically, in step S100, the original AIS data, original meteorological data, original TOS data, and original quay crane data of the port are synchronized using a digital twin to obtain a unified port state. It should be understood that the original AIS data, original meteorological data, original TOS data, and original quay crane data of the port originate from different business systems, and their data formats, update frequencies, timestamp standards, and semantic connotations all differ significantly, resulting in a lack of a unified and reliable data foundation for any subsequent intelligent analysis and decision-making. Therefore, in this embodiment, the original AIS data, original meteorological data, original TOS data, and original quay crane data of the port are further synchronized using a digital twin to construct a standardized, high-fidelity virtual port model that is consistent with the physical port in terms of key entities, state attributes, and spatiotemporal dimensions. This provides an unambiguous, globally consistent real-time state snapshot for all subsequent calculation steps such as disturbance prediction, impact domain assessment, and local re-optimization, thereby ensuring the accuracy and effectiveness of decision-making.

[0023] In practical implementation, step S100 can first subscribe to or poll the port's raw data stream in real time, including raw AIS data, raw meteorological data, raw TOS data, and raw quay crane data. Next, data preprocessing and standardization steps are performed. This step parses, cleans, and aligns the received heterogeneous data. For example, NMEA-formatted AIS messages are parsed into structured ship dynamic information; speed and heading data with jumps or drifts are smoothed using a Kalman filter; and timestamps from all data sources are uniformly converted to Coordinated Universal Time (UTC), unifying units to international standard units. Subsequently, the core state mapping and fusion step is performed. In this step, the preprocessed data is associated with virtual entities in the digital twin model. For example, based on the ship's unique identification code (IMO or MMSI), the parsed AIS dynamic position information is updated to the position attributes of the corresponding ship object in the digital twin model. Simultaneously, information such as the ship's planned berthing berth and planned operation time obtained from the TOS data is also bound to this ship object, achieving semantic fusion of multi-source data on the same entity. Finally, the unified port status aggregation and publication step is performed. In this step, all updated virtual entity objects, including ship, berth, quay crane and environmental field data, are aggregated into a complete port status data snapshot with a unified timestamp, which is the unified port status.

[0024] Specifically, in step S200, based on the baseline scheduling plan, machine learning-based disturbance prediction and baseline deviation calculation are performed on the unified port status to obtain a deviation vector. It should be understood that the baseline scheduling plan is essentially a static timetable based on idealized assumptions. It cannot inherently reflect or adapt to the real-world uncertainties caused by complex dynamic factors such as weather, sea conditions, ship performance, and equipment status, leading to inevitable deviations between the plan and actual execution. If these deviations are not predicted and quantified, they will evolve into scheduling conflicts and resource waste. Therefore, in this embodiment, based on the baseline scheduling plan, machine learning-based disturbance prediction and baseline deviation calculation are further performed on the unified port status. This transforms high-fidelity real-time status data into a more accurate prediction of future key events (such as actual ship arrival time and operation duration) through a prediction model capable of understanding complex nonlinear relationships. By comparing this prediction result with the baseline plan item by item, potential scheduling conflicts are made explicit and structured. In this way, the core capability of the scheduling system can be upgraded from a passive, delayed response to a proactive, forward-looking early warning, providing a precise and quantifiable data input, namely the deviation vector, for subsequent disturbance impact assessment and decision-making steps. This enables the entire scheduling system to proactively identify and measure potential risks before the disturbance event actually occurs.

[0025] Figure 3This document describes a flowchart illustrating a benchmark-based scheduling plan for a port multi-source data collaborative intelligent berth scheduling method according to an embodiment of the present invention. It describes a process of predicting disturbances and calculating benchmark deviations based on machine learning for a unified port status, resulting in a deviation vector. Figure 3 As shown, according to an embodiment of the present invention, the intelligent berth scheduling method for port multi-source data collaboration includes the following steps in step S200: S210, extracting dynamic data of the i-th vessel en route, static data of the i-th vessel en route, target route and berth data of the i-th vessel en route, and environmental data along the predicted path from a unified port status; S220, inputting the dynamic data of the i-th vessel en route, static data of the i-th vessel en route, target route and berth data of the i-th vessel en route, and environmental data along the predicted path into a pre-trained ETA prediction model. S230: Extract cargo operation data of the i-th vessel en route from the baseline scheduling plan, and allocate quay crane data and quay crane efficiency data to the i-th vessel en route; S240: Based on the cargo operation data of the i-th vessel en route, the quay crane data and quay crane efficiency data allocated to the i-th vessel en route, perform berth operation time prediction based on equipment efficiency and operational constraints, and obtain a predicted operation time list; S250: Vectorize the prediction-plan deviation of the predicted operation time list and the predicted ETA list to obtain the deviation vector.

[0026] Specifically, in step S210, the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th en route vessel are extracted from the unified port status. It should be understood that the accuracy of machine learning ETA prediction models highly depends on the completeness and relevance of the input features. Single vessel dynamic data is insufficient to support accurate time prediction for complex navigation processes; it is necessary to comprehensively consider multi-dimensional factors affecting navigation time, such as the vessel's own attributes, navigation objectives, and the external environment. Therefore, in this embodiment, the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th en route vessel are further extracted from the unified port status. This constructs a multi-dimensional, context-complete input dataset for each en route vessel requiring prediction, containing its current motion state, inherent physical characteristics, navigation task constraints, and the external environmental conditions it will face in the future.

[0027] It is worth mentioning that the feature vector of the i-th vessel en route includes its current ground speed, current ground heading, average speed over a preset time period (e.g., 30 minutes), standard deviation of speed over a preset time period (e.g., 30 minutes), instantaneous acceleration, turning rate, remaining distance, total planned distance, number of remaining waypoints, whether it is in a waterway, distance to the pilot station, vessel type code, overall length, beam, current draft, gross tonnage, average wind speed of segment k, average wind direction of segment k, average relative wind angle of segment k, average ocean current speed of segment k, average ocean current direction of segment k, average relative current angle of segment k, average wave height of segment k, average visibility of segment k, vessel punctuality score, number of vessels at current anchorage, and occupancy rate of the berth at the target berth. This ensures that the subsequent ETA prediction model obtains all the information necessary for high-precision inference, thereby significantly improving the reliability and accuracy of the prediction results and laying a solid data foundation for subsequent deviation calculations.

