Terminal truck dynamic optimization scheduling method and system combining machine learning and path planning

Abstract

This invention relates to the field of intelligent terminal technology, specifically to a method and system for dynamic optimization scheduling of container trucks at terminals, combining machine learning and path planning. It includes: a behavior data acquisition and feature coupling unit; a learnable incentive and behavior guidance unit; a scheduling demand prediction unit; and a path planning and scheduling unit. Based on coupled features, this invention optimizes incentive coefficients and dynamically pushes incentive instructions through reinforcement learning, achieving targeted guidance for non-operational stopover behavior of container trucks. Based on standardized time-series data, this invention learns the correlation between historical stopover periods and operating conditions through an LSTM model and optimizes prediction results by combining real-time data, providing a forward-looking constraint basis for scheduling. Simultaneously, it constructs a multi-dimensional path constraint system encompassing time, space, resources, and priority through a structural causal model, and uses an improved A* algorithm with integrated dynamic weights to complete container truck-berth matching and dynamic path planning, effectively avoiding scheduling conflicts.

Description

Technical Field

[0001] This invention relates to the field of smart terminal technology, and more specifically, to a method and system for dynamic optimization scheduling of terminal container trucks that combines machine learning and path planning. Background Technology

[0002] Terminal truck scheduling is a core component of efficient port logistics, directly impacting container handling and turnover efficiency, yard resource utilization, and overall operating costs. Its precision is crucial for ports to cope with ever-increasing throughput and complex, ever-changing operational scenarios. Currently, traditional scheduling models relying on manual experience or single, fixed rules are no longer adequate for dynamic and multi-variable operational needs, necessitating an upgrade to scheduling logic through data-driven and algorithm-optimized technologies.

[0003] In the existing technology, relevant patents have explored the field of terminal truck scheduling. For example, Chinese patent CN202210528907.X discloses a scheduling method for port automated trucks without terminal operations, including a terminal operation system, a vehicle management system, automated trucks, and automated operation equipment. By predicting the current and next operation type of the truck and allocating appropriate special tasks based on the direction of operation, the truck can move in advance to avoid waiting in place and improve operation efficiency. Another example is Chinese patent CN202210012431.4, which discloses a terminal vehicle scheduling system, method, equipment, and medium, including task allocation, single-vehicle scheduling, traffic control, and route planning modules. It realizes task allocation, single-vehicle scheduling, intersection priority passage, and optimal route planning for mixed operation scenarios, meeting the needs of loading / unloading and emergency pre-scheduling.

[0004] Despite the design advantages of the aforementioned technical solutions, they also suffer from the following technical shortcomings: First, they lack targeted guidance for non-operational truck dwell time and the ability to mine multi-source data features: Chinese patent CN202210528907.X only achieves pre-movement of trucks through task type prediction, failing to address the critical scenario of non-operational dwell time that affects scheduling efficiency, and also failing to integrate multi-source data mining and correlation features such as truck operation, yard operations, and task attributes, thus failing to accurately match guidance strategies with actual operational needs; Second, path planning lacks multi-dimensional constraints and forward... The path planning in Chinese patent CN202210012431.4 focuses solely on the "shortest distance" objective, neglecting crucial constraints such as truck dwell time, resource occupancy status, and task urgency. Furthermore, neither CN202210012431.4 nor CN202210528907.X provides forward-looking predictions of truck operating conditions, making it difficult to proactively avoid scheduling conflicts. Moreover, they lack a closed loop for execution data feedback and dynamic optimization of scheduling strategies, hindering rapid adaptation and adjustment after anomalies occur. Therefore, we propose a dynamic optimization scheduling method and system for terminal trucks that combines machine learning and path planning. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic optimization scheduling method and system for container trucks at terminals that combines machine learning and route planning, in order to solve the problems mentioned in the background art, such as the lack of targeted guidance for non-operational stop behavior of container trucks and the ability to mine multi-source data features, the lack of multi-dimensional constraints and forward-looking prediction support for route planning, and the lack of a closed-loop adjustment mechanism for the entire process.

[0006] To address the aforementioned technical problems, one objective of this invention is to provide a dynamic optimization scheduling system for terminal container trucks that combines machine learning and route planning, comprising:

[0007] The behavioral data acquisition and feature coupling unit collects relevant data on trucks, yards, and scheduling through GPS modules, RFID readers, driver-side APPs, and port management system interfaces. After data cleaning, it constructs behavioral-scheduling coupled features and stores standardized time-series data in a database.

[0008] The learnable incentive and behavior guidance unit is based on behavior-scheduling coupling characteristics. It uses "non-task dwell event + task urgency matching" as the trigger condition, optimizes the incentive coefficient through reinforcement learning to generate incentive amount, pushes real-time micro-incentives through the driver's mobile APP and executes automatic re-push; when multiple drivers dwell at the same time, the driver with the longer dwell time is given priority in incentive; after the incentive is triggered, the driver's dwell time information is fed back to the subsequent route planning and scheduling stage in real time, so that the route planning uses the dwell time as a dynamic constraint condition and automatically skips the dwell time to generate a new task route;

[0009] The scheduling demand prediction unit is based on standardized time series data and uses an LSTM model to learn the correlation between historical dwell time periods and working conditions. It predicts the possible dwell time periods of each truck in a future specified period and optimizes the model output by supplementing it with the latest real-time data, providing a forward-looking constraint basis for scheduling.

[0010] The route planning and scheduling unit receives the dwell time prediction results from the scheduling demand prediction unit, the real-time dwell events and driver response status from the learnable incentive and behavior guidance unit, and the multi-source operation data from the behavior data acquisition and feature coupling unit. It generates path constraints through a structural causal model, uses an improved A* algorithm to complete truck-berth matching and dynamic route planning, and combines resource pre-occupancy and elastic release mechanisms to synchronously feed back scheduling data to the scheduling instruction execution and dynamic adjustment unit and the learnable incentive and behavior guidance unit.

[0011] The scheduling instruction execution and dynamic adjustment unit converts the scheduling scheme into standardized instructions and pushes them to the truck terminal. It collects the truck's operating status, berth operation status, and scheduling scheme execution data in real time, and determines abnormal states through multi-dimensional verification. When an abnormality occurs, it calls the path planning function of the path planning and scheduling unit to generate an alternative scheme and update the instructions. Simultaneously, it feeds back the instruction execution results, status monitoring data, and abnormal handling information to the behavior data acquisition and feature coupling unit, the learnable incentive and behavior guidance unit, the scheduling demand prediction unit, and the path planning and scheduling unit according to functional requirements.

[0012] As a further improvement to this technical solution, the behavior data acquisition and feature coupling unit includes a multi-source data access module, a data cleaning and calibration module, a behavior-scheduling feature construction module, and a standardized data storage module, wherein:

[0013] The multi-source data access module collects truck location, speed and start / stop status data through GPS module, collects container loading and unloading status and berth occupancy information in yard through UHF RFID reader, collects driver-marked reasons for stopping data through driver-end APP interaction interface, and connects to the terminal management system through HTTPRESTful interface to synchronize task type, task urgency and berth operation status data. Each data is accompanied by a timestamp.

[0014] The data cleaning and calibration module, based on the output data of the multi-source data access module, performs trajectory continuity verification and corrects drift data for GPS data, performs adjacent state consistency verification and corrects conflict states for RFID data, fills in continuous and discrete data for missing data respectively, and removes time-series segments with an excessive proportion of invalid data.

[0015] The behavior-scheduling feature construction module extracts behavioral features such as non-operation dwell time and operation response time of container trucks, as well as scheduling business features such as task urgency and yard congestion index, based on the effective data from the data cleaning and calibration module, and constructs coupled features of behavior and scheduling requirements.

[0016] The standardized data storage module, based on the feature data of the behavior-scheduling feature construction module, converts the original data and coupled feature data into a unified columnar storage format, stores them in a distributed database partitioned by "truck number-date", and establishes indexes for timestamp, truck number, and berth number.

[0017] As a further improvement to this technical solution, the learnable incentive and behavior guidance unit includes an incentive trigger judgment module, a reinforcement learning optimization module, and an incentive push and data feedback module. Each module works collaboratively according to the "trigger-optimization-push-feedback" logic, wherein:

[0018] The incentive triggering judgment module receives the behavior-scheduling coupling features output by the behavior data acquisition and feature coupling unit as the core input for determining the triggering condition;

[0019] The reinforcement learning optimization module receives the trigger event data output by the incentive trigger judgment module, trains the policy network model using the PPO algorithm, and outputs the incentive coefficients.

[0020] The incentive push and data feedback module generates specific incentive instructions based on the incentive coefficients output by the reinforcement learning optimization module and pushes them; and synchronizes the incentive execution results to the path planning and scheduling unit and the behavior data acquisition and feature coupling unit to form a closed loop of data flow.

[0021] As a further improvement to this technical solution, the process of the incentive trigger judgment module performing the "non-job stay event + task urgency matching" judgment includes the following steps;

[0022] S21.1 Receive the behavior-scheduling coupling features output by the behavior data acquisition and feature coupling unit, and synchronously acquire the continuous location data of the truck from the GPS module, the operation status data of the corresponding cargo location in the yard from the RFID reader, and the current task deadline data from the terminal management system.

[0023] S21.2 Set a position deviation threshold and a continuous sampling point number threshold. When both conditions are met simultaneously, namely "the position deviation of the GPS module's continuous preset number of sampling points is less than or equal to the position deviation threshold" and "the RFID reader reports no loading or unloading action record for the corresponding cargo location", and the duration of these conditions is greater than or equal to the preset duration threshold, it is determined to be a non-operational stop event.

[0024] S21.3 Based on the difference between the task deadline and the current time, the urgency value is quantified into a 0-1 range using a linear mapping. The larger the value, the more urgent the task. The mapping benchmark is the longest task cycle preset by the dock.

[0025] S21.4 When the non-work stoppage event is determined to be true, and the task urgency value is... When the task is in the preset medium urgency range, the trigger condition is determined to be met, and the trigger signal, non-task dwell time, and task urgency value are output to the reinforcement learning optimization module. And the yard congestion index.

[0026] As a further improvement to this technical solution, the reinforcement learning optimization module receives the trigger event data output by the incentive trigger judgment module, trains the policy network model using the PPO algorithm, and outputs the incentive coefficients in the following steps;

[0027] S22.1 Construct a policy network model for the PPO agent. The network uses non-task dwell time and task urgency values ​​from the trigger event data. The yard congestion index is the input feature, and the output is the incentive coefficient corresponding to the dwell time weight. Incentive coefficients corresponding to the weights of the congestion index The network has multiple fully connected layers, and nonlinear feature mapping is achieved through ReLU and tanh activation functions;

[0028] S22.2 Define a two-dimensional reward signal: the core reward is negatively correlated with the driver's response time, the constraint reward is negatively correlated with the incentive amount, and the total reward signal is the weighted sum of the core reward and the constraint reward.

[0029] S22.3 Daily collection of "Trigger Condition Data - Incentive Coefficient" , The triplet of "driver response result" is used as a training sample and input into the policy network for incremental training. The parameter update range is limited by the parameter pruning mechanism of the PPO algorithm.

[0030] S22.4, Call the optimal activation coefficients output by the trained network. and The real-time incentive amount is calculated by combining the non-operational dwell time and yard congestion index in the trigger event data. .

[0031] As a further improvement to this technical solution, the process of the incentive push and data feedback module executing incentive instruction push and data synchronization includes the following steps;

[0032] S23.1 When multiple drivers simultaneously trigger non-operational stop events, calculate the priority score for each driver based on the stop duration and yard congestion index output by the incentive trigger judgment module. , Positively correlated with dwell time and congestion index, according to Incentive resources are allocated and pushed in descending order.

[0033] S23.2, Push incentive amounts to target drivers via driver-side app The instruction is sent, and a response timer is started. If no "location movement confirmation" signal is received from the app within the first preset response time, a second push is triggered. If there is still no response after the second push and a second preset response time, the push is terminated, and the incentive amount for the second push is adjusted to [amount missing]. , for The product of the preset incentive adjustment coefficient;

[0034] S23.3 After the excitation is triggered, the start and end timestamps of the dwell time period marked in the GPS data are extracted, encapsulated into a standardized data frame, and pushed to the path planning and scheduling unit in real time via wireless communication;

[0035] S23.4. Synchronously feed back the driver's response status and the actual dwell time change data to the standardized data storage module of the behavior data acquisition and feature coupling unit.

[0036] As a further improvement to this technical solution, the scheduling demand prediction unit includes a data input and preprocessing module, an LSTM model processing module, and a real-time output optimization module. Each module works collaboratively according to the logic of "data preprocessing - model training and inference - real-time optimization," wherein:

[0037] The data input and preprocessing module receives standardized time-series data output by the behavior data acquisition and feature coupling unit, performs format regularization and feature extraction on the historical dwell time period and working condition information in the data, and forms the basic data for the input of the LSTM model.

