Single gantry crane dispatching system based on intelligent optimization algorithm

Abstract

The invention discloses a single gantry crane dispatching system based on an intelligent optimization algorithm. The single gantry crane dispatching system comprises an urgency degree calculation module, a task urgency degree calculation module and a task urgency degree calculation module, the future task prediction module is used for determining a prediction task set through a sliding time window and calculating the discount urgency degree of a prediction task; the path planning and task sorting module is used for modeling a task scheduling problem into a state transition problem on a search tree, and gradually exploring a combined space of a task execution sequence and a moving path in the search tree by adopting an improved A * algorithm under the guidance of urgent degree discount cost to obtain an optimal task sequence; the self-adaptive parameter adjusting module is used for optimizing parameters on line based on reinforcement learning; and the event-driven control module is used for responding to task arrival, periodic refreshing, abnormal events and task completion events through the message queue and triggering dynamic rescheduling. According to the method, the problems of dynamics, diversity and burstiness of tasks in actual operation can be effectively solved, and the scheduling efficiency of large-scale logistics equipment is improved.

Description

Technical Field

[0001] This invention relates to the field of port logistics automation technology, and in particular to a single gantry crane scheduling system based on intelligent optimization algorithms. Background Technology

[0002] In container terminal yard operations, the single gantry crane (RMG) is a core loading and unloading equipment, and its scheduling efficiency directly affects the port's turnover capacity.

[0003] For example, Chinese patent document CN104077638A discloses a scheduling method for gantry cranes. The method aims to improve the utilization rate of cranes, uses crane task sorting and scheduling as the idea, and adopts the comprehensive integration optimization of crane usage task classification, crane usage task priority and crane selection as the technical means, with the goal of orderly production tasks.

[0004] Chinese patent document CN118396287A discloses a scheduling method for container operations at a terminal, comprising: determining a target container from containers corresponding to the same type of operation instructions; determining a target gantry crane based on the location information of the gantry crane corresponding to the target container; determining a target truck based on the instruction sorting result of the target container, the target gantry crane, and the positioning data of a truck in normal condition, wherein the instruction sorting result is the order in which the target gantry crane executes operation instructions on the corresponding target container; and outputting a scheduling instruction based on the instruction sorting result of the target container, the target gantry crane, and the target truck, wherein the scheduling instruction is used to control the target truck and the target gantry crane to execute operation instructions on the target container according to the instruction sorting result.

[0005] However, traditional scheduling methods have the following limitations:

[0006] 1. Static Priority Strategy: Existing systems often use fixed rules (such as first-come, first-served, shortest distance priority), which cannot dynamically quantify the urgency of tasks. For example, emergency transfer tasks are delayed due to long distances; VIP containers are scheduled together with ordinary tasks, causing high-value operations to time out.

[0007] 2. Short-sightedness in path planning: Classic algorithms (such as greedy algorithms) only optimize the current path and do not consider the future task distribution, which easily leads to the problem of "repeated back-and-forth". For example, after processing the current task, a long distance needs to be moved to the new task area, increasing the cumulative movement cost by more than 30%.

[0008] 3. Lack of predictive and adaptive capabilities: When sudden tasks are inserted, manual rescheduling is required, and the response delay exceeds 5 minutes; fixed weight parameters (such as) cannot adapt to changes in the intensity of work at different times, and the overtime rate can reach 15% during peak periods.

[0009] 4. Rigid anomaly handling: Sudden events such as track failures trigger global shutdowns, lacking local replanning mechanisms, resulting in a 40% decrease in equipment utilization.

[0010] Therefore, there is an urgent need for an intelligent scheduling system that integrates dynamic priority assessment, forward-looking path planning, online parameter optimization, and flexible event response to break through the efficiency bottleneck of single gantry crane operations. Summary of the Invention

[0011] This invention provides a single gantry crane scheduling system based on intelligent optimization algorithms, which can effectively handle the dynamic, diverse and sudden problems of tasks in actual operations and improve the efficiency of intelligent scheduling of large logistics equipment.

