RGV shuttle vehicle multi-task scheduling optimization method and system
By training a joint prediction model and optimizing RGV scheduling with dynamic heuristics, and combining ant colony optimization and execution bias feedback, the congestion and energy consumption problems in RGV scheduling are solved, and an efficient and robust scheduling scheme is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- KUNSHAN CIMC AUTOMATIC LOGISTIC EQUIP CO LTD
- Filing Date
- 2025-09-03
- Publication Date
- 2026-08-04
AI Technical Summary
Existing RGV shuttle scheduling algorithms suffer from congestion hotspots in real-world scenarios with multiple vehicles operating concurrently. Furthermore, they fail to comprehensively consider energy consumption costs, task priorities, and the system's ability to respond to unforeseen events, resulting in scheduling schemes that lack robustness and cost-effectiveness.
By training a joint prediction model to predict congestion risk and energy consumption cost, dynamic heuristic information is constructed. This information is combined with an ant colony algorithm to generate scheduling schemes. Closed-loop feedback correction is achieved by executing deviation vectors, thereby establishing self-learning and optimization capabilities.
It improves the actual execution efficiency, robustness, and economy of the RGV scheduling scheme, enabling efficient, stable, and economical material scheduling in complex and ever-changing industrial environments.
Smart Images
Figure CN121091818B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent manufacturing and logistics automation technology. More specifically, this invention relates to a method and system for multi-task scheduling optimization of RGV shuttles. Background Technology
[0002] RGV (Rail Guided Vehicle) shuttles are core material handling equipment in automated warehouses and flexible production lines. They are responsible for accurately and efficiently transporting materials to designated workstations, buffer areas, or inbound / outbound stations according to instructions on a pre-set track network. The efficiency and stability of the RGV scheduling system directly determine the material flow speed, equipment utilization rate, and order delivery capability of the entire production system, and are key indicators for measuring the automation level of a smart factory.
[0003] To address the RGV scheduling problem in complex production environments, ant colony optimization (ACO) algorithms are commonly employed. This algorithm simulates the collective intelligent behavior of ants, which use pheromones to find the shortest path, exhibiting good distributed computing characteristics and robustness. For example, Chinese patent document CN110554673B discloses a scheduling method and apparatus for an intelligent RGV processing system. It reveals a method that uses a genetic algorithm to globally optimize the workpiece loading sequence, and then utilizes an ant colony optimization algorithm to optimize the specific RGV path during the transfer process between information nodes. This complementary use of the two algorithms improves the optimization capability to a certain extent.
[0004] However, the scheduling decisions of these algorithms are mostly based on static or idealized information, such as considering only fixed track distances as the basis for path selection, while ignoring the dynamic nature of system operation. In real-world scenarios with multiple vehicles operating concurrently, such idealized optimal paths often become congestion hotspots, resulting in actual operating efficiency far lower than theoretical expectations. Furthermore, existing models typically use the shortest total time or shortest total distance as a single optimization objective, failing to comprehensively consider energy consumption costs, task priorities, and the system's ability to cope with unexpected events such as emergency order insertions or equipment failures, leading to a lack of robustness and economy in the scheduling scheme. Summary of the Invention
[0005] To address the aforementioned technical problem of poor multi-task scheduling performance of RGV shuttles, the present invention provides solutions in the following aspects.
[0006] In a first aspect, the present invention provides a multi-task scheduling optimization method for RGV shuttles, comprising:
[0007] Historical operational data is acquired and a joint prediction model is trained. This model is used to predict the congestion risk and energy cost of candidate scheduling schemes in response to these schemes. Dynamic heuristics are constructed for the ant colony algorithm. These heuristics are positively correlated with the expected business revenue of the task cluster and negatively correlated with the congestion risk predicted by the joint prediction model and the energy cost with an added energy aversion coefficient. The energy aversion coefficient is exponentially inversely proportional to the ratio of the product of the current RGV power and the full charge power. An ant colony algorithm incorporating the dynamic heuristics is used to generate and execute a scheduling scheme. The execution deviation vector between the actual execution result and the predicted result of the scheduling scheme is calculated, and the joint prediction model is corrected online using this deviation vector.