[0038] The LSTM model processing module, based on the basic data input and output of the preprocessing module, uses the LSTM model to learn the correlation between historical dwell time and working conditions, and infers the possible dwell time of each truck in the future specified period.

[0039] The real-time output optimization module receives the latest real-time operating condition data, supplements the output of the LSTM model processing module with the real-time operating condition data, dynamically optimizes the prediction results of possible dwell time periods, and forms the forward-looking constraint basis required for scheduling.

[0040] As a further improvement to this technical solution, the path planning and scheduling unit includes a data integration and preprocessing module, a path constraint generation module, a dynamic path planning and berth matching module, a resource management module, and a data feedback module. Each module works collaboratively according to the logic of "data integration - constraint generation - path planning - resource scheduling - data feedback," wherein:

[0041] The data integration and preprocessing module synchronously accesses the future stay time prediction results output by the scheduling demand prediction unit, the real-time stay event and driver response status data output by the learnable incentive and behavior guidance unit, and the multi-source operation data output by the behavior data collection and feature coupling unit. It performs format standardization conversion and timestamp alignment on the heterogeneous data, removes expired data that exceeds the effective time window, and forms a unified scheduling input dataset.

[0042] The path constraint generation module, based on the standardized dataset of the data integration and preprocessing module, constructs a multi-dimensional path constraint system through a structural causal model: it uses predicted dwell time and real-time dwell time as "time constraints" to clarify the unschedulable periods for trucks; it maps the yard congestion index as "spatial constraints" to define the path passage priority in high-congestion areas; it uses berth operation status and cargo space occupancy information as "resource constraints" to limit the range of berths that trucks can match; and it uses task urgency as "priority constraints" to set the execution order weight of scheduling tasks.

[0043] The dynamic route planning and berth matching module uses the multi-dimensional constraints of the route constraint generation module as boundary conditions and adopts an improved A* algorithm with integrated dynamic weights to realize truck-berth matching and route planning. The heuristic function design incorporates the task urgency coefficient and the dynamic weight of yard congestion. The higher the urgency and the lower the congestion index, the greater the weight of the route. During the route search process, the module avoids the dwell time period limited by "time constraints" and the high congestion area limited by "space constraints" in real time, and simultaneously verifies the availability of berth "resource constraints". The module outputs the optimal truck scheduling route and berth matching scheme that takes into account both the shortest route and the timeliness of the task.

[0044] The resource management module, based on the scheduling scheme output by the dynamic path planning and berth matching module, initiates a resource pre-reservation mechanism: after the truck scheduling path is generated, it automatically pre-reserves the specified operation time of the target berth and the passage resources of key nodes along the path. The pre-reservation duration is set according to the estimated operation time of the task plus a preset buffer time. When the truck completes the operation, the task is canceled, or an abnormal adjustment occurs, the elastic release mechanism is triggered to release the pre-reserved resources in real time and update the resource occupancy status database to ensure resource reuse efficiency.

[0045] The data feedback module feeds back the truck-berth matching results and dynamic path information output by the dynamic path planning and berth matching module, as well as the resource pre-occupancy and release status data from the resource management module, to the scheduling instruction execution and dynamic adjustment unit and the learnable incentive and behavior guidance unit, respectively. The data fed back to the learnable incentive and behavior guidance unit is used to optimize the training sample library of the incentive coefficient.

[0046] As a further improvement to this technical solution, the scheduling instruction execution and dynamic adjustment unit includes a standardized instruction generation and push module, a multi-source status data acquisition module, an anomaly verification and judgment module, an alternative scheme generation and update module, and a targeted data feedback module, wherein:

[0047] The standardized instruction generation and push module converts the scheduling scheme output by the path planning and scheduling unit into standardized instructions and pushes them to the truck terminal.

[0048] The multi-source status data acquisition module collects real-time data on truck operation status, berth operation status, and scheduling plan execution.

[0049] The anomaly verification and judgment module performs multi-dimensional verification to determine the anomaly status of the data collected by the multi-source state data acquisition module;

[0050] When the anomaly verification and judgment module determines that an anomaly exists, the alternative solution generation and update module calls the path planning function of the path planning and scheduling unit to generate an alternative solution and update the standardized instructions.

[0051] The directional data feedback module feeds back the instruction execution results, status monitoring data, and anomaly handling information to the behavior data acquisition and feature coupling unit, the learnable incentive and behavior guidance unit, the scheduling demand prediction unit, and the path planning and scheduling unit according to functional requirements.

[0052] The second objective of this invention is to provide a dynamic optimization scheduling method for terminal container trucks that combines machine learning and route planning. Based on the aforementioned dynamic optimization scheduling system for terminal container trucks that combines machine learning and route planning, the method includes the following steps:

[0053] S1. Behavioral data collection and feature coupling: Multi-source timestamped data is collected through GPS, RFID readers, driver's APP and port management system interface. After cleaning and calibration, behavioral and scheduling business features are extracted to construct coupled features. After standardization, the data is partitioned and stored in a distributed database according to "truck number-date".

[0054] S2, Learnable Incentive Trigger and Behavior Guidance: Based on the coupling features of S1, non-work-related stay events and medium-urgency tasks are determined. The incentive coefficient is output through the PPO algorithm to calculate the amount. The amount is pushed according to the priority of stay duration and congestion index. If there is no response, the amount is adjusted and the push is re-pushed. Relevant data is simultaneously fed back to subsequent stages.

[0055] S3. Forward-looking prediction of scheduling demand: Based on the standardized time-series data of S1, the correlation between historical dwell times and working conditions is learned through the LSTM model to predict the possible dwell times of trucks in the future, supplement the optimization results of real-time working condition data, and form a forward-looking constraint basis;

[0056] S4. Route Planning and Scheduling Execution: Integrate relevant data from S1 to S3, construct a four-dimensional constraint system through a structural causal model, use an improved A* algorithm to complete truck-berth matching and route planning, simultaneously activate resource pre-occupancy and elastic release mechanisms, and feed back scheduling data to the corresponding links;

[0057] S5. Execution and Dynamic Adjustment of Scheduling Instructions: The S4 scheduling scheme is converted into standardized instructions and pushed to the truck terminal. Real-time data on operation and work status is collected. After three-dimensional verification to determine anomalies, alternative schemes are generated and instructions are updated. Relevant information is fed back to each upstream link.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] 1. This invention collects multi-source timestamped data through GPS modules, RFID readers, driver-side APPs, and dock management system interfaces. After cleaning and calibration, it constructs behavior-scheduling coupling features. Based on these coupling features, it uses "non-operational stop events + task urgency matching" as trigger conditions, optimizes the incentive coefficient through reinforcement learning, and dynamically pushes incentive instructions. This achieves targeted guidance for non-operational stop behaviors of container trucks and accurately matches actual operational needs.

[0060] 2. Based on standardized time-series data, this invention learns the correlation between historical dwell time periods and working conditions through an LSTM model and optimizes prediction results by combining real-time data, providing a forward-looking constraint basis for scheduling. At the same time, it constructs a multi-dimensional path constraint system of time, space, resources, and priority through a structural causal model, and completes truck-berth matching and dynamic path planning with an improved A* algorithm that integrates dynamic weights, effectively avoiding scheduling conflicts.

[0061] 3. This invention constructs a closed-loop feedback system for the entire scheduling process. During the execution of scheduling instructions, various status data are collected in real time. Anomalies are identified through multi-dimensional verification and alternative solutions are generated. At the same time, relevant data is fed back to each functional unit in a targeted manner, realizing the dynamic optimization of scheduling strategies, incentive models and prediction results, and improving the adaptability and robustness of scheduling schemes. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0063] Figure 2 This is a schematic diagram of the method steps of the present invention;

[0064] Figure 3 This is a schematic diagram of the module composition of the behavioral data acquisition and feature coupling unit in this invention;

[0065] Figure 4 This is a schematic diagram of the modular composition of the learnable incentive and behavior guidance unit in this invention;

[0066] Figure 5 This is a schematic diagram of the module composition of the scheduling demand prediction unit in this invention;

[0067] Figure 6 This is a schematic diagram of the module composition of the path planning and scheduling unit in this invention;

[0068] Figure 7 This is a schematic diagram of the module composition of the scheduling instruction execution and dynamic adjustment unit in this invention;

[0069] The meanings of the labels in the diagram are as follows:

[0070] 1. Behavioral data acquisition and feature coupling unit; 11. Multi-source data access module; 12. Data cleaning and calibration module; 13. Behavioral-scheduling feature construction module; 14. Standardized data storage module;

[0071] 2. Learnable incentive and behavior guidance unit; 21. Incentive trigger judgment module; 22. Reinforcement learning optimization module; 23. Incentive push and data feedback module;

[0072] 3. Scheduling demand prediction unit; 31. Data input and preprocessing module; 32. LSTM model processing module; 33. Real-time output optimization module;

[0073] 4. Route planning and scheduling unit; 41. Data integration and preprocessing module; 42. Route constraint generation module; 43. Dynamic route planning and berth matching module; 44. Resource management module; 45. Data feedback module;

[0074] 5. Scheduling instruction execution and dynamic adjustment unit; 51. Standardized instruction generation and push module; 52. Multi-source status data acquisition module; 53. Anomaly verification and judgment module; 54. Alternative solution generation and update module; 55. Targeted data feedback module. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0076] like Figure 1 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown, this embodiment provides a dynamic optimization scheduling system for terminal container trucks that combines machine learning and route planning, including:

[0077] Behavior data acquisition and feature coupling unit 1: Behavior data acquisition and feature coupling unit 1 collects relevant data on container trucks, yards, and scheduling through GPS module, RFID reader, driver's APP and terminal management system interface. After data cleaning, behavior-scheduling coupled features are constructed and standardized time series data are stored in the database.

[0078] In this embodiment, the behavior data acquisition and feature coupling unit 1 includes a multi-source data access module 11, a data cleaning and calibration module 12, a behavior-scheduling feature construction module 13, and a standardized data storage module 14, wherein:

[0079] The multi-source data access module 11 collects truck location, speed and start / stop status data through GPS module, collects container loading and unloading status and occupancy information in yard through UHF RFID reader, collects driver-marked reasons for stopping data through driver-end APP interaction interface, and connects to terminal management system through HTTPRESTful interface to synchronize task type, task urgency and berth operation status data. Each data is accompanied by a timestamp.

[0080] Specifically, the multi-source data access module 11 adopts the implementation method of "multi-device collaboration + time synchronization" to ensure the complete collection and time consistency of four types of data: truck operation, yard status, driver behavior, and scheduling tasks. It selects an industrial-grade Beidou / GPS dual-mode positioning module adapted to the complex environment of the terminal (obstruction, metal interference) and collects data such as the truck's unique identifier, latitude and longitude, driving speed, and start / stop status at a fixed frequency (such as 1Hz). The data is then transmitted to the edge computing gateway through the UART serial port.

[0081] At the same time, UHF RFID readers conforming to ISO18000-6C standards are used to deploy reading antennas at the entrances and exits of each cargo location in the yard and in the operation areas of loading and unloading equipment (gantry cranes, quay cranes). The information such as container tag ID, cargo location number, loading and unloading status is collected in the form of "event triggering + fixed periodic supplementary collection (such as 30 seconds / time)" and transmitted to the gateway via Ethernet.

[0082] Furthermore, the driver's app has a built-in module for interacting with reasons for stopping, providing drop-down selection boxes for "equipment waiting, loading / unloading waiting, rest, malfunction, and other". After the driver triggers the truck to stop, a pop-up window will prompt a marker. The marker data (including truck number, marker timestamp, reason for stopping code, and driver ID) is uploaded via HTTPS protocol.

[0083] In addition, the system connects to the terminal management system via an HTTP RESTful interface to synchronize data such as task number, task type, task urgency, berth number, and berth operation status at fixed intervals (e.g., every 30 seconds). Interface requests include an authentication token for security. All collected data is accompanied by a millisecond-level UTC timestamp, and full data time synchronization is achieved through the gateway's built-in NTP service.

[0084] The data cleaning and calibration module 12, based on the output data of the multi-source data access module 11, performs trajectory continuity verification and corrects drift data for GPS data, performs adjacent state consistency verification and corrects conflict states for RFID data, fills in continuous and discrete data for missing data respectively, and removes time-series segments with an excessive proportion of invalid data.