[0012] A single-gantry crane scheduling system based on intelligent optimization algorithms includes:

[0013] Urgency Calculation Module: Used for dynamically weighted calculation of the urgency score U for each task. i The formula is:

[0014]

[0015] Among them, U i This represents the urgency score for the i-th task. Let T be the expected completion time for the i-th task. now D represents the current system time. (i) For the current gantry crane position B cur To the mission objective B i The distance, P (i) α, β, and γ are the task priority weights, and α, β, and γ are adjustable parameters.

[0016] Future task prediction module: using a sliding time window W predict =[T now ,T now +Δt] determines the prediction time period, and constructs a prediction task set based on the data source. And calculate the discount urgency of the prediction task, using the following formula:

[0017] Where Δt represents the width of the sliding window, This indicates the target position, expected type, and priority information for the prediction task. This represents the probability confidence level of the prediction task. θ represents the estimated arrival time of the prediction task, and θ represents the confidence threshold.

[0018] Path planning and task sequencing module: Based on urgency scoring, the task scheduling problem is modeled as a state transition problem on a search tree. An improved A* algorithm is used to explore the combination space of task execution order and movement path in the search tree, guided by urgency discount cost, to obtain the optimal task sequence.

[0019] Adaptive parameter tuning module: Based on reinforcement learning, online optimization of parameters α, β, γ is performed, with the loss function L driving parameter updates;

[0020] Event-driven control module: Responds to task arrival events, periodic refresh events, exception events, and task completion events through message queues, triggering dynamic rescheduling.

[0021] The system of this invention is particularly suitable for the real-time task scheduling and path planning of gantry cranes in container terminal yard operations. Through multi-module collaboration, the system achieves dynamic task priority assessment, path cost optimization, future task prediction, and adaptive parameter adjustment, solving the problems of minimizing movement costs and controlling task timeout rates in single gantry crane operation scenarios.

[0022] In the urgency calculation module, when Forced U i =∞;

[0023] Task priority weight P (i) The values ​​for VIP containers, regular containers, and emergency transport containers are set at 1.5, 1.0, and 2.0 respectively.

[0024] In the future task prediction module, the sliding window width Δt is 60-120 seconds, and the confidence threshold θ is 0.6-0.8.

[0025] In the future mission prediction module, the data sources include: historical mission arrival frequency distribution, container truck ETA forecast data, real-time detection results of port area sensing equipment, and time-based rules for operation plans.

[0026] In the path planning and task sorting module, an improved A* algorithm is used to evaluate the state n of each node in the search tree, and its evaluation function is defined as follows:

[0027] f(n) = g(n) + h(n)

[0028] Where f(n) represents the total cost assessment of the current path from the starting point to state n; g(n) is the cumulative actual cost of the completed task sequence; and h(n) is the estimated total cost of the remaining tasks.

[0029] g(n) is defined as:

[0030]

[0031] Where k represents the number of tasks completed in the current node state, and λ is the timeout penalty coefficient. Represents task T i Actual completion time Indicates the expected time for the task to be completed.

[0032] h(n) is defined as:

[0033]

[0034] in, T represents the set of tasks that have not yet been scheduled. i express The i-th task, Represents the set of prediction tasks. express The i-th task, This represents the scheduling cost required for the prediction task.

[0035] In the adaptive parameter tuning module, the Q-learning reinforcement learning algorithm is adopted; the state space includes the task intensity level and the timeout rate range; the action space is the ±0.1 step adjustment of α, β, and γ; the reward function is R = -L, and the update strategy adopts ∈-greedy exploration.

[0036] In the adaptive parameter tuning module, the loss function L is defined as follows:

[0037]

[0038] Where λ1 and λ2 are weighting factors, r represents the average time to complete the task. timeout This indicates the task timeout rate.