[0028] In practical implementation, the extraction process in step S210 can first use the unique identifier of a specific vessel en route, such as the vessel's MMSI code, as a query index to retrieve the structured data object of the unified port status. Then, the data extraction operation is performed step by step: First, extract the vessel's dynamic data, obtaining real-time ground speed, ground heading, current latitude and longitude coordinates, and corresponding timestamps synchronized from AIS data from the twin object associated with the vessel identifier; second, extract the vessel's static data, obtaining the vessel's inherent physical attributes that remain unchanged during this voyage from the same twin object, such as the vessel's overall length, beam, gross tonnage, vessel type code, and current draft; third, extract the target route and berth data, based on the vessel's TOS... The fourth step involves extracting environmental data along the predicted path. Based on the planned waypoint sequence generated in the unified port status, the precise geographical coordinates of the berth are retrieved from the unified port status. This is combined with the preset standard route to the port to generate a planned waypoint sequence from the ship's current position to the target berth. The fifth step involves extracting environmental data along the predicted path. Based on the planned waypoint sequence generated in the previous step, a spatiotemporal query is performed in the environmental data layer of the unified port status to obtain meteorological and hydrological forecast grid data covering the geographical path for a period of time in the future. This includes information such as segmented or time-divisional wind speed, wind direction, ocean current speed, ocean current direction, and wave height.

[0029] Specifically, in step S220, the dynamic data of the i-th vessel en route, the static data of the i-th vessel en route, the target route and berth data of the i-th vessel en route, and the environmental data along the predicted path are input into the pre-trained ETA prediction model to obtain a predicted ETA list. It should be understood that the actual arrival time of a vessel is comprehensively affected by multiple nonlinear factors, including its own dynamic and static attributes, the characteristics of the planned route, and the complex and variable meteorological and hydrological environment along the route. No single data source or simple linear extrapolation method can provide sufficiently accurate predictions, resulting in a lack of reliable time benchmarks for scheduling plans. Therefore, in this embodiment, the dynamic data of the i-th en route vessel, the static data of the i-th en route vessel, the target route and berth data of the i-th en route vessel, and the environmental data along the predicted path are further input into the pre-trained ETA prediction model to obtain a predicted ETA list. This allows for a complete computational process including feature engineering, model inference, and time conversion, systematically transforming this raw, heterogeneous, multi-dimensional information into a standardized feature vector that can be understood and processed by the pre-trained ETA prediction model. Utilizing the complex navigation patterns learned from massive historical data and embedded in this ETA prediction model, the model ultimately calculates a high-precision predicted arrival time for each vessel under the current comprehensive conditions. This generates a dynamically updated predicted ETA for each en route vessel in the scheduling system, far more reliable than ship reports or traditional estimates. This provides a high-quality, reliable data foundation for subsequent accurate plan-prediction deviation calculations, a fundamental prerequisite for proactive disturbance early warning and management.

[0030] Figure 4 The flowchart illustrates the process of inputting dynamic data of the i-th vessel en route, static data of the i-th vessel en route, target route and berth data of the i-th vessel en route, and environmental data along the predicted path into a pre-trained ETA prediction model to obtain a predicted ETA list, according to an embodiment of the present invention. Figure 4 As shown, according to an embodiment of the present invention, the intelligent berth scheduling method for port multi-source data collaboration includes the following steps in step S220: S221, performing feature engineering on the dynamic data of the i-th vessel en route, the static data of the i-th vessel en route, the target route and berth data of the i-th vessel en route, and the environmental data along the predicted path to obtain the feature vector of the i-th vessel en route; S222, inputting the feature vector of the i-th vessel en route into a pre-trained ETA prediction model to obtain the predicted time required for the i-th vessel to complete the remaining voyage; S223, adding the predicted time required for the i-th vessel to complete the remaining voyage to the current timestamp to obtain the predicted ETA of the i-th vessel en route.

[0031] Specifically, in step S221, feature engineering is performed on the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th en route vessel to obtain the feature vector of the i-th en route vessel. It should be understood that the original multidimensional data extracted from a unified port status are of diverse types and dimensions, and contain a large amount of unstructured information and implicit correlations that machine learning models cannot directly utilize. For example, the model cannot directly understand the concepts of headwind or tailwind from absolute wind direction and heading data. Therefore, in this embodiment, feature engineering is further performed on the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th en route vessel to obtain the feature vector of the i-th en route vessel. This transforms these original, heterogeneous data into a standardized, numerical, high-dimensional feature set that can be directly processed by machine learning models, and in this process, calculates and derives deeper correlation features that better reveal navigation patterns. This provides the ETA prediction model with a higher information density and stronger predictive power, thereby fundamentally improving the accuracy and robustness of the model's predictions.

[0032] In practical implementation, the feature engineering process in step S221 can first execute the derived feature calculation step. In this step, new features with stronger business implications are calculated based on the original data. For example, the remaining voyage is calculated using the spherical distance formula based on the ship's current position coordinates and target berth coordinates; the average relative wind angle and average relative current angle are calculated based on the ship's planned course for each segment of the predicted path and the average wind direction and ocean current direction of that segment, respectively, to directly quantify environmental resistance or boosting effects; and statistical features such as average speed, speed standard deviation, and instantaneous acceleration are calculated based on the speed sequence over a past period. Next, the data encoding and transformation step is executed. In this step, categorical data is numerically processed. For example, the non-numerical feature of ship type is converted into a multi-dimensional binary vector using one-hot encoding. Subsequently, a numerical normalization step is performed. In this step, to eliminate the impact of differences in dimensions between different features on model training, all numerical features, such as total length, total tonnage, and remaining range, are scaled to a preset range, such as 0 to 1, using max-min normalization or Z-score normalization. Finally, a feature vector construction step is performed. In this step, all the calculated, encoded, and normalized numerical features are concatenated in a predefined order to form a single, flattened one-dimensional array, namely the feature vector of the i-th in-transit vessel. This vector can be directly input into the subsequent ETA prediction model.