[0085] Specifically, the data cleaning and calibration module 12 designs differentiated processing rules based on the characteristics of different types of data to ensure data validity. The specific steps of each processing method are as follows:

[0086] GPS data cleaning and drift correction: First, extract the latitude, longitude, and timestamp of two adjacent sampling points, and calculate the straight-line distance between the two points using the Haversine formula; then, calculate the theoretical driving speed based on the distance and time difference; next, if the theoretical speed exceeds the reasonable driving speed range of the truck (e.g., 0-80km / h), it is judged as abnormal data and removed; finally, for drift points (e.g., the latitude and longitude of a single point deviates from the normal driving route by more than a preset distance threshold, such as 50m), the average of adjacent valid points is used for smoothing correction to maintain the continuity of the trajectory.

[0087] RFID data conflict correction: First, extract continuous collection period data for the same cargo location or container; then, compare the loading and unloading status and cargo location number of adjacent periods. If a logical conflict occurs (e.g., the previous period is "loading and unloading in progress", and the next period directly jumps to "idle" with no work completion record); finally, use the most recent valid status of the current period as the benchmark to correct the conflicting data and ensure the consistency of status flow.

[0088] Missing data imputation: First, identify missing items in the data and distinguish between continuous data (such as speed, latitude and longitude) and discrete data (such as loading and unloading status, reasons for stopping); then, for continuous missing data, use linear interpolation to calculate the imputation value based on the valid data before and after the missing item; finally, for discrete missing data, statistically analyze the historical valid data of the same truck or cargo location and select the mode with the highest frequency as the imputation value.

[0089] Invalid data removal: First, divide the time-series data according to "cargo number - time segment" (e.g., each segment is 10 minutes); then, calculate the total percentage of abnormal and missing data in each segment; finally, if the percentage exceeds the preset threshold (e.g., 30%), the data quality of that segment is deemed substandard and is directly removed to avoid affecting subsequent feature calculations.

[0090] The behavior-scheduling feature construction module 13 extracts behavioral features such as non-operation dwell time and operation response time of container trucks, as well as scheduling business features such as task urgency and yard congestion index, based on the effective data of the data cleaning and calibration module 12, and constructs coupled features of behavior and scheduling requirements.

[0091] Specifically, the behavior-scheduling feature construction module 13, based on cleaned and effective data, achieves "precise coupling of behavior features and scheduling business features": through correlation analysis of truck start / stop status and RFID loading / unloading status, it calculates the non-operation dwell time of trucks when they are stopped and have no loading / unloading actions; it determines the operation response time by the difference between the task release timestamp and the truck start operation timestamp; it directly extracts the task urgency information synchronized by the terminal management system, and calculates the yard congestion index based on the yard occupancy rate and the busy status of loading / unloading equipment (statistically calculated through RFID data); it associates and binds the behavior features such as the non-operation dwell time and operation response time of the same truck with the business features such as the task urgency and yard congestion index of the corresponding scheduling task to construct behavior-scheduling coupling features, thereby realizing the direct mapping between truck behavior and scheduling needs.

[0092] The standardized data storage module 14, based on the feature data of the behavior-scheduling feature construction module 13, converts the original data and coupled feature data into a unified columnar storage format, stores them in a distributed database partitioned by "truck number-date", and establishes indexes for timestamp, truck number, and berth number.

[0093] Specifically, the standardized data storage module 14 adopts a "column storage + partitioned index" approach to ensure data storage efficiency and retrieval speed: the original collected data (GPS, RFID, APP, and terminal management system data) and the constructed behavior-scheduling coupled feature data are uniformly converted into a columnar storage format (such as Parquet). This format can improve the reading efficiency of feature data and adapt to the needs of subsequent algorithm calculations; the data is partitioned and stored according to the "truck number-date" rule and deployed to a distributed database (such as HBase) to meet the storage needs of massive amounts of terminal data; secondary indexes are established for timestamps, truck numbers, and berth numbers to ensure that subsequent modules can quickly retrieve target data (such as querying behavior features of a specific date by truck number and matching scheduling task data of the same period by timestamp).

[0094] The learnable incentive and behavior guidance unit 2, based on behavior-scheduling coupling characteristics, uses "non-task dwell event + task urgency matching" as trigger conditions. It optimizes the incentive coefficient through reinforcement learning to generate incentive amounts, pushes instant micro-incentives via the driver's mobile app, and automatically re-pushes them. When multiple drivers dwell simultaneously, priority is given to incentivizing those with longer dwell times. After an incentive is triggered, the driver's dwell time information is fed back to the subsequent route planning and scheduling stages in real time, allowing route planning to use this dwell time as a dynamic constraint and automatically skip it to generate a new task route. The learnable incentive and behavior guidance unit 2 includes an incentive trigger judgment module 21, a reinforcement learning optimization module 22, and an incentive push and data feedback module 23. These modules work collaboratively according to the "trigger-optimization-push-feedback" logic.

[0095] In this embodiment, the incentive trigger judgment module 21 receives the behavior-scheduling coupling feature output by the behavior data acquisition and feature coupling unit 1 as the core input for trigger condition judgment; the process of the incentive trigger judgment module 21 performing the "non-job stay event + task urgency matching" judgment includes the following steps;

[0096] S21.1 Receive the behavior-scheduling coupling features output by the behavior data acquisition and feature coupling unit 1, and synchronously acquire the continuous location data of the truck from the GPS module, the operation status data of the corresponding cargo location in the yard from the RFID reader, and the current task deadline data from the terminal management system.

[0097] Specifically, the standardized data storage module 14 of the behavior data acquisition and feature coupling unit 1 pushes behavior-scheduling coupled feature data to the incentive trigger judgment module 21 via the HTTP / 2 protocol. The data format adopts JSON format (a common format for cross-module data transmission in the industry, taking into account both readability and transmission efficiency). The fields include "truck number, non-operation dwell time (seconds), operation response time (seconds), original task urgency level, yard congestion index (0-1), data generation timestamp (UTC millisecond level)". The push frequency is consistent with the data acquisition cycle (e.g., 1Hz).

[0098] Specifically, the incentive trigger judgment module 21 synchronously acquires three types of real-time data through the data stream subscription mechanism of the edge gateway:

[0099] GPS module data is subscribed via UART serial port. The data fields are "truck number, longitude (decimal), latitude (decimal), and collection timestamp (UTC millisecond level)". The subscription frequency is 1Hz (consistent with the GPS module's collection frequency to ensure the continuity of location data).

[0100] RFID reader data is subscribed via Ethernet TCP protocol. The data fields are "cargo location number, container tag ID, loading and unloading status (0=idle, 1=loading and unloading in progress, 2=completed), reader number, and collection timestamp (UTC millisecond level)". The subscription frequency is 30 seconds / time (set according to the RFID data update frequency. High-frequency synchronization is not required for changes in loading and unloading status. 30 seconds can meet the status determination requirements and reduce network transmission pressure).

[0101] The terminal management system actively retrieves task data via an HTTP RESTful interface. The retrieval frequency is consistent with the data synchronization cycle (e.g., 30 seconds / time). The data fields are "task number, truck number, task deadline (UTC seconds), task type, berth number, and task release timestamp (UTC seconds)".

[0102] At the same time, the received multi-source data is time-series aligned using "truck number + timestamp" as the key. The allowable timestamp deviation range is ≤1 second (set according to the data transmission delay characteristics of the port's local area network, with the industry's standard deviation threshold of 1-2 seconds to ensure data consistency in the same time dimension).

[0103] S21.2 Set a position deviation threshold and a continuous sampling point number threshold. When both conditions are met simultaneously, namely "the position deviation of the GPS module's continuous preset number of sampling points is less than or equal to the position deviation threshold" and "the RFID reader reports no loading or unloading action record for the corresponding cargo location", and the duration of these conditions is greater than or equal to the preset duration threshold, it is determined to be a non-operational stop event.

[0104] Specifically, three types of judgment thresholds are set (which can be adjusted according to the dock operation scenario, and the values ​​are based on industry-standard practices and actual needs), as follows:

[0105] Position deviation threshold = 5m (set based on the technical parameter of dynamic positioning accuracy ≤ 5m of the selected industrial-grade Beidou / GPS dual-mode positioning module, matching the actual performance of the equipment).

[0106] The threshold for the number of continuous sampling points is 5 (based on a 1Hz sampling frequency, 5 sampling points correspond to 5 seconds, to avoid misjudgment due to instantaneous positioning fluctuations; the industry standard for continuous sampling is 3-5 points).

[0107] Preset duration threshold = 3 minutes (180 seconds. According to common sense of dock operations, short stops, work gaps and other normal states usually do not exceed 2 minutes. 3 minutes can effectively distinguish between non-work stops and normal work gaps).

[0108] Specifically, for five consecutive GPS sampling points of the same truck, the straight-line distance (i.e., position deviation) between adjacent points is calculated sequentially using the Haversine formula. When the position deviation of all adjacent points is ≤5m, the truck is considered to be in a stable position.

[0109] Specifically, based on the cargo location number currently associated with the truck (obtained from the task data), the RFID data of the corresponding cargo location is queried three times consecutively (covering 90 seconds to ensure no loading or unloading actions are missed). If the loading and unloading status is "0 = idle" and there is no status change, it is determined that there is no loading or unloading action at the corresponding cargo location.

[0110] Specifically, when both "stabilized truck position" and "no loading / unloading at cargo location" are met simultaneously, the stability of the conditions is continuously monitored starting from the timestamp of the first sampling point that meets the conditions.

[0111] Furthermore, if the timing duration is ≥180 seconds, it is determined to be a non-operational stop event; if any condition is not met during the period, the timing is terminated and the status is reset.

[0112] S21.3 Based on the difference between the task deadline and the current time, the urgency value is quantified into a 0-1 range using a linear mapping. The larger the value, the more urgent the task. The mapping benchmark is the longest task cycle preset by the dock.

[0113] Specifically, firstly, obtain the maximum preset task cycle of the dock. (Example value 8 hours = 28800 seconds. The value is based on the characteristics of common task types at the terminal. The normal operation cycle of core tasks such as import unloading and in-yard transfer is 4-8 hours. The middle value is taken to take into account the universality of different task types. It can be configured separately according to task type.)

[0114] Then, the time difference is calculated. :

[0115] Extract the difference between the task deadline and the current system time. ,like ,but (The task has timed out; proceed with the highest urgency level.) ,but (Sufficient tasks, no incentive required);

[0116] Next, the urgency is quantified using a linear mapping formula:

[0117] ;

[0118] in The time difference is in the second range, and the calculation result is rounded to two decimal places (the standard precision for industry data quantification) to ensure that the value is in the range of 0-1.

[0119] S21.4 When the non-work stoppage event is determined to be true, and the task urgency value is... When the task is in the preset medium urgency range, the trigger condition is determined to be met, and the trigger signal, non-task dwell time, and task urgency value are output to the reinforcement learning optimization module 22. And the yard congestion index.

[0120] Specifically, the operations for trigger condition matching and data output are as follows:

[0121] First, the preset medium urgency range is 0.3-0.7 (the value is based on the scheduling incentive strategy logic: when the urgency is <0.3, there are enough tasks and no incentive is needed; when the urgency is >0.7, the task is urgent and requires forced scheduling rather than micro-incentives; the 0.3-0.7 range is suitable for guiding drivers to respond through incentives. The industry standard range for similar incentive scenarios is 0.2-0.8, which can be adjusted as needed).

[0122] Subsequently, a dual condition determination was made: the non-operational stay event was established and If the condition is met, a trigger signal (True) is generated; otherwise, the signal is set to False and the process is terminated.

[0123] Next, extract the output data: non-task dwell time. (Current timeout duration, in seconds), Task urgency Yard congestion index (Extracted from coupling features);

[0124] Finally, the data is encapsulated into a JSON format data frame (a common format for cross-module transmission) and transmitted to reinforcement learning optimization module 22 via TCP protocol. The transmission timeout is 5 seconds (the normal timeout threshold for local area network data transmission is 3-5 seconds to ensure timely delivery of data). If the timeout occurs, the data will be retransmitted once.

[0125] In this embodiment, the reinforcement learning optimization module 22 receives the trigger event data output by the stimulus trigger judgment module 21, trains the policy network model using the PPO algorithm, and outputs the stimulus coefficients. The process of the reinforcement learning optimization module 22 receiving the trigger event data output by the stimulus trigger judgment module 21, training the policy network model using the PPO algorithm, and outputting the stimulus coefficients includes the following steps.

[0126] S22.1 Construct a policy network model for the PPO agent. The network uses non-task dwell time and task urgency values ​​from the trigger event data. The yard congestion index is the input feature, and the output is the incentive coefficient corresponding to the dwell time weight. Incentive coefficients corresponding to the weights of the congestion index The network has multiple fully connected layers, and nonlinear feature mapping is achieved through ReLU and tanh activation functions;

[0127] Specifically, the construction of the policy network model involves the following steps:

[0128] First, define the network structure parameters (the values ​​are based on the standard configuration for reinforcement learning with few samples, and are adapted to the daily increase in samples at the port), as follows:

[0129] Input layer: Number of neurons = 3 (corresponding to 3 input features), input features need to be normalized to the 0-1 range by Min-Max (a common method for feature preprocessing in the industry), maximum dwell time = 1800 seconds (i.e. 30 minutes, based on the characteristics of the longest non-operational dwell time at the dock, exceeding 30 minutes is usually due to equipment failure or abnormal situation, which needs to be handled separately).