[0039] In the event-driven control module, when a task arrives, urgency recalculation and A* global replanning are triggered; when a periodic refresh event is triggered, the entire process scheduling is forced to refresh at a fixed period; when an abnormal event is triggered, local replanning is triggered and tasks in the fault area are eliminated; when a task completes, the gantry crane position is updated and the remaining task set is re-evaluated.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. Enhanced scientific scheduling: This invention uses an urgency function to model multiple factors in a unified manner, allowing for quantitative evaluation of scheduling priorities;

[0042] 2. Higher efficiency: This invention effectively reduces task skipping and backtracking through A* path planning and predictive task guidance;

[0043] 3. Enhanced adaptability: The scheduling parameters of this invention can be optimized online based on system operating conditions;

[0044] 4. Improved robustness: This invention adapts to unexpected situations and dynamic changes in tasks through event-driven and task prediction mechanisms;

[0045] 5. Easy to deploy and expand: The decoupled design of each module in this invention facilitates practical industrial implementation and module expansion and upgrade. Attached Figure Description

[0046] Figure 1 This is a block diagram of a single gantry crane scheduling system based on an intelligent optimization algorithm, according to an embodiment of the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0048] A single-gantry crane scheduling system based on intelligent optimization algorithms mainly includes an urgency calculation module, a future task prediction module, a path planning and task sequencing module, an adaptive parameter tuning module, and an event-driven control module. It coordinates scheduling from five aspects: task evaluation, task prediction, path planning, parameter optimization, and system scheduling control, forming an intelligent, real-time, and robust single-gantry crane scheduling system.

[0049] (1) Urgency Calculation Module

[0050] This module first quantifies the "urgency" of a task, with the urgency function being its core component:

[0051]

[0052] Among them, U i This represents the urgency score for the i-th task; T represents the expected completion time of the i-th task; now Indicates the current system time; D (i) =|B cur -B i | indicates the current position of the gantry crane B. cur With mission objective B i Distance of P; (i) The weights represent the importance of the task; α, β, and γ represent adjustment parameters that can be adaptively learned.

[0053] This formula takes into account time urgency, movement costs, and task priority, and is used to prioritize and evaluate tasks, forming the basis for system decision-making.

[0054] (2) Future Task Prediction Module

[0055] To overcome the short-sighted behavior of traditional scheduling methods in dynamic task scenarios, this invention introduces a future task prediction module before the path planning module is executed. A sliding time window mechanism is used to probabilistically model the tasks that are about to arrive, thereby improving the foresight and robustness of the scheduling system's path decision.

[0056] In actual port operations, tasks are not generated all at once, but dynamically as external events such as truck arrivals and the issuance of operational instructions occur. If the scheduling system only sorts and plans routes based on the current task set, delays in task arrival can easily cause cranes to have to turn back as soon as they leave the target area, increasing unnecessary movement costs and scheduling delays. Therefore, it is necessary to introduce a future task prediction mechanism before scheduling execution to model tasks that may occur within a certain time frame.

[0057] This module uses a sliding time window mechanism for prediction. Let the current system time be T. now The predicted time period is denoted as:

[0058] W predict =[T now ,T now +Δt]

[0059] Where Δt is the window width, the specific value of which can be set according to the actual workload, such as 60 to 120 seconds.

[0060] This module builds a prediction model based on one or more of the following data sources: task arrival frequency distribution based on historical scheduling data; vehicle arrival time forecast data from the truck scheduling system; real-time target detection results provided by port area sensing devices (such as AGVs, cameras, RFID); and time-based task generation rules in the work plan.

[0061] Based on the above information, the system constructs a set of prediction tasks: in, For the target position, expected type, and priority information of the prediction task; The probability confidence level for predicting the task; θ represents the estimated arrival time of the prediction task; θ∈[0,1] is the confidence threshold set by the system, which is usually 0.6 to 0.8.

[0062] To avoid drastic disturbances to scheduling results caused by prediction tasks, the system does not directly include prediction tasks in the scheduling execution list. Instead, it uses a discounted approach in the urgency scoring and path cost estimation. The definition is as follows:

[0063]

[0064] in, Let T be the expected completion time for the i-th prediction task; now The current system time; This is the estimated distance from the current position to the predicted target position. The importance weights for the prediction task are α, β, and γ, which are adjustment parameters that can be adaptively learned.