[0033] Specifically, in step S222, the feature vector of the i-th en route vessel is input into a pre-trained ETA prediction model to obtain the predicted time required for the i-th en route vessel to complete the remaining voyage. It should be understood that the feature vector of the i-th en route vessel, obtained through feature engineering, is a high-dimensional numerical array. There are complex and non-linear coupling relationships between its various features and the final voyage time, relationships that cannot be effectively fitted by simple mathematical formulas or linear models. Therefore, in this embodiment, the feature vector of the i-th en route vessel is further input into the pre-trained ETA prediction model to obtain the predicted time required for the i-th en route vessel to complete the remaining voyage. This utilizes a complex function mapper that has learned and solidified these complex patterns from massive amounts of historical voyage data to perform an efficient, non-linear regression calculation on the current vessel's feature vector. In this way, static, multi-dimensional feature information can be transformed into a single, physically meaningful, and dynamically predicted time value that comprehensively considers all relevant factors, thus providing the most crucial core input for calculating the final estimated arrival time.

[0034] In practical implementation, the prediction process in step S222 can first perform the loading and instantiation of a pre-trained model. In this step, a pre-trained and validated machine learning model file, such as a gradient boosting decision tree model or a long short-term memory network model, is loaded from a specified model repository as the ETA prediction model, and then instantiated as an executable prediction object. Next, the input data formatting step is performed. In this step, the feature vector of the i-th in-transit ship, which is a one-dimensional array generated in the previous process, is converted into the standard input tensor format required by the ETA prediction model interface, such as a two-dimensional matrix of shape (1, N), where N is the dimension of the feature vector of the i-th in-transit ship. Subsequently, the model inference calculation step is performed. In this step, the prediction method of the instantiated prediction object is called, and the formatted input tensor is passed in as a parameter. The ETA prediction model then performs a forward propagation calculation, mapping the input tensor to an output value through its internal complex network layers or decision tree structure. Finally, the output result parsing step is executed. In this step, the prediction result is extracted from the output tensor returned by the ETA prediction model. This prediction result is a scalar value representing the time required for the ship to complete the remaining voyage, for example, in hours. This value is used as the final output of this step, which is the predicted time required for the i-th ship in transit to complete the remaining voyage.

[0035] Specifically, in step S223, the predicted time required for the i-th en route vessel to complete its remaining voyage is added to the current timestamp to obtain the predicted ETA for the i-th en route vessel. It should be understood that the predicted time output by the ETA prediction model is a relative time length, such as 5.5 hours, and does not possess a time coordinate itself, making it impossible to directly compare with the absolute time defined in the baseline scheduling plan, such as 3 PM. Therefore, in this embodiment, the predicted time required for the i-th en route vessel to complete its remaining voyage is further added to the current timestamp to obtain the predicted ETA for the i-th en route vessel. This anchors this relative, fluctuating time prediction to the current real time base, thereby converting it into an absolute time point with a specific date and time. In this way, a specific predicted arrival time that can be directly mathematically compared with the planned arrival time in the baseline scheduling plan on the same time dimension can be generated, thus providing final, consistent data for the subsequent accurate calculation of deviation values.

[0036] In practical implementation, the calculation process in step S223 can first execute the input data acquisition step. In this step, the predicted time required for the i-th en route vessel to complete its remaining voyage is obtained from the previous process. This predicted time is a floating-point number in hours or minutes. The current absolute timestamp is also obtained from the system clock or the global timestamp of the unified port status. Next, the time unit conversion and standardization step is executed. In this step, the predicted duration in floating-point format is converted into a standard time interval object. For example, 5.5 hours is converted into a time interval structure containing 5 hours, 30 minutes, and 0 seconds. Subsequently, the arithmetic addition operation of the timestamp is executed. In this step, a standard time library function is called to add the current timestamp object to the time interval object generated in the previous step. This addition operation accurately handles carry logic across minutes, hours, days, months, and years. Finally, the formatted output step of the result is executed. In this step, the future timestamp object obtained after the addition operation is formatted into a string conforming to a predefined standard or a standard date and time data type, and this is used as the predicted ETA output for the i-th en route vessel.

[0037] Specifically, in step S230, cargo operation data of the i-th vessel en route is extracted from the baseline scheduling plan, and quay crane data and quay crane efficiency data are allocated to the i-th vessel en route. It should be understood that the operation time of a vessel at the berth is not a fixed value, but directly depends on the total workload and the overall working efficiency of the allocated loading and unloading equipment in its current state. Without accurately obtaining these decisive parameters, it is impossible to effectively predict the operation time. Therefore, in this embodiment, cargo operation data of the i-th vessel en route is further extracted from the baseline scheduling plan, and quay crane data and quay crane efficiency data are allocated to the i-th vessel en route. This accurately gathers the three core calculation elements—the workload, available resource allocation, and real-time resource performance—for each vessel scheduled to berth. This provides a complete, unambiguous input dataset closely related to a specific vessel and a specific time point for the subsequent berth operation time prediction step based on equipment efficiency and operational constraints, thereby ensuring that the calculation basis for time prediction is accurate and complete.

[0038] In practical implementation, the extraction process in step S230 can begin by using the unique identifier of a specific vessel as the primary key to query the baseline scheduling plan data structure to locate the vessel's scheduling record. Next, the extraction of sub-items is performed: First, cargo operation data is extracted by reading the pre-loaded total operation volume field from the vessel's scheduling record, such as the total number of containers loaded and unloaded; second, allocated quay crane data is extracted by reading a list containing the unique identifiers of all quay cranes allocated for the vessel's operations from the same scheduling record; third, quay crane performance data is extracted by using the quay crane identifier list obtained in the previous step to query each one in the unified port status, obtaining the standard rated efficiency (a preset static value) and real-time performance correction factor (a dynamic value reflecting current equipment status, operator proficiency, etc.) from the twin object corresponding to each quay crane. Finally, the extracted total operation volume, quay crane identifier list, and performance data corresponding to each quay crane are aggregated into a structured dataset, which serves as the final output of step S230.