[0130] Hidden layers: 2 fully connected layers with 64 and 32 neurons respectively (a common neuron count configuration for small-dimensional feature mapping in the industry; decreasing from 64 to 32 aligns with feature dimensionality reduction logic and avoids overfitting); the input layer and the first hidden layer use the ReLU activation function (formula: ,in This represents the weighted sum of the inputs to the current neuron (i.e., the output of the previous layer's neuron, multiplied by the corresponding weight of the current layer, and then added to the bias term of the current neuron). This indicates that the ReLU function is applied to the input. Processed output value (if) Output itself; if (output 0), a common activation function to solve the gradient vanishing problem), the first and second hidden layers use the tanh activation function (formula: ,in It is a natural constant in mathematics (a fixed value, approximately equal to 2.71828), which enhances the nonlinear mapping capability of features.

[0131] Output layer: Number of neurons = 2 (corresponding to 2 activation coefficients), output value is constrained to the range of 0.1-0.9 by the Sigmoid function (to avoid the activation amount being too low due to the coefficient being too small, or the activation cost being too high and thus out of control).

[0132] Subsequently, the network was initialized: the weight parameters were initialized using Xavier (a general initialization method adapted to fully connected layers to ensure consistent output variance across layers), and the bias parameters were initialized to 0.1 (to avoid initial output of 0); the network framework used was PyTorch 1.10 or later (the mainstream deep learning framework in the industry, supporting GPU acceleration and adapting to the massive data training needs of the platform).

[0133] Finally, define the input / output interface: the input is a 3D vector [normalized, paused, ...] , The output is a 2D vector. The data type is float32 (a common data type for deep learning models).

[0134] S22.2 Define a two-dimensional reward signal: the core reward is negatively correlated with the driver's response time, the constraint reward is negatively correlated with the incentive amount, and the total reward signal is the weighted sum of the core reward and the constraint reward.

[0135] Specifically, the definition and operation of the two-dimensional reward signal are as follows:

[0136] First, define the core rewards. (Time taken for driver response) negative correlation The time difference between push notification and confirmation (in seconds) is calculated using the following formula:

[0137] ;

[0138] in Seconds (10 minutes, set within a reasonable timeframe for driver response decisions). If there is no response within seconds 0; The value is set to 0-10 (the normal quantization range for reinforcement learning reward signals, to avoid excessively large reward values ​​that could cause model oscillations).

[0139] Subsequently, define the constraint reward. (and incentive amount) (Negative correlation), calculation formula:

[0140] ;

[0141] in Yuan (set based on the upper limit of port incentive cost control). Yuanze 0; Values ​​range from 0 to 5 (weighted in accordance with core rewards, balancing efficiency and cost);

[0142] Finally, calculate the total reward signal. :

[0143] ;

[0144] The weighting coefficients of 0.7 and 0.3 are set based on the dock scheduling requirements of "efficiency first, cost controllable". The conventional weighting ratio for similar multi-objective optimization scenarios in the industry is 7:3 or 8:2, which can be adjusted.

[0145] S22.3 Daily collection of "Trigger Condition Data - Incentive Coefficient" , The triplet of "driver response result" is used as a training sample and input into the policy network for incremental training. The parameter update range is limited by the parameter pruning mechanism of the PPO algorithm.

[0146] Specifically, the specific operations for training sample collection and incremental training in this embodiment are as follows:

[0147] First, sample collection: Collect the previous day's "trigger condition data - incentive coefficient" between 00:00 and 00:30 every day. , - Driver response result triplet sample (the terminal workload is low during this period and does not affect the system operation), the sample fields cover the core input and output data, the storage format is CSV (the industry sample storage format is common and easy to preprocess), and it is partitioned by "date-truck number" and stored in a distributed database (such as HBase, which is suitable for massive time series data storage needs).

[0148] Subsequently, sample preprocessing: removal of outlier samples ( The training set and the validation set are divided into the training set and the validation set in a 7:3 ratio (a standard division ratio for industry model training to ensure training effectiveness and generalization ability).

[0149] Next, incremental training configuration:

[0150] The Adam optimizer (a commonly used optimizer in deep learning, known for its fast convergence speed) is used. The learning rate is 0.001 (a standard value for the Adam optimizer, avoiding oscillations caused by an excessively high learning rate and slow convergence caused by an excessively low learning rate). The number of iterations is 100 epochs (a standard number of iterations suitable for the daily sample size). The batch size is 32 (a standard configuration that balances training speed and memory usage).

[0151] Enable the PPO algorithm parameter pruning mechanism to constrain the probability ratio of the new and old strategies within the range of [0.8, 1.2] (standard configuration of the PPO algorithm to avoid excessive parameter update amplitude);

[0152] Finally, model update: The validation set loss is calculated using the MSE loss function. If the loss value is ≤0.5, the old model is replaced (this value is based on the conventional threshold for judging the model's convergence effect; a loss value <0.5 indicates that the model has a good fit). Otherwise, the old model is retained and the samples are accumulated for retraining the next day.

[0153] S22.4, Call the optimal activation coefficients output by the trained network. and The real-time incentive amount is calculated by combining the non-operational dwell time and yard congestion index in the trigger event data. .

[0154] Specifically, the calculation of real-time incentive amounts is performed as follows:

[0155] First, the benchmark incentive amount (Based on the standard cost of the terminal's micro-incentive program, the benchmark amount for similar immediate incentives in the industry is 5-20 yuan, and 10 yuan balances attractiveness and cost controllability).

[0156] Then, the optimal coefficients of the trained model are called. and (Keep two decimal places, standard numerical precision requirements);

[0157] Next, calculate the incentive amount. :

[0158] ;

[0159] in Convert to minute-level (to avoid excessively large second-level values) abnormal);

[0160] Finally, the rounding rule: keep one decimal place. The amount is calculated at 5 yuan (the minimum incentive threshold to ensure drivers' willingness to respond). The cost is calculated at 200 yuan (maximum cost).

[0161] In this embodiment, the incentive push and data feedback module 23 generates specific incentive instructions based on the incentive coefficients output by the reinforcement learning optimization module 22 and pushes them; and synchronizes the incentive execution results to the path planning and scheduling unit 4 and the behavior data acquisition and feature coupling unit 1, forming a closed loop of data flow. The process of the incentive push and data feedback module 23 executing the incentive instruction push and data synchronization includes the following steps;

[0162] S23.1 When multiple drivers simultaneously trigger non-operational stop events, the priority score for each driver is calculated based on the stop duration and yard congestion index output by the incentive trigger judgment module 21. , Positively correlated with dwell time and congestion index, according to Incentive resources are allocated and pushed in descending order.

[0163] Specifically, the priority score calculation and resource allocation operations are as follows:

[0164] First, the statistical baseline data:

[0165] Average non-operational stay time at the port in the past 7 days (Second-level, dynamically adapting to fluctuations in terminal operations; industry statistical cycles typically range from 7 to 14 days), maximum threshold for yard congestion index. (Default maximum value after normalization);

[0166] Subsequently, the priority scoring formula is as follows:

[0167] ;

[0168] The calculated score, based on the above formula, is rounded to two decimal places. The weights of 0.6 and 0.4 are set according to the scheduling logic that "stay duration has a greater impact," which is the standard weight allocation for multi-factor priority calculation in the industry.

[0169] Next, resource allocation: when N drivers trigger simultaneously, press... Sort in descending order and push to the top first. name( =10), set according to the terminal's concurrent processing capacity; a regular terminal can have no more than 10 non-operational vehicles parked at the same time.

[0170] Finally, drivers who are not selected are added to the waiting queue, and their priority is recalculated every 30 seconds (matching data update cycle), and they are pushed to other drivers in order of priority.

[0171] S23.2, Push incentive amounts to target drivers via driver-side app The instruction is sent, and a response timer is started. If no "location movement confirmation" signal is received from the app within the first preset response time, a second push is triggered. If there is still no response after the second push and a second preset response time, the push is terminated, and the incentive amount for the second push is adjusted to [amount missing]. , for The product of the preset incentive adjustment coefficient;

[0172] Specifically, the operation of incentive command push and automatic re-push is as follows:

[0173] First, the content pushed to the target driver is as follows: The APP pop-up notification format is "[Dispatch Incentive] Your vehicle (truck number: Smart Terminal) is currently in a non-operational stop state. Smart Terminal can receive an incentive amount upon completion of the task." "Click to confirm and go now!" The pop-up window lasts for 10 seconds;

[0174] Next, configure the response timer: the first preset response duration. Minutes (180 seconds, the driver's decision time after seeing the pop-up), second preset response time Minutes (300 seconds, secondary waiting time), preset incentive adjustment coefficient (Secondary incentives typically increase by 10%-30%, 1.2 balancing attractiveness and cost).

[0175] Next, the specific push notification process is as follows:

[0176] First push: Send the command and record the timestamp, then start a 180-second timer;

[0177] Response judgment: If a "location movement confirmation" signal is received (judgment criteria: the position deviation of 3 consecutive GPS sampling points is >5m and they are facing the target berth, matching the actual state of non-operational stop cancellation), then record "responded";

[0178] Second push: If there is no response within 180 seconds, it will be calculated. (Limited to 5-200 yuan), send a second notification and start a 300-second timer;

[0179] Terminate push notifications: Record "No Response" if there is no response for 300 seconds and terminate push notifications.

[0180] Finally, the push logs are stored locally and synchronized to the behavior data collection and feature coupling unit 1 (ensuring data traceability).

[0181] S23.3 After the excitation is triggered, the start and end timestamps of the dwell time period marked in the GPS data are extracted, encapsulated into a standardized data frame, and pushed to the path planning and scheduling unit 4 in real time via wireless communication;

[0182] Specifically, the steps for providing feedback on the duration of stay are as follows:

[0183] First, extract the timestamp:

[0184] Start and end timestamps : That is, the sampling point timestamp (UTC millisecond level) when the condition in S21.2 is first met;

[0185] End of stay timestamp A response indicates a confirmation of the sampling point timestamp; no response indicates the termination timestamp of the second push.

[0186] Subsequently, JSON format (a universal format for cross-unit transmission) was adopted, with fields including "truck number, task number, ... , "Incentive trigger status, yard number", data frame ≤ 1KB (to avoid transmission delay);

[0187] Next, configure the communication: use the MQTT protocol, the Broker address is the edge gateway IP, the port is 1883 (the default MQTT port), and the QoS level is 1 (to ensure that messages are delivered at least once, to meet the reliability requirements of scheduling data).

[0188] Finally, send immediately. If no receipt is received within 5 seconds, resend once (the standard retransmission mechanism for LAN transmission). If it fails, record the exception in the log and retransmit later.

[0189] S23.4. The driver's response status and the actual dwell time change data after the response are synchronously fed back to the standardized data storage module of the behavior data acquisition and feature coupling unit 1.

[0190] Specifically, the data feedback to the standardized data storage module is performed as follows:

[0191] First, organize the feedback data:

[0192] Response status: Enumerated values ​​(0 = no response, 1 = first response, 2 = second response, industry-standard data enumeration method).

[0193] Changes in dwell time:

[0194] (seconds) ;

[0195] Incentive execution data: (If applicable), push count, timestamp; subsequently, data standardization: encapsulated in JSON format, with field names consistent with the fields stored in the behavioral data collection and feature coupling unit 1 (to ensure compatibility).

[0196] Next, data transmission: The data is sent to the interface of Behavior Data Acquisition and Feature Coupling Unit 1 via an HTTP POST request. The request header carries an authentication token (which is consistent with the authentication method of the interface of Behavior Data Acquisition and Feature Coupling Unit 1 to ensure data security). The timeout period is 10 seconds (the normal timeout threshold for cross-unit transmission).

[0197] Finally, the success is determined by the feedback: if a "200 OK" response is received, it is successful; otherwise, it will retry 3 times (the standard retry mechanism). If it fails, it will be logged for manual verification.

[0198] The scheduling demand prediction unit 3 is based on standardized time series data. It uses an LSTM model to learn the correlation between historical dwell time and working conditions, predicts the possible dwell time of each truck in a specified future period, and optimizes the model output by supplementing it with the latest real-time data, providing a forward-looking constraint basis for scheduling.

[0199] In this embodiment, the scheduling demand prediction unit 3 includes a data input and preprocessing module 31, an LSTM model processing module 32, and a real-time output optimization module 33. Each module works collaboratively according to the logic of "data preprocessing - model training and inference - real-time optimization," wherein:

[0200] The data input and preprocessing module 31 receives the standardized time-series data output by the behavior data acquisition and feature coupling unit 1, and performs format regularization and feature extraction on the historical dwell time period and working condition information in the data to form the basic data for the input of the LSTM model.