[0065] The above expression introduces the predicted task into the ranking evaluation in a flexible form without changing the existing path planning framework. This allows the scheduling system to avoid blind movement of subsequent task-intensive areas in advance when deciding the current task order, thereby effectively reducing global path redundancy and task timeout rate.

[0066] This module operates in a fixed-period manner (e.g., refreshes every 10 seconds) and supports an online correction mechanism based on task status feedback. It can work together with the parameter tuning module to achieve a rolling optimization closed loop of prediction-planning-execution.

[0067] (3) Path planning and task sequencing module

[0068] Based on urgency scoring, this module models the task scheduling problem as a state transition problem on a search tree: each node represents a sequence of completed tasks and the current crane state, and each state transition represents selecting an unfinished task as the next target, forming a path in the search tree.

[0069] This module employs an improved A* heuristic search algorithm, guided by urgency discount costs, to gradually explore the combination space of task execution order and movement path in the search tree, thereby obtaining the task sequence with the optimal or near-optimal global scheduling effect.

[0070] 1. Definition of State and Nodes in the Search Tree

[0071] In this system, each node n in the A* search tree represents an intermediate state in the scheduling process, which is defined as:

[0072]

[0073] Among them, B cur This refers to the current position of the gantry crane; T now The current system time; This is a collection of completed tasks. This is the set of tasks that have not yet been scheduled.

[0074] Starting node (root node): At the initial moment, the gantry crane is in the initial position B0, the initial time is T0, the task set is empty, and all tasks are in the waiting list.

[0075] Node expansion (state transition): Select a task from the current node n. As the next scheduling target, after executing this task, a new state node n' is generated, and the crane position, system time, and task status are updated.

[0076] 2. Definition of A* path cost function

[0077] The classic A* search strategy defines its evaluation function for each node state n in the search tree as follows:

[0078] f(n) = g(n) + h(n)

[0079] Where: f(n) represents the total cost assessment of the current path from the starting point to state n, which is a key indicator used to guide the search direction; the A* algorithm will prioritize expanding the node with the smallest f(n) each time, thus taking into account both the actual cost g(n) already paid and the estimated cost h(n) in the future.

[0080] The two parts of this function have the following meanings:

[0081] g(n) represents the cumulative true cost of the completed task sequence, indicating the total actual cost of the task execution path from the initial state to the current state n. The formula is:

[0082]

[0083] Among them, P (i) As the importance weight of the task, D i To move from the position of the previous task to task T i Distance moved by the location; The actual time it took to complete the task. The expected time for the task to be completed.

[0084] h(n) is the estimated total cost of the remaining tasks (heuristic function). To estimate the minimum total cost required to complete the remaining tasks starting from the current state, this invention designs the following heuristic function:

[0085]

[0086] Among them, D (i) =|B cur -B i | represents the estimated distance traveled from the current position to the target location; U i Rate the urgency of the task to indicate its level of importance; This is the estimated distance from the current position to the predicted target position. Rate the urgency of the task discount.

[0087] This design reflects the urgent task (U) i Larger tasks are given smaller valuations to encourage early execution; tasks that are far away (D) (i) This will increase costs and balance path efficiency; task sequences with lower overall costs will be prioritized for expansion, thereby approaching the optimal scheduling path.

[0088] 3. Module Output

[0089] A path planning scheme includes a sorted sequence of task executions, along with the corresponding movement path for each step and the task completion timestamp.

[0090] (4) Adaptive parameter tuning module

[0091] To adapt to different operating environments and scheduling objectives, this module introduces reinforcement learning-based methods (such as Q-learning) to perform online adaptive optimization of the parameters α, β, and γ in the urgency model, so as to dynamically adjust the task priority evaluation criteria, thereby realizing the self-adjustment and performance improvement of the scheduling strategy under various operating conditions.

[0092] This module defines the following system performance metrics: average task completion time. Task timeout rate r timeout .