[0008] This invention trains a joint prediction model capable of forecasting future congestion risks and energy costs, and uses this model to construct dynamic heuristic information. This transforms the decision-making basis of the ant colony algorithm from static physical distance to prediction of the future dynamic system state. Simultaneously, by introducing an energy aversion coefficient related to electricity consumption, intelligent perception of the RGV's energy status is achieved. This invention establishes a closed-loop feedback mechanism from execution results to model correction, using execution deviation vectors to correct the prediction model online. This enables the system to continuously learn and optimize, thereby improving the actual execution efficiency, robustness, and economy of the scheduling scheme.
[0009] Preferably, training the joint prediction model includes: extracting completed task clusters and their corresponding RGV, actual path, actual completion time, actual energy consumption, and system resource occupancy density during execution from historical operation data to form a training dataset; using the feature vectors of the task clusters and the feature vectors of the planned paths as model inputs, using the difference between the actual completion time and the theoretical completion time as a congestion risk label, and using the actual energy consumption as an energy cost label, and training the model using a multi-objective regression algorithm.
[0010] This invention utilizes real historical data, including actual routes, times, energy consumption, and system resource occupancy density, for training. This data-driven approach enables the model to learn the complex, nonlinear system operating patterns hidden behind the data, such as congestion patterns of specific road segments during specific time periods. Therefore, its predictions of congestion risk and energy costs are more accurate, providing a reliable data foundation for the subsequent construction of dynamic heuristic information, which is a prerequisite for the effectiveness of the entire method.
[0011] Preferably, the dynamic heuristic information satisfies the expression:
[0012] ;
[0013] In the formula, This represents the dynamic heuristic information for the k-th RGV to select the m-th task cluster; This represents the number of tasks in the m-th task cluster; In the m-th task cluster, the first task is the first task. The static priority of each task; This represents the dynamic value of the j-th task over time, which is a normalized value of the reciprocal of the remaining completion time of the j-th task. , This represents the normalized value of the congestion risk and energy cost generated by the k-th RGV executing the m-th task cluster, as predicted by the joint prediction model. Let represent the energy aversion coefficient of the k-th RGV at time t.
[0014] Preferably, the energy aversion coefficient satisfies the expression:
[0015] ;
[0016] In the formula, Let represent the energy aversion coefficient of the k-th RGV at time t; This is the adjustment constant; This represents the current remaining power of the k-th RGV; This represents the full charge level of the RGV. This represents the natural exponential function.
[0017] This invention uses an exponential function to make the energy aversion coefficient nonlinearly inversely proportional to the remaining power of the RGV; this means that when the RGV has sufficient power, it is not sensitive to energy consumption; but when the power decreases, its aversion to energy consumption will increase sharply, which can prevent the RGV from shutting down due to power depletion and enhance the operational stability and reliability of the entire system.
[0018] Preferably, the method further includes constructing a spatiotemporal resource network and a task association graph; the spatiotemporal resource network is used to plan conflict-free paths when executing the scheduling scheme; the task association graph is used to divide the tasks to be assigned into multiple task clusters.
[0019] Preferably, the association strength of the task association graph satisfies the expression:
[0020] ;
[0021] In the formula, This indicates the strength of the association between task i and task j; This indicates an order association indicator function; if task i and task j have the same order number, then... =1, otherwise =0; , Indicates the start and end points of task i. , Indicates the start and end points of task j; , This represents the physical path distance from the starting point of task i to the ending point of task j, and from the starting point of task j to the ending point of task i. This represents the maximum path distance within the warehouse; , This represents the dynamic urgency of tasks i and j at the current time t, where the dynamic urgency is the reciprocal of the remaining time. Represents the natural exponential function; Represents the maximum value function; It is a local minimum.