[0201] Specifically, the core of the data input and preprocessing module 31 is to transform the raw time series data into standardized data that can be directly input into the LSTM model, ensuring data quality and format compatibility. The specific operations are as follows:

[0202] First, the system receives standardized time-series data output from the behavioral data acquisition and feature coupling unit 1. This data is transmitted using the HTTP RESTful protocol with end-to-end encryption to ensure data security and integrity. The data format is JSONLines, UTF-8 encoded, and stored in aggregates of one minute per record, covering nearly six months of historical data (meeting the model's sample size requirements for learning historical patterns). Core fields include static metadata such as truck number, load class, cargo location type, and task type, as well as dynamic time-series data such as collection timestamp, dwell time, real-time location, equipment operating status, congestion index, and weather data.

[0203] Subsequently, standardized preprocessing was performed on the received data: abnormal data (such as unusual dwell time and invalid congestion index) were identified using the 3σ principle, and after verification in conjunction with industry operation logic, the data was replaced with the median of the same type of data in the same period; sporadic missing data were filled with linear interpolation, and consecutive missing data of ≤3 were filled with forward filling, while those exceeding 3 were marked as invalid time periods and removed; the time deviation of multi-source data was corrected based on the collection timestamp to ensure time sequence consistency.

[0204] Next, feature engineering was carried out to meet the requirements of LSTM modeling: discrete features were encoded using Embedding (a conventional dimension in this field), and ordered features were mapped to the 0-1 interval through min-max normalization; statistical features such as average dwell time, longest dwell time, and distribution entropy were extracted based on a 7-day sliding window, while the weights of working condition-related features and environmental impacts were calculated. The working condition-related features are actual terminal operation parameters, including real-time crane operation efficiency, number of truck queues, and real-time occupancy rate of cargo spaces; the environmental impact features are actual port monitoring data, including real-time wind speed and rainfall level; the weights were determined based on the feature importance analysis of the terminal's historical scheduling data and dynamically adjusted according to real-time working conditions (such as increased rainfall and truck queue congestion); and the variance thresholding method was used to remove invalid features and retain core features for model input.

[0205] Finally, the data is grouped by set card number, and a 3D input tensor (batch_size, time_step, input_size) is constructed. Here, time_step is the number of time steps corresponding to 7 days, input_size is the number of core features after filtering, and batch_size is set to 32-64 (the common range for model training in this field). The training set, validation set, and test set are divided by stratified sampling in a 7:2:1 ratio to ensure data distribution consistency. The preprocessed data is stored in a distributed file system with "set card number + date" as the partition key, and is also cached in an in-memory database for fast reading by the model.

[0206] Based on the basic data output by the data input and preprocessing module 31, the LSTM model processing module 32 uses the LSTM model to learn the correlation between historical dwell time and working conditions, and infers the possible dwell time of each truck in the future specified period.

[0207] Specifically, the LSTM model processing module 32 is used to learn patterns in historical data and infer future dwell times. Incremental training ensures the timeliness of the model. The specific operation is as follows:

[0208] First, an LSTM policy network model adapted for time series prediction is constructed: the input layer dimension is consistent with the number of core features after preprocessing; the hidden layer consists of two fully connected layers with 32-128 neurons (a common value for time series prediction in this field, balancing model expressive power and computational cost), and the dropout coefficient is set to 0.1-0.3 (to suppress overfitting); the input layer and the first hidden layer use the ReLU activation function, and the first layer and the second hidden layer use the tanh activation function to achieve non-linear feature mapping; the output layer corresponds to the relevant prediction parameters for future dwell time periods, and the network framework uses mainstream deep learning frameworks such as TensorFlow or PyTorch.

[0209] Next, configure the core parameters for model training: use the Adam optimizer, set the learning rate to around 0.001 (a common value for deep learning to avoid gradient oscillation); select mean squared error (MSE) as the loss function, and adapt it to the prediction error calculation of continuous values ​​such as dwell time; enable the early stopping mechanism during training (stop if the validation set loss does not decrease for 10 consecutive rounds), and use L2 regularization to suppress overfitting. The batch size is adapted to hardware resources, and the number of iterations is set to 100-200 epochs.

[0210] Next, the model training process of "offline initialization + incremental update" is executed: 6 months of historical preprocessed data are loaded to complete offline initialization training. The initialization parameters adopt a uniform distribution of Xavier. Gradient clipping is enabled during training to prevent gradient explosion. The baseline model with the optimal loss on the validation set is saved. During the low-work hours of each day, the new data from the previous day is loaded to perform incremental training on the baseline model (about 50 epochs). After the update, it is validated through the test set. The mean absolute percentage error (MAPE) is required to meet the industrial prediction accuracy requirements. If the requirements are not met, it is rolled back to the previous effective model.

[0211] Finally, inference is performed based on the trained model: with the current time as the endpoint, the "working condition-stay" correlation pattern is learned by sliding a 7-day historical window to predict the possible stay time of each truck in the next 24 hours; each truck outputs 3-5 high confidence (≥85%) stay time periods, each time period includes the stay start time (UTC seconds), stay end time (UTC seconds), confidence level and main influencing factors. The data is pushed to the real-time output optimization module 33 in JSON format and stored in the time series database for 30 days.

[0212] The real-time output optimization module 33 receives the latest real-time operating condition data and supplements the output of the LSTM model processing module 32 with the real-time operating condition data. It dynamically optimizes the prediction results of possible dwell time periods to form the forward-looking constraint basis required for scheduling.

[0213] Specifically, the real-time output optimization module 33 is used to integrate real-time operating data to correct prediction deviations, ensuring the accuracy and timeliness of the output results. The specific operation is as follows:

[0214] First, it receives real-time operating data from multiple sources and transmits it via a Kafka message queue (end-to-end latency is controlled within the industrial real-time response threshold). Data sources include real-time GPS location of container trucks (updated at 1Hz), RFID cargo loading and unloading status (every 30 seconds), equipment operating status (every 10 seconds), port area weather data (every 5 minutes), and TMS task change notifications (real-time push). The data is parsed according to a preset format and key indicators are extracted. The integrity and timeliness of the data are verified, and failed data is stored in a dead-letter queue for subsequent investigation.

[0215] Then, a hybrid algorithm of "Kalman filtering + sliding window update" (a common solution for industrial real-time forecasting) is used to optimize the forecast results, dynamically allocating the weights of historical model results and real-time data (historical results weight 0.6, real-time data weight 0.4). The optimization triggering mechanism is divided into active and passive: real-time data is synchronized every 30 seconds to optimize the forecast period before it expires; when key operating conditions change abruptly, such as equipment failure, a sudden increase in congestion index, or severe weather, the single-unit truck forecast result optimization is triggered immediately.

[0216] Next, the specific correction logic is executed: when the truck deviates from the predicted path (distance > 10m, based on industrial-grade GPS positioning accuracy ≤ 5m, reserving 2 times redundancy to avoid misjudgment of positioning errors, conforming to the academic theory of spatial error tolerance, which is the conventional threshold for port path monitoring), the dwell start time is extended by 1.2 times the deviation duration; when the cargo space is suddenly occupied, the dwell time is extended by the average release time of the terminal cargo space (about 30 minutes); when the real-time operating conditions match the historical patterns ≥ 90% (the core threshold of "high consistency" in the field of time series prediction, 90% or more represents a high degree of scenario overlap), the confidence level is increased by 5%-10% (the conventional range of confidence level correction for industrial prediction, a gentle positive adjustment to reflect reliability); when the matching degree is < 60% (the key dividing point for identifying high deviation risk in industrial scenarios), the confidence level is reduced by 10%-20% (the conventional range of confidence level reduction for high-risk scenarios, conveying deviation warnings); the period below 60% is marked as "low confidence" and supplemented with alternative dwell time periods (the core means of fault-tolerant design in industrial prediction, adapting to scheduling redundancy requirements).

[0217] Finally, the optimized results are grouped by "truck number - task ID", including dwell time (start time, end time), confidence level (≥70%), risk warning and scheduling suggestions. After being serialized in ProtoBuf format, they are pushed to the path planning and scheduling unit 4. If the push fails, it will automatically retry 3 times. All optimization processes (trigger conditions, correction logic, working condition data) are recorded in the audit log and retained for 180 days to support model iteration and problem traceability. The mean absolute error (MAE) of the optimized prediction results is ≤30 seconds, which meets the scheduling decision requirements.

[0218] The route planning and scheduling unit 4 receives the dwell time prediction results from the scheduling demand prediction unit 3, the real-time dwell events and driver response status from the learnable incentive and behavior guidance unit 2, and the multi-source operation data from the behavior data collection and feature coupling unit 1. It generates path constraints through a structural causal model, uses an improved A* algorithm to complete truck-berth matching and dynamic route planning, and combines resource pre-occupancy and elastic release mechanisms to synchronously feed back scheduling data to the scheduling instruction execution and dynamic adjustment unit 5 and the learnable incentive and behavior guidance unit 2.

[0219] In this embodiment, the route planning and scheduling unit 4 includes a data integration and preprocessing module 41, a route constraint generation module 42, a dynamic route planning and berth matching module 43, a resource management module 44, and a data feedback module 45. Each module works collaboratively according to the logic of "data integration - constraint generation - route planning - resource scheduling - data feedback," wherein:

[0220] The data integration and preprocessing module 41 synchronously accesses the future stay time prediction results output by the scheduling demand prediction unit 3, the real-time stay event and driver response status data output by the learnable incentive and behavior guidance unit 2, and the multi-source operation data output by the behavior data collection and feature coupling unit 1. It performs format standardization conversion and timestamp alignment on the heterogeneous data, removes expired data that exceeds the effective time window, and forms a unified scheduling input dataset.

[0221] Specifically, the data integration and preprocessing module 41 is used to form a unified scheduling input dataset, and the specific operations are as follows:

[0222] First, clarify the core configuration for multi-source data access:

[0223] Data sources include: the future dwell time prediction results output by the scheduling demand prediction unit 3, the real-time dwell events and driver response status output by the learnable incentive and behavior guidance unit 2, and the multi-source operation data (including truck location, working condition, task information, etc.) output by the behavior data acquisition and feature coupling unit 1.

[0224] Transmission protocols: Commonly used industrial protocols such as TCP, HTTPRESTful, and MQTT are employed to complete cross-unit data transmission and ensure data transmission stability;

[0225] Access frequency: Real-time status data is synchronized at 1Hz, and predictive and statistical data is updated every 30 seconds to match the timeliness requirements of scheduling decisions.

[0226] Subsequently, data standardization processing was performed:

[0227] Unified Field Standardization: All key fields of the data are named uniformly according to "cargo number, task number, timestamp", the character encoding is uniformly UTF-8, and the time format is uniformly UTC seconds.

[0228] Heterogeneous data conversion: Enumerated types are converted to integer encoding, latitude and longitude are retained to 6 decimal places, and floating-point data (congestion index, urgency) are retained to 2 decimal places to ensure data format consistency.

[0229] Next, complete the timestamp alignment and data filtering:

[0230] Timestamp alignment: Based on the timestamp of the running data of the behavior data acquisition and feature coupling unit 1, the time deviation of other data sources is corrected. When the deviation is ≤3 seconds, it is completed by interpolation. Data exceeding the deviation is marked as abnormal.

[0231] Valid time window setting: retain data within the range of "current time - 30 minutes" to "current time + 24 hours", and remove expired and outdated invalid data;

[0232] Data validation: Validate the integrity and logical rationality of the data (e.g., start time of stay < end time of stay, congestion index ∈ [0,1]). Failed data is stored in the exception table and an alarm is triggered.

[0233] Finally, a unified scheduling input dataset is output: structured data frames are constructed by grouping according to the "card-task" dimension and stored in a distributed database and memory cache commonly used in industry, supporting fast reading by subsequent modules.

[0234] The path constraint generation module 42, based on the standardized dataset of the data integration and preprocessing module 41, constructs a multi-dimensional path constraint system through a structural causal model: it uses the predicted dwell time and real-time dwell time as "time constraints" to clarify the unschedulable periods for trucks; it maps the yard congestion index as "spatial constraints" to define the path passage priority in high-congestion areas; it uses the berth operation status and cargo space occupancy information as "resource constraints" to limit the range of berths that trucks can match; and it uses the task urgency as "priority constraints" to set the execution order weight of scheduling tasks.