[0039] Specifically, in step S240, based on the cargo operation data of the i-th vessel en route, the quay crane data and quay crane efficiency data allocated to the i-th vessel are used to predict the berth operation time based on equipment efficiency and operational constraints, resulting in a predicted operation time list. It should be understood that the cargo operation data, allocated quay crane data, and efficiency data of each quay crane obtained from the previous process are discrete and multi-dimensional. They do not directly reflect the overall efficiency of multiple pieces of equipment working together or the total time required to complete all operations. This time is another key dimension for judging whether the scheduling plan will deviate. Therefore, in this embodiment, based on the cargo operation data of the i-th vessel en route, the quay crane data and quay crane efficiency data allocated to the i-th vessel are used to predict the berth operation time based on equipment efficiency and operational constraints, resulting in a predicted operation time list. This is achieved by first summing the individual efficiencies of all quay cranes allocated to the vessel after real-time efficiency factor correction, to calculate an expected total efficiency reflecting the current comprehensive loading and unloading capacity. Then, the vessel's total workload is divided by this expected total efficiency. In this way, the static operation plan and dynamic equipment performance can be transformed into a single predicted operation time value, corresponding to the vessel's ETA prediction, and directly used for subsequent deviation calculations, through a deterministic physical model. This completes a comprehensive prediction of the two core dimensions of future scheduling execution.

[0040] Specifically, in this embodiment, step S240 includes: multiplying the standard rated efficiency of each quay crane assigned to the i-th vessel en route by its corresponding real-time efficiency correction factor to obtain the real-time expected efficiency of each quay crane; then summing the real-time expected efficiency of all assigned quay cranes to obtain the expected total efficiency of the i-th vessel en route; subsequently, dividing the total workload of the i-th vessel en route by the expected total efficiency of the i-th vessel en route to obtain the predicted operation time of the i-th vessel en route. The real-time efficiency correction factor is a dynamic value reflecting factors such as the current equipment status and operator proficiency, and its specific value can be determined according to the actual situation. The standard rated efficiency is a preset static value.

[0041] In practical implementation, the berth operation duration prediction process can first initialize an empty predicted operation duration list for the i-th vessel en route in the baseline scheduling plan. Next, the expected total efficiency calculation step is executed. In this step, the list of quay cranes assigned to the i-th vessel en route is traversed. For the j-th quay crane in this list, its standard rated efficiency and real-time efficiency correction factor are obtained. The obtained standard rated efficiency and the corresponding real-time efficiency correction factor are multiplied to calculate the real-time expected efficiency of the j-th quay crane. Subsequently, the real-time expected efficiencies of all assigned quay cranes are summed to obtain the expected total efficiency of the i-th vessel en route. Then, the predicted operation duration calculation step is executed. In this step, the total workload of the i-th vessel en route is obtained and divided by the corresponding expected total efficiency to obtain a predicted operation duration in hourly floating-point form. Finally, the result aggregation step is performed. In this step, the calculated predicted operation duration is associated with the unique identifier of the i-th vessel in transit, and this association result is added as an element to the predicted operation duration list. After repeating the above steps for all vessels that need to be predicted, the predicted operation duration list is completed.

[0042] Specifically, in step S250, the predicted job duration list and the predicted ETA list are vectorized to obtain a deviation vector. It should be understood that the predicted ETA list and the predicted job duration list are merely independent sets of raw predicted values ​​about future states. The differences between them and the planned values ​​used as the scheduling benchmark have not yet been revealed, quantified, or structured, and therefore cannot be directly used for subsequent risk assessment and decision-making. Therefore, in this embodiment, the predicted job duration list and the predicted ETA list are further vectorized to obtain the deviation vector. This systematically pairs and mathematically compares the predicted results of each key event with its predetermined target in the benchmark scheduling plan, thereby transforming all potential, discrete time deviations into a unified, machine-readable, structured data set containing complete information such as the deviation subject, type, and magnitude. This generates a comprehensive, quantitative risk snapshot of all potential future scheduling conflicts at the port—the deviation vector—providing direct and accurate input for subsequent sensitive deviation screening and cascading impact assessment. It serves as a crucial bridge connecting the prediction and decision-making stages.

[0043] In practical implementation, the prediction-plan deviation vectorization process in step S250 can first initialize an empty deviation vector data structure. Next, iterate through each planned event in the baseline scheduling plan, such as the berthing operation of the i-th en route vessel. During the iteration, a step-by-step deviation calculation and encapsulation step is performed. For the currently traversed i-th en route vessel, its planned arrival time and planned operation duration are first extracted from the baseline scheduling plan. Subsequently, using the unique identifier of the en route vessel, the corresponding predicted arrival time and predicted operation duration are retrieved from the predicted ETA list and predicted operation duration list, respectively. Deviation calculation is then performed by subtracting the corresponding predicted value from the planned value (i.e., subtracting the predicted arrival time from the corresponding planned arrival time, and subtracting the predicted operation duration from the corresponding planned operation duration) to obtain the ETA deviation value and operation duration deviation value. Each non-zero deviation value, along with its associated vessel identifier, parameter type, planned value, and predicted value, is encapsulated into a structured deviation record. Finally, a vector construction step is performed, in which the deviation record generated in the previous step is added to the initialized deviation vector. After traversing all planned events, the deviation vector is constructed, containing all quantitative differences between forecasts and plans.

[0044] Specifically, in step S300, disturbance event triggering and impact domain assessment are performed based on the deviation vector to obtain the triggered rescheduling event. It should be understood that the deviation vector is merely a raw data set containing all potential scheduling deviations; it does not distinguish the urgency of different deviations or the severity of their cascading effects. Responding to all deviations would lead to frequent unnecessary calculations by the scheduling system, thereby compromising the stability of the plan. Therefore, in this embodiment, disturbance event triggering and impact domain assessment are further performed based on the deviation vector to obtain the triggered rescheduling event. This allows for the extraction of the raw deviation data into high-priority rescheduling tasks with clear action instructions through a multi-layered intelligent decision funnel that includes sensitive deviation screening, cascading effect quantification and scoring, and triggering and impact domain determination. This ensures that the system only intervenes when the impact of the disturbance truly exceeds the absorbable range, and the scope of intervention is precisely limited to the smallest affected task set. This significantly improves the efficiency and accuracy of decision-making while ensuring the robustness of the scheduling plan, avoiding the response delays and resource waste caused by blindly performing global recalculation in traditional methods.