[0022] This invention comprehensively considers three dimensions: order association, spatial overlap and proximity, and time urgency similarity. It can more accurately identify tasks that are suitable for bundled execution in terms of business logic, physical path and time window, thereby forming more efficient and reasonable task clusters and laying a solid foundation for improving overall scheduling efficiency.
[0023] Preferably, the edge weights of the spatiotemporal resource network satisfy the expression:
[0024] ;
[0025] In the formula, Represents spatiotemporal resource nodes arrive Directed edge The weights; Indicates that RGV is through directed edges The required time cost; Indicates that RGV is through directed edges Required energy costs; This indicates the preset baseline energy consumption value.
[0026] Preferably, the pheromone update of the ant colony algorithm further includes a correction term based on the system disturbance amount, which is positively correlated with the difference between the current scheduling scheme and the previous round's optimal scheduling scheme, and negatively correlated with the correction term.
[0027] Preferably, the execution deviation vector includes the following components:
[0028] The execution deviation vector includes total time deviation, congestion risk prediction deviation, and energy consumption prediction deviation;
[0029] The total time deviation is obtained by subtracting the planned completion time from the actual completion time; the congestion risk prediction deviation is equal to the residual time deviation in the total time deviation that exceeds the congestion risk predicted by the joint prediction model; the energy consumption prediction deviation is obtained by subtracting the energy consumption predicted by the joint prediction model from the actual energy consumption.
[0030] This invention decomposes the total deviation into multiple components such as congestion risk prediction deviation and energy consumption prediction deviation, enabling the system to accurately pinpoint the root cause of poor performance. This provides a clear reason for subsequent online correction of the joint prediction model, improving the efficiency and accuracy of the system's self-learning and correction.
[0031] Secondly, the present invention provides an RGV shuttle multi-task scheduling optimization system, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned RGV shuttle multi-task scheduling optimization method is implemented.
[0032] By adopting the above technical solution, a computer program is generated from the above-mentioned RGV shuttle multi-task scheduling optimization method and stored in the memory so that it can be loaded and executed by the processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.
[0033] The beneficial effects of this invention are as follows:
[0034] (1) This invention constructs an RGV intelligent scheduling system with prediction, multi-objective optimization and closed-loop self-learning capabilities. It predicts future congestion risks and energy costs by training a joint prediction model and integrates it into the dynamic heuristic information of the ant colony algorithm, thereby realizing an early response to dynamic changes in the system.
[0035] (2) By decomposing the execution deviation vector, the present invention establishes a precise feedback mechanism that can correct the prediction model online, enabling the system to have the ability to continuously self-evolve, thereby achieving more efficient, robust and economical material scheduling in complex and ever-changing industrial environments. Attached Figure Description
[0036] Figure 1 This is a flowchart illustrating an RGV shuttle multi-task scheduling optimization method according to the present invention;
[0037] Figure 2 This is a schematic diagram illustrating the spatiotemporal resource network structure.
[0038] Figure 3 This is a schematic diagram illustrating the physical layout of the RGV operating environment. Detailed Implementation
[0039] This invention discloses a multi-task scheduling optimization method for RGV shuttles, referring to... Figure 1 This includes steps S1-S4:
[0040] S1: Obtain historical operational data of the RGV system and construct a spatiotemporal resource network and task association graph.
[0041] It should be noted that, in order to achieve efficient collaborative scheduling of RGVs, a comprehensive and accurate digital model of the RGV system's physical constraints, historical operational patterns, and the inherent logic between tasks is first required. This invention constructs two interrelated graph models: a spatiotemporal resource network and a task association graph. For example... Figure 2 This is a diagram of the spatiotemporal resource network structure.
[0042] Specifically, obtain historical operational data from the RGV scheduling system, including:
[0043] First, obtain the technical parameters of all physical equipment, including RGVs, elevators, track sections, intersections, and charging stations, as well as the topological connections of the physical track network, which together constitute the physical environment and equipment data. For example... Figure 3 This is a schematic diagram of the physical layout of the RGV operating environment.