[0235] Specifically, the core of the path constraint generation module 42 is to quantify multi-dimensional constraint relationships through the structural causal model (SCM) to construct a computable and implementable path constraint system. The specific operations are as follows:

[0236] First, clarify the variable definitions and causal relationships in the structural causal model, specifically including:

[0237] Core variable set:

[0238] Endogenous variable (output to be constrained): Path access priority Task execution order weight Matchable berth set Unschedulable periods ;

[0239] Exogenous variables (input features): Predicted duration of stay Real-time stay period Yard congestion index Berth operation status Storage space occupancy status urgency of the task ;

[0240] Causal relationship construction: Variable associations are defined through a directed acyclic graph (DAG), which conforms to the direct causal logic of "condition-constraint". Those skilled in the art can implement this based on conventional causal modeling methods.

[0241] Subsequently, the constraints of each dimension were quantified, specifically including:

[0242] Time constraints (unschedulable periods) ):

[0243] The time constraints clearly define the periods during which trucks cannot perform scheduled tasks, and are generated by fusing the predicted dwell time period with the real-time dwell time period:

[0244] Real-time dwell time period: If a real-time dwell event has been triggered, the real-time unschedulable period is... ;in This is the real-time start time of the stay. This refers to the real-time end time of your stay.

[0245] Predicted stay duration: If the confidence level of the predicted stay is ≥70%, the predicted unschedulable stay duration is... ;in To predict the start time of the stay, To predict the end time of stay, =300 seconds =180 seconds is the buffer time, which is set according to the start-stop time of the truck;

[0246] Total unschedulable period: (Take the union of two time periods; overlapping time periods are automatically merged).

[0247] Spatial constraints (path priority) ):

[0248] Spatial constraints map the yard congestion index to path priority, guiding the algorithm to avoid highly congested areas:

[0249] Congestion level classification: based on congestion index Divided into 3 levels, For low congestion, For the congestion, High congestion;

[0250] Priority quantification formula: ( =1.2 is the congestion impact coefficient, which is set based on the statistical impact of port congestion on traffic efficiency.

[0251] Priority rules: High priority (preferred selection) Medium priority (second choice) Low priority (avoidance).

[0252] Resource constraints (matchable berth set) ):

[0253] Resource constraints limit the range of berths that can be matched with trucks, and filtering is based on berth operation status and cargo occupancy information:

[0254] Berth availability rating formula:

[0255] ;

[0256] in The weights are set based on the priority of the impact of berth status and cargo space occupancy.

[0257] In "resource constraints (matchable berth set)" In the quantification logic of "), the berth status normalized value Normalized value of warehouse space occupancy The mapping rules are set based on a general logic that combines the actual operational scenarios of the dock with resource matching, as follows:

[0258] Berth status normalized value (The berth operation status) (Converted to a quantitative score)

[0259] when When (berth status: idle), This assignment corresponds to the high matching value of idle berths—these berths can directly undertake container truck operations, are the most efficient available resources, and are also the common practice in industrial scenarios where "high availability resources correspond to high quantitative scores."

[0260] when When berth status: busy, Based on operational data from similar terminals, the operation connection time for busy berths is about three times that of idle berths, indicating a significant risk of delay. The value of 0.3 reflects both its limited availability and the low matching priority of this type of resource.

[0261] when When (berth status: fault), Faulty berths cannot provide operational support and are invalid options in resource matching. Assigning a value of 0.0 is a common practice in the industry to exclude invalid resources and prevent faulty berths from being included in the matching range.

[0262] Normalized value of warehouse space occupancy (Convert the target storage space vacancy rate into a quantitative score):

[0263] When the target storage location vacancy rate is ≥80%, Based on the actual operation data of the terminal cargo berths over the past three months, cargo berths with an idle rate of ≥80% can support trucks to quickly complete berthing and operation (usually taking ≤5 minutes), which is a typical scenario for efficient cargo berth matching. Therefore, a value of 1.0 is assigned to reflect its high availability.

[0264] When 50% ≤ target storage location vacancy rate < 80%, The cargo space in this area is partially occupied, and the efficiency of truck operation is about 50% of that of high-idleness cargo space. The value of 0.5 is just right to match its actual availability.

[0265] When the target storage location vacancy rate is less than 50% Cargo locations with an idle rate of less than 50% are experiencing severe congestion, preventing trucks from quickly docking and completing operations. This represents an inefficient scenario in cargo location resource matching, and assigning a value of 0.0 can exclude such inefficient cargo locations.

[0266] The specific filtering rules are as follows: The berths are included in the matchable set. Otherwise, exclude it.

[0267] Priority constraints (task execution order weights) ):

[0268] Priority constraints set the task execution order, determined by task urgency. Quantification:

[0269] Weight calculation formula: ( (Mapped to weight values ​​from 0 to 10 for easy sorting).

[0270] Sort by: Sort in descending order. The larger the value, the higher the task execution priority, and the more resources are allocated during path planning.

[0271] Finally, output a multi-dimensional constraint system: The constraint matrix is ​​encapsulated and transmitted to the dynamic path planning and berth matching module 43. The constraint update frequency is consistent with the data integration frequency.

[0272] The dynamic route planning and berth matching module 43 uses the multi-dimensional constraints of the route constraint generation module 42 as boundary conditions and adopts an improved A* algorithm with integrated dynamic weights to realize truck-berth matching and route planning. The heuristic function design incorporates the task urgency coefficient and the dynamic weight of yard congestion. The higher the urgency and the lower the congestion index, the greater the weight of the route. During the route search process, it avoids the dwell time period limited by "time constraint" and the high congestion area limited by "space constraint" in real time, and simultaneously verifies the availability of berth "resource constraint". It outputs the optimal truck scheduling route and berth matching scheme that takes into account both the shortest route and the timeliness of the task.

[0273] Specifically, the core of the dynamic path planning and berth matching module 43 is an improved A* algorithm that integrates dynamic weights. Conventional A* algorithms have shortcomings in port truck scheduling, such as "focusing only on the shortest path and ignoring task priority and real-time operating conditions" and "separation of berth matching and path planning easily leading to resource conflicts." The dynamic path planning and berth matching module 43 adopts an improved A* algorithm that integrates dynamic constraints and unified decision-making to achieve optimal matching and path planning between trucks and berths. The specific implementation is as follows:

[0274] First, the heuristic function combines "task urgency + dynamic congestion weights" to balance timeliness and traffic efficiency:

[0275] The conventional A* heuristic function, based solely on geometric distance, cannot adapt to the varying task priorities in port scheduling. The algorithm designs the heuristic function as follows:

[0276] ;

[0277] in, Represents a node Arrive at the target berth The heuristic function estimate (unitless) is used to guide the path search direction and balance path length with scheduling requirements; , for nodes Arrive at the target berth Euclidean distance (basic path length reference); (Mission urgency) (Dynamic congestion weighting) satisfies — By dynamically allocating weights, we can achieve scenario adaptation by prioritizing timeliness for urgent tasks and avoiding congestion for less urgent tasks; Weights for task execution order. For nodes Path access priority ( The node congestion index is obtained based on real-time traffic flow statistics in the area where the node is located. The congestion impact coefficient is set based on port congestion statistics—for every 0.1 increase in the congestion index, traffic efficiency decreases by an average of 12%. This coefficient can accurately quantify the impact of congestion on route priority.

[0278] The above design solves the problem that conventional A* cannot distinguish task priorities, making path planning more in line with actual scheduling needs.

[0279] Secondly, the actual cost introduces a "constraint penalty term" to achieve a hard limit on dynamic constraints:

[0280] To prevent trucks from planning routes that fall into unschedulable periods or high-congestion areas, a constraint penalty term is added to the actual cost:

[0281] ;

[0282] in, ; From the starting point to the node The cumulative path length; (Meets time / space constraints) or (Violation of constraints, such as entering unschedulable periods or low-priority nodes); 0.8 represents the impact coefficient of congestion on path costs. According to statistics from similar port operation data, when the congestion index is 0.5, the transit time for container trucks increases by 40% compared to smoother traffic. This coefficient can accurately map the impact of congestion on actual transit costs.

[0283] This design allows the algorithm to automatically avoid invalid paths during the search process, ensuring the feasibility of the planning results.

[0284] Then, berth matching and route planning are integrated to reduce resource conflicts:

[0285] The traditional process of "matching berths first, then planning routes" is prone to the problem of "the route to the matched berth is not feasible". The algorithm links berth matching scores with route feasibility:

[0286] ;

[0287] in, Indicates the container truck and the target berth. Match rating (unitless), the higher the rating, the better the fit; This indicates the weight allocation ratio, set according to the port scheduling decision logic—task priority ( The factor that has the greatest impact on scheduling rationality (weight 0.5) is berth availability, followed by berth availability. (Weight 0.3), and finally path feasibility ( (with a weight of 0.2), which conforms to the scheduling priority order of "task first, resource adaptation, and path reachability"; This indicates the berth availability score (no unit). The calculation formula is: , The setting is based on the statistical law that the impact of berth status on operations is greater than that of cargo space occupancy; Prioritize the average route access from trucks to berths—simultaneously assess route feasibility when matching berths, and only select berths with "high scores and route accessibility" as targets, significantly reducing the risk of conflict in subsequent route planning.

[0288] Finally, real-time constraint dynamic verification adapts to changes in operating conditions:

[0289] During the path search process, each expanded node is checked to see if it meets the "time constraints (unschedulable periods)" and "spatial constraints (high congestion areas)"; at the same time, the latest operating data is synchronized every 30 seconds to dynamically update the constraints—solving the problem that static planning cannot adapt to real-time operating conditions (such as sudden congestion or equipment failure).

[0290] The specific execution flow of the algorithm is as follows:

[0291] Initialization: Set the current position of the truck as the starting point. The target berth selected based on the integrated matching score is set as the endpoint. Initialize the "open list (nodes to be explored)" and the "closed list (nodes already explored)";

[0292] Node expansion and constraint verification: Select from the open list ( For nodes The node with the smallest total cost Expand its neighboring nodes and verify constraints:

[0293] If the node is in an unschedulable period or a low-priority area ( Add penalty items ;

[0294] Calculate the nodes according to the formula. Update the open list (if the node is already in the open list and the new one is...). Smaller, therefore update cost);

[0295] Termination and Path Extraction: When the target node When adding to the close list, from Backtracking The optimal route is obtained; the target berth ID, route node sequence, and estimated arrival time are output (calculation formula: estimated arrival time = total route length / average truck speed, where the average truck speed = 5km / h, set according to the speed limit standard in the terminal) are obtained.

[0296] Based on the scheduling scheme output by the dynamic path planning and berth matching module 43, the resource management module 44 initiates a resource pre-occupancy mechanism: after the truck scheduling path is generated, it automatically pre-occupies the specified operation time of the target berth and the passage resources of key nodes along the path. The pre-occupancy time is set according to the estimated operation time of the task plus the preset buffer time. When the truck completes the operation, the task is canceled, or an abnormal adjustment occurs, the elastic release mechanism is triggered to release the pre-occupied resources in real time and update the resource occupancy status database to ensure resource reuse efficiency.

[0297] Specifically, the resource management module 44 addresses the pain points of resource conflicts and low utilization rates in port scheduling through a "dynamic pre-allocation-elastic release" mechanism, ensuring the feasibility of scheduling plans. The specific implementation is as follows:

[0298] Dynamic resource pre-allocation mechanism:

[0299] Resources are locked in a tiered priority system based on "berth resources > route node resources > cargo space resources" to avoid high conflicts caused by low-priority resource occupancy. The pre-occupancy duration is dynamically adjusted using the following formula:

[0300] Pre-occupancy duration = + ;

[0301] in, The estimated operation time for the task (30 minutes for import unloading, 25 minutes for export loading, and 15 minutes for in-yard transfer) is based on the average time statistics of similar tasks in the past 3 months. =10 minutes (preset buffer time), adapting to truck arrival deviation (average ±8 minutes) and work preparation time (average 2 minutes), avoiding resource conflicts caused by minor delays.

[0302] If the target resource has already been reserved, the query will be re-queried 10 minutes later, with a maximum of 3 re-queries; if there is still a conflict, the algorithm module will be notified to replan the path.

[0303] Elastic resource release mechanism:

[0304] The specific conditions for triggering the release include:

[0305] Normal release: The truck will be automatically released when GPS confirms that it has left the target berth and the mission is completed;

[0306] Active release: Released immediately upon task cancellation or scheduling adjustment;

[0307] Release after timeout: If no operation is detected to start within 3 minutes after the end of the pre-occupied time period, the operation will be automatically released (3 minutes is the normal buffer time for starting dock operations).

[0308] After resources are released, the resource occupancy status database is updated in real time (1Hz update) and pushed to the available resource pool to ensure the resource reuse efficiency of other scheduling tasks.

[0309] Resource utilization monitoring:

[0310] Daily statistics are calculated based on "Resource Utilization Rate = Actual Resource Occupancy Time / Expected Occupancy Time". The target utilization rate is ≥85% (meeting the industry standard for efficient utilization of port resources). If the utilization rate is lower than 60%, the expected occupancy buffer time is automatically adjusted to continuously optimize the rationality of resource allocation.