[0045] Figure 5 This document describes a flowchart illustrating a port multi-source data collaborative intelligent berth scheduling method based on deviation vectors for triggering disturbance events and evaluating their impact domain, resulting in a triggered rescheduling event. Figure 5As shown, step S300 includes the following steps: S310, based on the dynamic threshold configuration library, performing sensitive deviation screening on the deviation vector based on multi-dimensional dynamic thresholds to obtain a sensitive deviation list; S320, performing chain impact quantification scoring on the sensitive deviation list based on the port resource dependency graph to obtain a scoring event list; S330, based on the rescheduling trigger score threshold, performing trigger decision and influence domain breadth priority determination on the scoring event list based on score sorting to obtain the triggered rescheduling events.

[0046] Specifically, in step S310, based on the dynamic threshold configuration library, the deviation vector is screened for sensitive deviations using multi-dimensional dynamic thresholds to obtain a sensitive deviation list. It should be understood that the deviation vector contains a large number of minor fluctuations that can be absorbed by the system's internal buffers and do not require immediate intervention. Furthermore, the tolerable degree of deviation varies for different types or objects of different importance. Using a single, fixed threshold for judgment would lead to rigid and distorted decision-making. Therefore, in this embodiment, the deviation vector is further screened for sensitive deviations using multi-dimensional dynamic thresholds based on the dynamic threshold configuration library to obtain a sensitive deviation list. This establishes a filtering mechanism that can adaptively judge based on the specific context of the deviation, such as the deviation type and the importance of the affected object, thereby accurately identifying truly potentially risky deviations from the raw data stream. This significantly reduces the computational load of subsequent analysis, avoids overreacting to non-critical disturbances, and generates a more focused and high-value sensitive deviation list, laying the foundation for subsequent accurate quantitative assessment of cascading effects.

[0047] In practical implementation, the filtering process in step S310 can begin by initializing an empty sensitive deviation list. Next, each deviation record in the deviation vector is traversed. For the currently traversed deviation record, its parameter type (e.g., ETA or operation duration) and the contextual attributes of its associated object (e.g., vessel type or target berth's busy level) are extracted. Then, these extracted types and attributes are used as composite query keys to perform a query match in the dynamic threshold configuration library to retrieve dynamic thresholds specifically for this particular scenario. A comparison and judgment step is then executed, where the absolute value of the deviation record is compared with the matched dynamic threshold. If the absolute value of the deviation record is greater than the corresponding dynamic threshold, the deviation record is added as a sensitive deviation to the sensitive deviation list. After traversing all deviation records, the sensitive deviation list is completed, containing only those deviation events that, based on their specific context, are determined to require further attention.

[0048] Specifically, in step S320, the sensitive deviation list is quantitatively scored based on the cascading impact of the port resource dependency graph to obtain a scored event list. It should be understood that the sensitive deviation list only reflects the isolated magnitude of each deviation event and cannot reveal the true destructive power of the deviation on the entire scheduling plan. This is because a small deviation occurring at a critical bottleneck node can trigger a cascading effect far exceeding that of a larger deviation occurring during a less busy period. Therefore, in this embodiment, the sensitive deviation list is further quantitatively scored based on the cascading impact of the port resource dependency graph to obtain a scored event list. This establishes a quantitative assessment model that can penetrate the surface of deviations and understand their propagation potential. This model analyzes the sequential constraints between tasks in the port resource dependency graph and weights the magnitude of each sensitive deviation with its cascading impact potential on downstream tasks. In this way, a deviation list sorted only by size can be transformed into a scored event list sorted by actual threat level, providing an objective and quantifiable ranking basis for subsequent triggering decisions and ensuring that the most destructive disturbance events are addressed first.

[0049] Figure 6 This document presents a flowchart illustrating the process of quantifying and scoring the chain effects of a sensitive deviation list based on a port resource dependency graph, using a port multi-source data collaborative intelligent berth scheduling method according to an embodiment of the present invention. For example... Figure 6 As shown, step S320 includes the following steps: S321, extracting the first sensitive deviation from the sensitive deviation list; S322, finding the source task node corresponding to the first sensitive deviation in the port resource dependency graph, and finding the set of all downstream task nodes directly reachable from the source task node; S323, calculating the time difference between the planned start time of each downstream task node in the set of downstream task nodes and the planned end time of the source task node to obtain a set of safety buffer times; S324, calculating the total impact score of the first sensitive deviation based on the set of safety buffer times.

[0050] Specifically, in steps S321 and S322, a first sensitive deviation is extracted from the sensitive deviation list, the source task node corresponding to the first sensitive deviation is found in the port resource dependency graph, and the set of all downstream task nodes directly reachable from the source task node is identified. The first sensitive deviation can be any sensitive deviation in the sensitive deviation list. It should be understood that the first sensitive deviation itself is merely an isolated event record, which does not contain contextual information about the event within the entire scheduling plan, particularly its direct dependency on subsequent scheduled tasks, making it impossible to directly assess its potential cascading effects. Therefore, in this embodiment, the first sensitive deviation is further extracted from the sensitive deviation list, the source task node corresponding to the first sensitive deviation is found in the port resource dependency graph, and the set of all downstream task nodes directly reachable from the source task node is identified. This precisely associates the abstract deviation event with the concrete port resource dependency graph and clearly identifies the group of subsequent tasks at the forefront that are directly affected by the deviation event in the scheduling logic. This provides a clear and unambiguous set of targets for calculating the safety buffer time of each downstream task and quantifying the cascading impact score of the sensitive deviation, thereby transforming the impact assessment from an abstract concept into an executable computational process.

[0051] In practical implementation, steps S321 and S322 can begin by locating the source task node. In this step, a first sensitive deviation to be processed is extracted from the sensitive deviation list, and its associated object identifier, such as a ship's unique ID, is parsed. This object identifier is then used as an index to search the node set of the port resource dependency graph to find a unique node that matches the object identifier and the current planning stage; this node is the source task node. Next, the downstream task node search step is executed. Starting from the located source task node, all outgoing edges of that node are traversed. For each outgoing edge, the target node it points to is obtained, and this target node is added to an initially empty set of downstream task nodes. After traversing all outgoing edges, the set of downstream task nodes contains all downstream tasks that are logically dependent on the source task node, and this serves as the final output of this step.