[0044] Then, retrieve the historical task logs that record all previously issued handling tasks. Each task includes a task ID, associated order number, material information, starting location, target location, creation time, expected completion time, and static priority.
[0045] Simultaneously, historical status logs of all RGVs and key equipment, recorded at a fixed time frequency, are obtained. These historical status logs include location, speed, cargo status, remaining battery power, and occupancy status. For example, the fixed time frequency can be set to 1 second, and the key equipment may be, for example, an elevator.
[0046] Finally, historical event logs were retrieved, which recorded all unexpected system disturbances in history, such as emergency task insertions, equipment failures, and network communication interruptions.
[0047] Preferably, based on the physical orbit network in the physical environment and equipment data, a spatiotemporal resource network is constructed, including:
[0048] The physical track network and key equipment are discretized into a series of resource nodes; the scheduling time domain is discretized into a series of time steps; the nodes of the spatiotemporal network are spatiotemporal resource nodes. , representing the time step Resource-consuming nodes In a spatiotemporal resource network, if RGV can retrieve resources from a resource node within a time step... Move to Then in all adjacent time steps Between them, there exists a line from arrive The directed edge.
[0049] It should be noted that in scheduling optimization, time and energy consumption are two core and interrelated cost dimensions. Simple weighted summation relies on manually set weights and lacks adaptability. Therefore, this invention uses time cost as the base cost and transforms energy consumption cost into a dimensionless multiplication factor, making the influence of energy consumption related to the baseline energy consumption, which is more in line with physical reality.
[0050] The weights of the directed edges satisfy the expression:
[0051] ;
[0052] In the formula, Represents spatiotemporal resource nodes arrive Directed edge The weights; Indicates that RGV is through directed edges The required time cost; Indicates that RGV is through directed edges Required energy costs; This represents a preset baseline energy consumption value. For example, the baseline energy consumption value is 500 joules.
[0053] Preferably, based on the historical task logs, a task association graph is constructed, including:
[0054] Using all tasks to be assigned as nodes, calculate the association strength between any two task nodes i and j. If the association strength is greater than a preset association threshold, connect the corresponding two task nodes to form an edge, and the weight of the edge is the association strength value. It should be noted that the preset association threshold is a threshold value used to determine whether the association strength between tasks is sufficient to form an associated edge. This value can be preset based on historical data statistical analysis or through experimental optimization. For example, the preset association threshold is set to 0.6.
[0055] It should be noted that the correlation between tasks is reflected in multiple dimensions. The correlation in business logic is the strongest and should be considered the fundamental factor. The similarity in physical proximity and temporal urgency jointly determine the convenience and necessity of collaborative execution. Therefore, this invention constructs a correlation strength calculation formula that integrates multiple factors in a product form.
[0056] The correlation strength satisfies the expression:
[0057] ;
[0058] In the formula, This indicates the strength of the association between task i and task j; This indicates an order association indicator function; if task i and task j have the same order number, then... =1, otherwise =0; , Indicates the start and end points of task i. , Indicates the start and end points of task j; , This represents the physical path distance from the starting point of task i to the ending point of task j, and from the starting point of task j to the ending point of task i. This represents the maximum path distance within the warehouse; , This represents the dynamic urgency of tasks i and j at the current time t, where the dynamic urgency is the reciprocal of the remaining time. Represents the natural exponential function; Represents the maximum value function; This is a minimum value used to avoid a denominator of 0. For example, .
[0059] In the formula, This indicates spatial proximity, which measures the ease of performing one task along the same route after completing another. This indicates the similarity in time urgency. The closer the value is to 1, the more similar the urgency of the two tasks are, and the more they need to be considered together.
[0060] A community detection algorithm is run on the task association graph to dynamically divide the tasks to be assigned into several task clusters. It should be noted that the community detection algorithm is an existing technology, such as the LPA algorithm, and will not be elaborated upon here.