[0311] The data feedback module 45 feeds back the truck-berth matching results and dynamic path information output by the dynamic path planning and berth matching module 43, as well as the resource pre-occupancy and release status data of the resource management module 44, to the scheduling instruction execution and dynamic adjustment unit 5 and the learnable incentive and behavior guidance unit 2, respectively. The data fed back to the learnable incentive and behavior guidance unit 2 is used to optimize the training sample library of the incentive coefficient.

[0312] Specifically, the data feedback module 45 is key to building the "scheduling-execution-optimization" closed loop. By providing targeted feedback of scheduling data, it supports the iterative optimization of other units. The specific implementation is as follows:

[0313] Targeted feedback content:

[0314] Feedback is sent to the scheduling instruction execution and dynamic adjustment unit 5: truck-berth matching results, optimal route node sequence, estimated departure / arrival time, and resource pre-occupancy period, to support the accurate execution of scheduling instructions;

[0315] Feedback is provided to the learnable incentive and behavior guidance unit 2: truck scheduling status (scheduled / unscheduled), target berth information, and resource pre-occupancy / release status—supplementing the training sample library for incentive coefficients and optimizing the targeting of incentive strategies.

[0316] Highly reliable feedback mechanism:

[0317] Feedback timeliness: The scheduling plan is pushed out in real time after it is generated, and updated immediately when the resource status changes (delay ≤ 50ms, adapting to the real-time data transmission requirements of industry).

[0318] Retry upon failure: Automatically retry 3 times (the normal number of retries for industrial data transmission) if push fails. If it still fails, an alarm log is recorded (retained for 90 days) to ensure the reliability of data transmission.

[0319] Value of feedback data:

[0320] Feedback data is used for subsequent "scheduling effect evaluation" (such as on-time arrival rate of container trucks and resource utilization) and "algorithm iteration" (such as adjusting the weight of berth matching score and congestion impact coefficient) to continuously improve the adaptability of the scheduling system to port operation scenarios.

[0321] The scheduling instruction execution and dynamic adjustment unit 5 converts the scheduling plan into standardized instructions and pushes them to the truck terminal. It collects the truck's operating status, berth operation status, and scheduling plan execution data in real time, and determines abnormal status through multi-dimensional verification. When an abnormality occurs, it calls the path planning function of the path planning and scheduling unit 4 to generate an alternative plan and update the instructions. Simultaneously, it feeds back the instruction execution results, status monitoring data, and abnormal handling information to the behavior data acquisition and feature coupling unit 1, the learnable incentive and behavior guidance unit 2, the scheduling demand prediction unit 3, and the path planning and scheduling unit 4 according to functional requirements.

[0322] In this embodiment, the scheduling instruction execution and dynamic adjustment unit 5 includes a standardized instruction generation and push module 51, a multi-source status data acquisition module 52, an anomaly verification and judgment module 53, an alternative solution generation and update module 54, and a targeted data feedback module 55, wherein:

[0323] The standardized instruction generation and push module 51 converts the scheduling scheme output by the path planning and scheduling unit 4 into standardized instructions and pushes them to the truck terminal.

[0324] Specifically, the core of the standardized instruction generation and push module 51 is to complete the conversion and reliable push of the scheduling plan into executable instructions for the terminal. It first extracts the core information output by the path planning and scheduling unit 4, such as the target berth ID, optimal path node sequence, estimated departure / arrival time, and resource pre-occupancy period, and converts it into structured instructions according to the general industrial control specifications. The instruction format adopts the highly compatible JSON, and the character encoding is UTF-8. Then, it pushes the instructions to the container truck on-board terminal through the MQTT protocol commonly used in port IoT. The first push is completed within 1 second after the scheduling plan is generated. At the same time, a 3-retry mechanism is configured (each time with an interval of 5 seconds). After receiving the terminal's "instruction received" receipt, the push is marked as successful. If it is not received, an alarm log is recorded and a pop-up prompt is triggered in the scheduling management background to ensure that the instruction is not lost.

[0325] The multi-source status data acquisition module 52 collects real-time data on the truck's operating status, berth operation status, and scheduling plan execution.

[0326] Specifically, the multi-source status data acquisition module 52 is responsible for capturing real-time status data of the entire scheduling execution chain to support anomaly detection. Its collection scope includes truck operation status (real-time latitude and longitude updated at 1Hz, driving speed, start-stop status, remaining fuel and driver operation feedback), berth operation status (idle / busy / fault status updated every 30 seconds, operation time consumed and remaining time), and scheduling execution data (instruction reception time, actual departure time, nodes traveled, and estimated arrival time deviation). The acquisition method adopts a combination of "terminal active reporting + system passive retrieval". Truck data is reported through the vehicle GPS and 4G module, berth data is obtained through cargo space sensors and PLC system, and scheduling execution data is extracted from the vehicle terminal log. Data transmission adopts the HTTPRESTful protocol. After receiving, timestamp alignment (based on the system reference time, direct alignment if deviation ≤ 1 second) and integrity verification are performed (core fields are retained if no missing, otherwise marked as abnormal and the terminal is required to retransmit), finally forming a standardized status dataset.

[0327] The anomaly detection and judgment module 53 performs multi-dimensional verification to determine the anomaly status of the data collected by the multi-source status data acquisition module 52;

[0328] Specifically, the anomaly detection and judgment module 53 completes anomaly judgment based on multi-dimensional thresholds and logical rules. The core verification dimensions include three categories: time deviation, position deviation, and state conflict, as detailed below:

[0329] Time deviation verification uses "the deviation between the actual departure time and the estimated departure time" and "the deviation between the estimated arrival time and the time required for the remaining distance" as indicators. If either deviation is ≥5 minutes, a time-related anomaly is triggered.

[0330] The position deviation verification uses the well-known Haversine formula to calculate the shortest distance between the real-time position of the truck and the planned path. If the distance is greater than 10 meters and there are 3 consecutive sampling points, it is judged as a position anomaly.

[0331] The status conflict check compares the truck's operating status with the berth status. For example, if there is a logical contradiction, such as the truck being "awaiting operation" but the corresponding berth being "faulty" or the truck being "operating" but the berth being "idle", it will be directly judged as a status exception.

[0332] The above three types of checks are performed at a frequency of 1 second / time. The anomaly judgment result is immediately synchronized to the alternative solution generation and update module 54, along with the anomaly type, trigger time and related data.

[0333] When the anomaly verification and judgment module 53 determines that an anomaly exists, the alternative solution generation and update module 54 calls the path planning function of the path planning and scheduling unit 4 to generate an alternative solution and update the standardized instructions.

[0334] Specifically, the alternative solution generation and update module 54 is responsible for rapid iteration of solutions and instruction updates in abnormal scenarios. After receiving the abnormal signal from the abnormal verification and judgment module 53, it calls the path planning function of the path planning and scheduling unit 4 through the preset API interface within 1 second. At the same time, it uploads supplementary data such as the abnormality type, the current status of the truck, and the real-time information of the berth, clarifying that the alternative solution must avoid the original abnormal factors (such as replanning the path for abnormal location and matching a new berth for berth failure). After calling the alternative solution returned by the path planning and scheduling unit 4, it generates an update instruction according to the instruction format standard of the standardized instruction generation and push module 51. The instruction ID is suffixed with "-V1" and "-V2" to distinguish the version. Then, it prioritizes pushing the update instruction through the pop-up window of the vehicle terminal with an "abnormal adjustment" prompt. At the same time, it uses the 3-retry mechanism to ensure reception. After the update is successful, the new instruction is associated with the alternative solution and stored. If it fails, it triggers the manual dispatch intervention process.

[0335] The directional data feedback module 55 feeds back the instruction execution results, status monitoring data and anomaly handling information to the behavior data acquisition and feature coupling unit 1, the learnable incentive and behavior guidance unit 2, the scheduling demand prediction unit 3 and the path planning and scheduling unit 4 according to functional requirements.

[0336] Specifically, the directional data feedback module 55 constructs a closed-loop feedback channel for scheduling execution data, and organizes the feedback data according to the functional requirements of each unit, as follows:

[0337] Raw state data and instruction execution logs are pushed to the behavior data acquisition and feature coupling unit 1 for feature extraction and data aggregation.

[0338] Feedback on the truck scheduling completion status and driver response time to the learnable incentive and behavior guidance unit 2 supports the optimization of incentive coefficients;

[0339] Push anomaly type, actual dwell time period and prediction deviation data to scheduling demand prediction unit 3 for model iteration;

[0340] Feedback on the execution effect of the scheme (such as the route travel time and berth matching success rate) and the results of anomaly handling is provided to the route planning and scheduling unit 4 to assist in the adjustment of algorithm parameters;

[0341] Furthermore, the feedback adopts a combination of "real-time triggering + timed aggregation" mode. Key data (such as abnormal results) are pushed in real time, while regular data is transmitted in batches every 5 minutes. The transmission protocol is adapted to each receiving unit (TCP / HTTP optional) to ensure that the data flows in a targeted manner without redundancy.

[0342] like Figure 2As shown, this embodiment also provides a dynamic optimization scheduling method for terminal container trucks that combines machine learning and path planning. Based on the above-mentioned dynamic optimization scheduling system for terminal container trucks that combines machine learning and path planning, it includes the following steps:

[0343] S1. Behavioral data collection and feature coupling: Multi-source timestamped data is collected through GPS, RFID readers, driver's APP and port management system interface. After cleaning and calibration, behavioral and scheduling business features are extracted to construct coupled features. After standardization, the data is partitioned and stored in a distributed database according to "truck number-date".

[0344] S2, Learnable Incentive Trigger and Behavior Guidance: Based on the coupling features of S1, non-work-related stay events and medium-urgency tasks are determined. The incentive coefficient is output through the PPO algorithm to calculate the amount. The amount is pushed according to the priority of stay duration and congestion index. If there is no response, the amount is adjusted and the push is re-pushed. Relevant data is simultaneously fed back to subsequent stages.

[0345] S3. Forward-looking prediction of scheduling demand: Based on the standardized time-series data of S1, the correlation between historical dwell times and working conditions is learned through the LSTM model to predict the possible dwell times of trucks in the future, supplement the optimization results of real-time working condition data, and form a forward-looking constraint basis;

[0346] S4. Route Planning and Scheduling Execution: Integrate relevant data from S1 to S3, construct a four-dimensional constraint system through a structural causal model, use an improved A* algorithm to complete truck-berth matching and route planning, simultaneously activate resource pre-occupancy and elastic release mechanisms, and feed back scheduling data to the corresponding links;

[0347] S5. Execution and Dynamic Adjustment of Scheduling Instructions: The S4 scheduling scheme is converted into standardized instructions and pushed to the truck terminal. Real-time data on operation and work status is collected. After three-dimensional verification to determine anomalies, alternative schemes are generated and instructions are updated. Relevant information is fed back to each upstream link.