[0052] Specifically, in step S323, the time difference between the planned start time of each downstream task node in the set of downstream task nodes and the planned end time of the source task node is calculated to obtain a set of safety buffer times. It should be understood that simply identifying downstream task nodes is insufficient to quantify the specific impact of deviations in the source task on these downstream tasks, because this impact directly depends on the size of the time redundancy reserved in the plan to absorb disturbances, which is not directly represented in the dependency graph. Therefore, in this embodiment, the time difference between the planned start time of each downstream task node in the set of downstream task nodes and the planned end time of the source task node is further calculated to obtain a set of safety buffer times. This makes the implicit time coupling relationship in the scheduling plan explicit as a set of precise, measurable time interval values ​​representing the robustness of each dependency chain. This provides a crucial and decisive input for subsequent quantification of the severity of cascading effects based on this time difference, thereby enabling the differentiation of different dependency tightness levels, which is a prerequisite for accurate scoring of cascading effects.

[0053] In its implementation, step S323 first performs a data preparation and initialization step. In this step, the planned end time is extracted from the data attributes of the source task node, and an initially empty safety buffer time set is created to store the calculation results. Next, an iterative calculation step is performed, traversing each downstream task node in the downstream task node set. For the currently traversed downstream task node, its planned start time is extracted from its data attributes, and an arithmetic subtraction operation is performed on the time difference, subtracting the planned end time of the source task node from the planned start time of the downstream task node to obtain a value representing the time interval. This calculated time difference value is then added to the safety buffer time set. Finally, after traversing all downstream task nodes, the safety buffer time set is completed, containing the planned buffer times of the source task for all its directly downstream tasks.

[0054] Specifically, in step S324, the total impact score of the first sensitive deviation is calculated based on the set of safety buffer times. It should be understood that the set of safety buffer times and the value of the first sensitive deviation are independent quantities describing different physical dimensions. They cannot provide a comprehensive, single metric for ranking the overall severity of the deviation event, especially failing to reflect the non-linear relationship that shorter buffer times lead to higher chain reaction risks. Therefore, in this embodiment, the total impact score of the first sensitive deviation is further calculated based on the set of safety buffer times. This is used to fuse the absolute magnitude of the deviation with a cumulative chain reaction factor, calculated based on an exponential decay function, reflecting the tightness of buffer times for all directly downstream tasks, through a weighted summation model. This generates a comprehensive, quantifiable total impact score for each sensitive deviation event. This total impact score not only reflects the magnitude of the deviation but also reveals its potential risk of triggering chain reactions in the scheduling network, thus providing an objective and reliable basis for subsequent triggering decisions based on score ranking.

[0055] Specifically, in practical implementation, step S324 includes: for each safety buffer time in the safety buffer time set, performing natural exponential calculation using the negative of the product of a positive attenuation constant and the safety buffer time as the exponent, to obtain the chain reaction factor corresponding to that safety buffer time, and aggregating all the chain reaction factors corresponding to all safety buffer times into a chain reaction factor set. Because natural exponential calculation is used, the shorter the safety buffer time, the larger the resulting chain reaction factor, thus reflecting a nonlinear relationship where a shorter buffer time leads to a higher risk of chain reaction. The attenuation constant is used to adjust the attenuation rate of this nonlinear relationship. Subsequently, the absolute value of the first sensitive deviation is multiplied by a preset weight for the deviation size to obtain a weighted value for the deviation size; all chain reaction factors in the chain reaction factor set are summed, and the summation result is multiplied by a preset weight for the chain reaction to obtain a weighted value for the chain reaction; finally, the weighted value for the deviation size is added to the weighted value for the chain reaction to obtain the total impact score of the first sensitive deviation.

[0056] Specifically, in step S330, based on the rescheduling trigger score threshold, the scoring event list is subjected to a score-based ranking trigger decision and a breadth-first determination of the impact domain to obtain the triggered rescheduling events. It should be understood that the scoring event list is merely a set of potential risks ranked by severity; it does not provide a clear action trigger standard, nor does it define the specific scope of adjustments needed to address the risk, making it impossible to make a final, executable intervention decision. Therefore, in this embodiment, based on the rescheduling trigger score threshold, the scoring event list is further subjected to a score-based ranking trigger decision and a breadth-first determination of the impact domain to obtain the triggered rescheduling events. This establishes a final, automated decision-making and scope definition checkpoint. This checkpoint first makes a deterministic decision on whether to initiate rescheduling by comparing the highest impact score with a preset threshold. Upon decision initiation, a breadth-first search algorithm is used to systematically trace and delineate the set of tasks affected by the entire chain reaction triggered by the highest-scoring event along the topological structure of the resource dependency graph. In this way, a list containing multiple potential risks can be refined and transformed into a single, directly executable rescheduling event with a clearly defined rescheduling scope. This ensures that only the most threatening disturbances are responded to, and provides a solution problem with clear boundaries and minimal scale for subsequent local re-optimization steps, greatly improving the decision-making efficiency and response accuracy of the entire scheduling system.

[0057] In specific implementation, step S330 can first execute a score-based sorting step, in which all events in the scoring event list are sorted in descending order according to their total impact score. Next, a trigger decision step is executed, in which the highest-scoring event at the top of the sorted list is extracted, and its total impact score is compared with a preset rescheduling trigger score threshold. If the total impact score is greater than the rescheduling trigger score threshold, a rescheduling is triggered, and the process proceeds to the next step; otherwise, the process terminates. Subsequently, after determining that a rescheduling has been triggered, a breadth-first determination step for the impact domain is executed. In this step, a queue and a set for storing impact domain nodes are first initialized, and the source task node corresponding to the highest-scoring event is simultaneously added to both the queue and the set of impact domain nodes. Then, a loop is entered. As long as the queue is not empty, a current node is retrieved from the queue, and all direct downstream neighbor nodes of that current node in the port resource dependency graph are traversed. For each neighbor node that has not yet been added to the impact domain set, it is simultaneously added to both the queue and the set of impact domain nodes. This loop continues until the queue is empty. Finally, the rescheduling event encapsulation step is performed, which encapsulates the original information of the highest-scoring event and the set of influence domains constructed in the previous step into a structured, triggered rescheduling event object, which serves as the final output of this process.