[0093] Based on this, construct the scheduling loss function:

[0094]

[0095] Wherein, λ1 and λ2 are weighting factors for the system performance objective, representing the weights of the average task completion time and task timeout rate in the total loss, respectively.

[0096] The system collects scheduling execution results feedback based on the policy gradient method or Q-learning policy evaluation mechanism, taking each scheduling cycle as a unit. Then, it updates the core parameters α, β, γ in the urgency model, gradually approaching the optimal scheduling policy parameter set in the current scenario, thereby improving the system's scheduling quality and environmental adaptability.

[0097] (5) Event-driven control module

[0098] To enhance the system's dynamic response capability, this module adopts an event-driven mechanism for scheduling control, supporting the following scheduling events:

[0099] Task arrival event: Triggers urgency recalculation and task reordering;

[0100] Time-period events: Triggering schedule refresh at a fixed frequency;

[0101] Abnormal events, such as track failures or mission cancellations, trigger rescheduling.

[0102] Task completion event: Update gantry crane location and reassess the remaining task set.

[0103] The system connects the modules using a message queue to ensure real-time scheduling decisions and decoupling between modules.

[0104] To verify the effectiveness of the present invention, such as Figure 1 The diagram illustrates a specific implementation process, which includes the following steps:

[0105] S01: System Initialization and Parameter Configuration

[0106] Hardware deployment: Deploy a scheduling system on the gantry crane control terminal, connecting yard positioning equipment (such as UWB base stations), task management system (TOS), and sensor network (LiDAR / RFID).

[0107] Initial parameters are set as follows: sliding window width Δt = 90; prediction confidence threshold θ = 0.7; initial values ​​for urgency parameters α = 0.3, β = 0.6, γ = 0.2; loss function weights λ1 = 0.3, λ2 = 0.4.

[0108] S02: Real-time Data Acquisition and Task Injection

[0109] Dynamically receive external task sources: Real-time parsing of loading and unloading instructions (container number, target bay B) issued by TOS. i Deadline ); Access ETA data from the truck dispatching system to generate a task set. The actual arrival location of the container is verified by RFID scanning, and the mission coordinates are corrected.

[0110] S03: Dynamic Calculation of Urgency

[0111] For each new task, calculate its urgency score:

[0112]

[0113] Priority P (i) Value assignment rules: VIP containers are set to 1.5, regular containers to 1.0, and emergency transfer containers to 2.0;

[0114] Error handling: If the task has timed out That is, then force U i =∞ (in practice, directly assign the value 100), directly giving it the highest priority.

[0115] S04: Future Mission Prediction and Discount Integration

[0116] Analyze the arrival frequency of tasks within the same time period over a historical 60-minute period to generate a Poisson distribution model; integrate truck ETA data to output a predicted task set. Then calculate the urgency of the discount.

[0117] S05: A* Global Path Planning

[0118] Building the search tree:

[0119] Root node:

[0120] Node expansion: Select The three tasks with the smallest cost function f(n) = g(n) + h(n) generate child nodes. The calculation of the cost function f(n) = g(n) + h(n) is divided into...

[0121] Cumulative cost g(n):

[0122]

[0123] Heuristic function h(n):

[0124]

[0125] Output the generated task sequence {T1→T2→…}.

[0126] S06: Online Parameter Adaptive Optimization

[0127] Q-learning parameter tuning mechanism (updated every 5 minutes):

[0128] Status: Work intensity (low / medium / peak), overtime rate range (<5%, 5-10%, >10%);

[0129] Action: Adjust α, β, γ (step size ±0.1);

[0130] Reward function:

[0131] Update strategy: ∈-greedy exploration, learning rate η = 0.2.

[0132] S07: Event-Driven Scheduling Execution

[0133] The event response logic is shown in Table 1 below.

[0134] Table 1

[0135] Event type Trigger action New task arrival Immediate recalculation S03-S05, generate new path Task completion Update B cur , check remaining task urgency Track failure (exception) Local re-planning: exclude failed zone tasks, re-search by S05 Periodic refresh (every 30 seconds) Force full flow execution S03-S06

[0136] S08: Rolling Optimization and Termination Conditions

[0137] Loop mechanism: After completing a task, return to S02, until...