[0348] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0349] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A terminal tractor dynamic optimization scheduling system combining machine learning and path planning, characterized in that, include: The behavior data acquisition and feature coupling unit (1) collects relevant data on trucks, yards, and scheduling through GPS module, RFID reader, driver's APP and terminal management system interface. After data cleaning, it constructs behavior-scheduling coupling features and stores standardized time series data in the database. The learnable incentive and behavior guidance unit (2) is based on the behavior-scheduling coupling feature. It uses "non-job stay event + task urgency matching" as the trigger condition, optimizes the incentive coefficient through reinforcement learning to generate incentive amount, pushes instant micro-incentives through the driver's mobile APP and executes automatic re-push; when multiple drivers stay at the same time, the driver with the longer stay time is given priority in incentive; after the incentive is triggered, the driver's stay time information is fed back to the subsequent path planning and scheduling stage in real time, so that the path planning uses the stay time as a dynamic constraint condition and automatically skips the stay time to generate a new task path; the learnable incentive and behavior guidance unit (2) includes an incentive trigger judgment module (21), a reinforcement learning optimization module (22) and an incentive push and data feedback module (23), wherein: The incentive trigger judgment module (21) receives the behavior-scheduling coupling features output by the behavior data acquisition and feature coupling unit (1) as the core input for trigger condition judgment; the process of the incentive trigger judgment module (21) performing the "non-job stay event + task urgency matching" judgment includes the following steps: S21.1 Receive the behavior-scheduling coupling features output by the behavior data acquisition and feature coupling unit (1), and synchronously acquire the continuous location data of the truck from the GPS module, the yard corresponding cargo location operation status data from the RFID reader, and the current task deadline data from the terminal management system. S21.2 Set a position deviation threshold and a continuous sampling point number threshold. When both conditions are met simultaneously, and the duration of the condition is greater than or equal to the preset duration threshold, it is determined to be a non-operational stop event. S21.3, based on the difference between the task deadline and the current time, quantize it into a task urgency value in the interval [0, 1] through linear mapping The larger the value, the more urgent the task is, wherein the mapping reference is the longest task cycle preset by the terminal. S21.4、when the non-job stay event determination is established, and the task urgency value is in the preset medium urgency interval, it is determined that the trigger condition matches, and a trigger signal, a non-job stay duration, a task urgency value and a yard congestion index are output to the reinforcement learning optimization module (22). The reinforcement learning optimization module (22) receives the trigger event data output by the incentive trigger judgment module (21), trains the policy network model using the PPO algorithm, and outputs the incentive coefficients, including the following steps: S22.1 Construct a policy network model for the PPO agent. The network uses non-task dwell time and task urgency values ​​from the trigger event data. The yard congestion index is the input feature, and the output is the incentive coefficient corresponding to the dwell time weight. Incentive coefficients corresponding to the weights of the congestion index The network has multiple fully connected layers, and nonlinear feature mapping is achieved through ReLU and tanh activation functions; S22.2 Define a two-dimensional reward signal: the core reward is negatively correlated with the driver's response time, the constraint reward is negatively correlated with the incentive amount, and the total reward signal is the weighted sum of the core reward and the constraint reward. S22.3 Daily collection of "trigger condition data - incentive coefficient" , The triplet of "driver response result" is used as a training sample and input into the policy network for incremental training. The parameter update range is limited by the parameter pruning mechanism of the PPO algorithm. S22.4, Call the optimal activation coefficients output by the trained network. and The real-time incentive amount is calculated by combining the non-operational dwell time and yard congestion index in the trigger event data. ; The incentive push and data feedback module (23) generates specific incentive instructions and pushes them based on the incentive coefficients output by the reinforcement learning optimization module (22); and synchronizes the incentive execution results to the path planning and scheduling unit (4) and the behavior data acquisition and feature coupling unit (1) to form a data flow closed loop; The scheduling demand prediction unit (3) is based on standardized time series data, uses an LSTM model to learn the correlation between historical dwell time and working conditions, predicts the possible dwell time of each truck in a specified future period, and optimizes the model output by supplementing the latest real-time data, providing a forward-looking constraint basis for scheduling. The route planning and scheduling unit (4) receives the dwell time prediction results from the scheduling demand prediction unit (3), the real-time dwell events and driver response status from the learnable incentive and behavior guidance unit (2), and the multi-source operation data from the behavior data collection and feature coupling unit (1). It generates path constraints through a structural causal model, completes truck-parking space matching and dynamic route planning using an improved A* algorithm, and, combined with the resource pre-occupancy and elastic release mechanism, synchronously feeds back scheduling data to the scheduling instruction execution and dynamic adjustment unit (5) and the learnable incentive and behavior guidance unit (2). The scheduling instruction execution and dynamic adjustment unit (5) converts the scheduling scheme into standardized instructions and pushes them to the truck terminal. It collects the truck operation status, berth operation status and scheduling scheme execution data in real time, and judges the abnormal status through multi-dimensional verification. When there is an abnormality, it calls the path planning function of the path planning and scheduling unit (4) to generate an alternative scheme and update the instructions. Simultaneously, it feeds back the instruction execution results, status monitoring data and abnormal handling information to the behavior data collection and feature coupling unit (1), the learnable incentive and behavior guidance unit (2), the scheduling demand prediction unit (3) and the path planning and scheduling unit (4) according to the functional requirements.

2. The dynamic optimization scheduling system for terminal container trucks combining machine learning and path planning as described in claim 1, characterized in that, The behavior data acquisition and feature coupling unit (1) includes a multi-source data access module (11), a data cleaning and calibration module (12), a behavior-scheduling feature construction module (13), and a standardized data storage module (14), wherein: The multi-source data access module (11) collects truck location, driving speed and start / stop status data through GPS module, collects container loading and unloading status and cargo space occupancy information through UHF RFID reader, collects driver manually marked reasons for stopping data through driver APP interaction interface, and connects to the terminal management system through HTTPRESTful interface to synchronize task type, task urgency and berth operation status data. Each data is accompanied by a timestamp. The data cleaning and calibration module (12) uses the output data of the multi-source data access module (11) to perform trajectory continuity verification and correct drift data for GPS data, and to perform adjacent state consistency verification and correct conflict state for RFID data. It fills in continuous and discrete data for missing data respectively and removes time-series segments with an excessive proportion of invalid data. The behavior-scheduling feature construction module (13) extracts the behavioral features of non-operation dwell time and operation response time of container trucks, as well as the scheduling business features of task urgency and yard congestion index, based on the effective data of the data cleaning and calibration module (12), and constructs the coupling features of behavior and scheduling requirements. The standardized data storage module (14) converts the original data and coupled feature data into a columnar storage format based on the feature data of the behavior-scheduling feature construction module (13).

3. The dynamic optimization scheduling system for terminal container trucks combining machine learning and path planning as described in claim 2, characterized in that, The process of the incentive push and data feedback module (23) executing incentive instruction push and data synchronization includes the following steps: S23.1 When multiple drivers trigger non-operational stop events simultaneously, the priority score of each driver is calculated based on the stop duration and yard congestion index output by the incentive trigger judgment module (21). , Positively correlated with dwell time and congestion index, according to Incentive resources are allocated and pushed in descending order. S23.2, Push incentive amounts to target drivers via driver-side app The instruction is sent, and a response timer is started. If no "location movement confirmation" signal is received from the app within the first preset response time, a second push is triggered. If there is still no response after the second preset response time, the push is terminated, and the incentive amount for the second push is adjusted to [amount missing]. , for The product of the preset incentive adjustment coefficient; S23.3 After the excitation is triggered, the start and end timestamps of the stay period marked in the GPS data are extracted, encapsulated into a standardized data frame, and pushed to the path planning and scheduling unit in real time via wireless communication (4). S23.

4. The driver's response status and the actual dwell time change data after the response are synchronously fed back to the standardized data storage module of the behavior data acquisition and feature coupling unit (1).

4. The dynamic optimization scheduling system for terminal container trucks combining machine learning and path planning as described in claim 3, characterized in that, The scheduling demand prediction unit (3) includes a data input and preprocessing module (31), an LSTM model processing module (32), and a real-time output optimization module (33), wherein: The data input and preprocessing module (31) receives the standardized time series data output by the behavior data acquisition and feature coupling unit (1), performs format regularization and feature extraction on the historical dwell time period and working condition information in the data, and forms the basic data for the input of the LSTM model. The LSTM model processing module (32) uses the basic data output by the data input and preprocessing module (31) to learn the correlation between historical dwell time and working conditions using the LSTM model, and infers the possible dwell time of each truck in the future specified period. The real-time output optimization module (33) receives the latest real-time operating data, supplements the real-time operating data into the output of the LSTM model processing module (32), dynamically optimizes the prediction results of possible dwell time periods, and forms the forward-looking constraint basis required for scheduling.

5. The dynamic optimization scheduling system for terminal container trucks combining machine learning and path planning according to claim 4, characterized in that, The path planning and scheduling unit (4) includes a data integration and preprocessing module (41), a path constraint generation module (42), a dynamic path planning and berth matching module (43), a resource management module (44), and a data feedback module (45), wherein: The data integration and preprocessing module (41) synchronously accesses the future stay time prediction results output by the scheduling demand prediction unit (3), the real-time stay events and driver response status data output by the learnable incentive and behavior guidance unit (2), and the multi-source operation data output by the behavior data collection and feature coupling unit (1). It performs format standardization conversion and timestamp alignment on the heterogeneous data, removes expired data that exceeds the effective time window, and forms a unified scheduling input dataset. The path constraint generation module (42) constructs a multi-dimensional path constraint system based on the standardized dataset of the data integration and preprocessing module (41) through a structural causal model: it uses the predicted dwell time and real-time dwell time as "time constraints" to clarify the unschedulable time periods for trucks; it maps the yard congestion index as "spatial constraints" to define the path passage priority in high-congestion areas; it uses the berth operation status and cargo space occupancy information as "resource constraints" to limit the range of berths that trucks can match; and it uses the task urgency as "priority constraints" to set the execution order weight of scheduling tasks. The dynamic path planning and berth matching module (43) uses the multi-dimensional constraints of the path constraint generation module (42) as boundary conditions and adopts the improved A* algorithm with integrated dynamic weights to realize truck-berth matching and path planning: the heuristic function design incorporates the task urgency coefficient and the dynamic weight of the yard congestion. The higher the urgency and the lower the congestion index, the greater the weight of the path. During the path search process, the "time constraint" limited dwell time period and the "space constraint" limited high congestion area are avoided in real time. The availability of the berth "resource constraint" is checked at the same time, and the optimal truck scheduling path and berth matching scheme that takes into account both the shortest path and the timeliness of the task are output. The resource management module (44) initiates a resource pre-occupancy mechanism based on the scheduling scheme output by the dynamic path planning and berth matching module (43): after the truck scheduling path is generated, it automatically pre-occupies the specified operation time of the target berth and the passage resources of key nodes along the path. The pre-occupancy time is set according to the estimated operation time of the task plus the preset buffer time. When the truck completes the operation, the task is canceled or an abnormal adjustment occurs, the elastic release mechanism is triggered to release the pre-occupied resources in real time and update the resource occupancy status database to ensure resource reuse efficiency. The data feedback module (45) feeds back the truck-berth matching results and dynamic path information output by the dynamic path planning and berth matching module (43), as well as the resource pre-occupancy and release status data of the resource management module (44), to the scheduling instruction execution and dynamic adjustment unit (5) and the learnable incentive and behavior guidance unit (2), respectively. The data fed back to the learnable incentive and behavior guidance unit (2) is used to optimize the training sample library of incentive coefficients.

6. The dynamic optimization scheduling system for terminal container trucks combining machine learning and path planning as described in claim 5, characterized in that, The scheduling instruction execution and dynamic adjustment unit (5) includes a standardized instruction generation and push module (51), a multi-source status data acquisition module (52), an anomaly verification and judgment module (53), an alternative solution generation and update module (54), and a targeted data feedback module (55), wherein: The standardized instruction generation and push module (51) converts the scheduling scheme output by the path planning and scheduling unit (4) into standardized instructions and pushes them to the truck terminal. The multi-source status data acquisition module (52) collects real-time data on truck operation status, berth operation status, and scheduling scheme execution. The anomaly verification and judgment module (53) performs anomaly judgment on the data collected by the multi-source state data acquisition module (52) through multi-dimensional verification; When the anomaly verification and judgment module (53) determines that there is an anomaly, the alternative solution generation and update module (54) calls the path planning function of the path planning and scheduling unit (4) to generate an alternative solution and update the standardized instructions. The directional data feedback module (55) feeds back the instruction execution results, status monitoring data and anomaly handling information to the behavior data acquisition and feature coupling unit (1), the learnable incentive and behavior guidance unit (2), the scheduling demand prediction unit (3) and the path planning and scheduling unit (4) according to functional requirements.

7. A dynamic optimization scheduling method for terminal container trucks combining machine learning and path planning, based on the dynamic optimization scheduling system for terminal container trucks combining machine learning and path planning as described in any one of claims 1-6, characterized in that, Includes the following steps: S1. Behavioral data collection and feature coupling: Multi-source timestamped data is collected through GPS, RFID readers, driver's APP and port management system interface. After cleaning and calibration, behavioral and scheduling business features are extracted to construct coupled features. After standardization, the data is partitioned and stored in a distributed database according to "truck number-date". S2, Learnable Incentive Trigger and Behavior Guidance: Based on the coupling features of S1, non-work-related stay events and medium-urgency tasks are determined. The incentive coefficient is output through the PPO algorithm to calculate the amount. The amount is pushed according to the priority of stay duration and congestion index. If there is no response, the amount is adjusted and the push is re-pushed. Relevant data is simultaneously fed back to subsequent stages. S3. Forward-looking prediction of scheduling demand: Based on the standardized time-series data of S1, the correlation between historical dwell times and working conditions is learned through the LSTM model to predict the possible dwell times of trucks in the future, supplement the optimization results of real-time working condition data, and form a forward-looking constraint basis; S4. Route Planning and Scheduling Execution: Integrate relevant data from S1 to S3, construct a four-dimensional constraint system through a structural causal model, use an improved A* algorithm to complete truck-berth matching and route planning, simultaneously activate resource pre-occupancy and elastic release mechanisms, and feed back scheduling data to the corresponding links; S5. Execution and Dynamic Adjustment of Scheduling Instructions: The S4 scheduling scheme is converted into standardized instructions and pushed to the truck terminal. Real-time data on operation and work status is collected. After three-dimensional verification to determine anomalies, alternative schemes are generated and instructions are updated. Relevant information is fed back to each upstream link.

Images