[0058] Specifically, in step S400, based on the unified port status and baseline scheduling plan, the triggered rescheduling event is subjected to local re-optimization based on the influence domain to obtain a scheduling change proposal. It should be understood that the triggered rescheduling event only defines a scheduling conflict that must be resolved and its scope of influence; it does not provide any specific solution. The scheduling system still needs a clear, executable, and optimal new plan under the current constraints to replace the original, ineffective local plan. Therefore, in this embodiment, based on the unified port status and baseline scheduling plan, the triggered rescheduling event is further subjected to local re-optimization based on the influence domain to obtain a scheduling change proposal. This allows for the construction and solution of a manageable mathematical optimization model aimed at minimizing local disturbances within the precisely defined influence domain. This model uses the latest port status as input constraints and aims to find a new, optimal scheduling arrangement for all tasks within the influence domain. In this way, an abstract rescheduling requirement can be transformed into a specific, quantifiable, and operationally feasible scheduling change proposal. This proposal not only resolves the direct conflict caused by the initial disturbance, but also ensures that resource utilization and work processes within the affected area are restored to their optimal state through optimization. It is the final execution step to achieve dynamic self-healing of the scheduling plan.

[0059] In its implementation, step S400, the local re-optimization process, first executes the optimization model construction step. In this step, a local mathematical optimization model is constructed only for the task entities within the influence domain of the triggered rescheduling event. This model uses the new berthing time and new berth of each ship within the influence domain as decision variables, and aims to minimize the total ship delay time, berth idle time, or plan change cost within the influence domain. It employs resource constraints (e.g., berths cannot be repeatedly occupied at the same time), timing constraints (e.g., ships cannot berth before their predicted arrival time), and influence domain boundary constraints (i.e., adjustments to tasks within the influence domain cannot conflict with fixed plans for tasks outside the domain) as constraints. Next, the model solution step is executed. In this step, an embedded, efficient mathematical programming solver, such as a mixed-integer programming solver, is invoked to quickly solve the local mathematical optimization model constructed in the previous step, obtaining a set of decision variable values ​​that satisfy all constraints and optimize the objective function within seconds or minutes. Finally, the step of generating the scheduling change proposal is executed. In this step, the optimal solution returned by the solver, i.e. the new optimal scheduling arrangement for each task in the influence domain, is compared with the original arrangement in the baseline scheduling plan, and the differences are formatted into a clear and structured scheduling change proposal as the final output of the process.

[0060] In summary, the intelligent berth scheduling method based on multi-source data collaboration in ports according to embodiments of the present invention is explained. Firstly, it uses a machine learning model to transform multi-source real-time data into accurate predictions of future events (such as ship arrivals and operation durations), and proactively calculates the deviation vector between the prediction and the baseline plan, thereby solving the core problems of passive response and lack of foresight in existing technologies. Secondly, it does not process all deviation vectors, but rather determines their severity by quantifying the chain reaction impact score of each deviation vector, and dynamically defines the minimized impact domain based on a resource dependency graph. Only when the impact score exceeds a preset threshold is a rapid re-optimization triggered specifically for that local impact domain. This precise adjustment method completely changes the predicament of traditional technologies blindly performing global recalculation, leading to delayed response and unstable plans, ensuring the real-time performance, robustness, and low cost of scheduling decisions in highly dynamic environments.

[0061] Another embodiment of the present invention provides an intelligent berth scheduling system for port multi-source data collaboration.

[0062] Figure 7 This is a block diagram of a port multi-source data collaborative intelligent berth scheduling system according to an embodiment of the present invention. Figure 7 As shown, the intelligent berth scheduling system 100 for port multi-source data collaboration according to an embodiment of the present invention includes: a digital twin state synchronization module 110, used to perform digital twin state synchronization of the port's original AIS data, original meteorological data, original TOS data and original quay crane data to obtain a unified port state; a deviation vector calculation module 120, used to perform machine learning-based disturbance prediction and benchmark deviation calculation on the unified port state based on the benchmark scheduling plan to obtain a deviation vector; a rescheduling event triggering module 130, used to trigger disturbance events and evaluate the impact domain based on the deviation vector to obtain triggered rescheduling events; and a local re-optimization module 140, used to perform local re-optimization based on the impact domain on the triggered rescheduling events based on the unified port state and the benchmark scheduling plan to obtain scheduling change proposals.

[0063] The deviation vector calculation module 120 is specifically used for: extracting the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th vessel en route from the unified port status; inputting the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th vessel en route into the pre-trained ETA prediction model to obtain a predicted ETA list; extracting the cargo operation data of the i-th vessel en route from the baseline scheduling plan, allocating quay crane data and quay crane efficiency data to the i-th vessel en route; based on the cargo operation data, the quay crane data and quay crane efficiency data allocated to the i-th vessel en route, predicting berth operation time based on equipment efficiency and operational constraints to obtain a predicted operation time list; and vectorizing the predicted operation time list and the predicted ETA list to obtain a deviation vector.

[0064] The specific implementation method of the intelligent berth scheduling system based on multi-source data collaboration in ports provided in this embodiment of the invention can be found in the description of the intelligent berth scheduling method based on multi-source data collaboration in ports provided in this embodiment of the invention, and will not be repeated here.

[0065] The port multi-source data collaborative intelligent berth scheduling system 100 according to embodiments of the present invention can be implemented in various wireless terminals, such as servers with multi-sensor-based bucket wheel excavator assisted driving algorithms. In one possible implementation, the port multi-source data collaborative intelligent berth scheduling system 100 according to embodiments of the present invention can be integrated into the wireless terminal as a software module and / or hardware module. For example, the port multi-source data collaborative intelligent berth scheduling system 100 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the port multi-source data collaborative intelligent berth scheduling system 100 can also be one of many hardware modules of the wireless terminal.