[0138] Interruption scenarios: manual emergency shutdown command; equipment failure lasting more than 10 minutes; yard operation plan termination signal.

[0139] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A single gantry crane scheduling system based on intelligent optimization algorithms, characterized in that, include: Urgency Calculation Module: Used for dynamically weighted calculation of the urgency score U for each task. i The formula is: Among them, U i This represents the urgency score for the i-th task. Let T be the expected completion time for the i-th task. now D represents the current system time. (i) For the current gantry crane position B cur To the mission objective B i The distance, P (i) α, β, and γ are the task priority weights, and α, β, and γ are adjustable parameters. Future task prediction module: using a sliding time window W predict =[T now ,T now +Δt] determines the prediction time period, and constructs a prediction task set based on the data source. And calculate the discount urgency of the prediction task, using the following formula: Where Δt represents the width of the sliding window, This indicates the target position, expected type, and priority information for the prediction task. This represents the probability confidence level of the prediction task. θ represents the estimated arrival time of the prediction task, and θ represents the confidence threshold. Path planning and task sequencing module: Based on urgency scoring, the task scheduling problem is modeled as a state transition problem on a search tree. An improved A* algorithm is used to explore the combination space of task execution order and movement path in the search tree, guided by urgency discount cost, to obtain the optimal task sequence. Adaptive parameter tuning module: Based on reinforcement learning, online optimization of parameters α, β, β, driven by loss function L; Event-driven control module: Responds to task arrival events, periodic refresh events, exception events, and task completion events through message queues, triggering dynamic rescheduling.

2. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 1, characterized in that, In the urgency calculation module, when Forced U i =∞; Task priority weight P (i) The values ​​for VIP containers, regular containers, and emergency transport containers are set at 1.5, 1.0, and 2.0 respectively.

3. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 1, characterized in that, In the future task prediction module, the sliding window width Δt is 60-120 seconds, and the confidence threshold θ is 0.6-0.

8.

4. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 1, characterized in that, In the future mission prediction module, the data sources include: historical mission arrival frequency distribution, container truck ETA forecast data, real-time detection results of port area sensing equipment, and time-based rules for operation plans.

5. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 1, characterized in that, In the path planning and task sorting module, an improved A* algorithm is used to evaluate the state n of each node in the search tree, and its evaluation function is defined as follows: f(n) = g(n) + h(n) Where f(n) represents the total cost assessment of the current path from the starting point to state n; g(n) is the cumulative actual cost of the completed task sequence; and h(n) is the estimated total cost of the remaining tasks.

6. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 5, characterized in that, g(n) is defined as: Where k represents the number of tasks completed in the current node state, and λ is the timeout penalty coefficient. Represents task T i Actual completion time Indicates the expected time for the task to be completed.

7. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 5, characterized in that, h(n) is defined as: in, T represents the set of tasks that have not yet been scheduled. i express The i-th task, Represents the set of prediction tasks. express The i-th task, This represents the scheduling cost required for the prediction task.

8. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 1, characterized in that, In the adaptive parameter tuning module, the Q-learning reinforcement learning algorithm is adopted; the state space includes the task intensity level and the timeout rate range; the action space is the ±0.1 step adjustment of α, β, and γ; the reward function is R = -L, and the update strategy adopts ∈-greedy exploration.

9. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 1, characterized in that, In the adaptive parameter tuning module, the loss function L is defined as follows: Where λ1 and λ2 are weighting factors, r represents the average time to complete the task. timeout This indicates the task timeout rate.

10. The single gantry crane scheduling system based on intelligent optimization algorithm according to claim 1, characterized in that, In the event-driven control module, when a task arrives, urgency recalculation and A* global replanning are triggered; when a periodic refresh event is triggered, the entire process scheduling is forced to refresh at a fixed period; when an abnormal event is triggered, local replanning is triggered and tasks in the fault area are eliminated; when a task completes, the gantry crane position is updated and the remaining task set is re-evaluated.

Images