[0066] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for intelligent berth scheduling based on multi-source data collaboration in ports, characterized in that, include: The port's original AIS data, original meteorological data, original TOS data, and original quay crane data are synchronized digitally to obtain a unified port status. Based on the baseline scheduling plan, machine learning-based disturbance prediction and baseline deviation calculation are performed on the unified port status to obtain the deviation vector; Based on the deviation vector, the triggering of disturbance events and the evaluation of the impact domain are performed to obtain the triggered rescheduling events; Based on the unified port status and baseline scheduling plan, the triggered rescheduling events are subjected to local re-optimization based on the influence domain to obtain scheduling change proposals.

2. The intelligent berth scheduling method for port multi-source data collaboration according to claim 1, characterized in that, Based on the baseline scheduling plan, machine learning-based disturbance prediction and baseline deviation calculation are performed on the unified port status to obtain the deviation vector, including: Extract the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th vessel from the unified port status. The dynamic data of the i-th vessel in transit, the static data of the i-th vessel in transit, the target route and berth data of the i-th vessel in transit, and the environmental data along the predicted path are input into the pre-trained ETA prediction model to obtain the predicted ETA list. Extract cargo operation data of the i-th vessel en route from the baseline scheduling plan, and allocate quay crane data and quay crane efficiency data to the i-th vessel en route. Based on the cargo operation data of the i-th vessel in transit, the quay crane data and quay crane efficiency data allocated to the i-th vessel in transit, the berth operation time is predicted based on equipment efficiency and operational constraints, resulting in a list of predicted operation times. The predicted job duration list and the predicted ETA list are vectorized to obtain the prediction-planning deviation vector.

3. The intelligent berth scheduling method for port multi-source data collaboration according to claim 2, characterized in that, The dynamic data of the i-th vessel en route, the static data of the i-th vessel en route, the target route and berth data of the i-th vessel en route, and the environmental data along the predicted path are input into the pre-trained ETA prediction model to obtain a list of predicted ETAs, including: Feature engineering is performed on the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th vessel in transit to obtain the feature vector of the i-th vessel in transit. Input the feature vector of the i-th ship en route into the pre-trained ETA prediction model to obtain the predicted time required for the i-th ship en route to complete the remaining voyage. Add the predicted time required for the i-th vessel to complete the remaining voyage to the current timestamp to obtain the predicted ETA for the i-th vessel.

4. The intelligent berth scheduling method for port multi-source data collaboration according to claim 3, characterized in that, The feature vector of the i-th vessel en route includes its current ground speed, current ground heading, average speed over a preset time period, standard deviation of speed over a preset time period, instantaneous acceleration, turning rate, remaining distance, total planned distance, number of remaining waypoints, whether it is in a waterway, distance to pilot station, vessel type code, overall length, beam, current draft, gross tonnage, average wind speed of segment k, average wind direction of segment k, average relative wind angle of segment k, average ocean current speed of segment k, average ocean current direction of segment k, average relative current angle of segment k, average wave height of segment k, average visibility of segment k, vessel punctuality score, number of vessels at current anchorage, and occupancy rate of the wharf where the target berth is located.

5. The intelligent berth scheduling method for port multi-source data collaboration according to claim 1, characterized in that, Based on the deviation vector, disturbance event triggering and impact domain assessment are performed to obtain triggered rescheduling events, including: Based on the dynamic threshold configuration library, the deviation vector is filtered for sensitive deviations based on multi-dimensional dynamic thresholds to obtain a list of sensitive deviations. The list of sensitive deviations is quantified and scored based on the cascading effects of port resource dependence graphs to obtain a list of scored events; Based on the rescheduling trigger score threshold, the triggering decision and influence domain breadth priority determination are performed on the scoring event list to obtain the triggered rescheduling events.

6. The intelligent berth scheduling method for port multi-source data collaboration according to claim 5, characterized in that, The list of sensitive deviations is quantified and scored based on the cascading effects of port resource dependence graphs, resulting in a list of scored events, including: Extract the first sensitive deviation from the list of sensitive deviations; In the port resource dependency graph, find the source task node corresponding to the first sensitive deviation, and find the set of all downstream task nodes directly reachable from the source task node. Calculate the time difference between the planned start time of each downstream task node in the set of downstream task nodes and the planned end time of the source task node to obtain a set of safety buffer times; The total impact score of the first sensitive deviation is calculated based on the set of safety buffer times.

7. A port multi-source data collaborative intelligent berth scheduling system, characterized in that, include: The digital twin status synchronization module is used to synchronize the port's original AIS data, original meteorological data, original TOS data and original quay crane data with digital twins to obtain a unified port status. The deviation vector calculation module is used to perform machine learning-based disturbance prediction and baseline deviation calculation on the unified port status based on the baseline scheduling plan, and obtain the deviation vector. The rescheduling event triggering module is used to trigger disturbance events and evaluate the impact domain based on the deviation vector, so as to obtain the triggered rescheduling events. The local re-optimization module is used to perform local re-optimization based on the influence domain of triggered rescheduling events, based on the unified port status and baseline scheduling plan, to obtain scheduling change proposals.

8. The intelligent berth scheduling system for port multi-source data collaboration according to claim 7, characterized in that, The deviation vector calculation module is specifically used for: Extract the dynamic data, static data, target route and berth data, and environmental data along the predicted path of the i-th vessel from the unified port status. The dynamic data of the i-th vessel in transit, the static data of the i-th vessel in transit, the target route and berth data of the i-th vessel in transit, and the environmental data along the predicted path are input into the pre-trained ETA prediction model to obtain the predicted ETA list. Extract cargo operation data of the i-th vessel en route from the baseline scheduling plan, and allocate quay crane data and quay crane efficiency data to the i-th vessel en route. Based on the cargo operation data of the i-th vessel in transit, the quay crane data and quay crane efficiency data allocated to the i-th vessel in transit, the berth operation time is predicted based on equipment efficiency and operational constraints, resulting in a list of predicted operation times. The predicted job duration list and the predicted ETA list are vectorized to obtain the prediction-planning deviation vector.

Images