[0061] Thus, historical operational data of the RGV system were obtained, and a spatiotemporal resource network representing the physical and temporal constraints of the system, as well as a task association graph representing the inherent logical relationships of the tasks, were constructed.
[0062] S2: Based on historical operational data, construct a joint prediction model of the risk and cost of the scheduling scheme; based on the joint prediction model, define heuristic information for the ant colony algorithm.
[0063] It should be noted that existing technologies use an idealized total operation time as the optimization objective, and their mathematical models ignore resource competition and uncertain delays caused by multiple vehicles operating in parallel. Theoretically, the scheduling scheme with the shortest time may, in actual implementation, cause severe congestion at key road sections or elevators, resulting in actual performance far worse than expected. In addition, this model does not consider the energy consumption cost of RGVs. Therefore, this invention constructs a data-driven joint prediction model that can simultaneously predict the potential congestion risk and energy consumption cost of scheduling schemes.
[0064] It should be further explained that in ant colony optimization (ACO) algorithms, heuristic information guides ants in choosing their direction at decision points and is crucial to algorithm performance. Traditional ACO algorithms typically rely on static information such as distance, lacking consideration of the overall task value and execution risk. Therefore, this invention establishes dynamic heuristic information that integrates expected returns, congestion risk, and energy costs as a core component of the improved ACO algorithm.
[0065] Specifically, based on historical operational data, a joint prediction model for the risks and costs of scheduling schemes is constructed, including:
[0066] From historical operational data, we extract all completed task clusters and their corresponding RGVs, actual paths, actual completion times, actual energy consumption, as well as the activity and resource density of other RGVs in the system during execution, which together constitute the model training dataset.
[0067] The difference between the theoretical completion time and the actual completion time of a task cluster is recorded as the congestion risk index of the task cluster.
[0068] A multi-objective regression model is used for training. The inputs are the feature vectors of the candidate task clusters and the feature vectors of the planned paths; the outputs are the expected congestion risk and expected energy cost of the task clusters; after training, a joint prediction model is obtained. The multi-objective regression model, such as gradient boosting decision trees or deep neural networks, is existing technology and will not be described in detail here.
[0069] Preferably, when an ant needs to select a task cluster from the set of candidate task clusters for RSV, dynamic heuristic information is defined for the ant, including:
[0070] It should be noted that heuristic information aims to determine the expected degree of choice; a good choice should possess the characteristics of high returns, low risk, and low cost. Therefore, this heuristic information is constructed by using the expected business returns of the task cluster as the numerator and the risks and energy consumption predicted by the joint prediction model as the denominator.
[0071] Dynamic heuristic information satisfies the expression:
[0072] ;
[0073] ;
[0074] In the formula, This represents the dynamic heuristic information for the k-th RGV to select the m-th task cluster; This represents the number of tasks in the m-th task cluster; In the m-th task cluster, the first task is the first task. The static priority of each task; This represents the dynamic value of the j-th task over time, which is a normalized value of the reciprocal of the remaining completion time of the j-th task. , This represents the normalized value of the congestion risk and energy cost generated by the k-th RGV executing the m-th task cluster, as predicted by the joint prediction model.
[0075] In the formula, Let represent the energy aversion coefficient of the k-th RGV at time t; This represents the expected total revenue from executing the m-th task cluster; The adjustment constant is based on ensuring that the coefficient has a significant, non-linear growth range as the charge decreases from full to empty. For example, The value is 2; This represents the current remaining power of the k-th RGV; This represents the full charge level of the RGV; when the RGV has sufficient charge, Approaching 1, Approaching 1; when the battery level decreases, Decrease The exponential increase makes RGV more sensitive to energy costs when making decisions.
[0076] Thus, dynamic heuristics for improving the ant colony algorithm were obtained.
[0077] S3: Adaptive ant colony algorithm is used to generate a cooperative scheduling scheme; dynamic replanning is performed based on system disturbance events.
[0078] It should be noted that existing technologies involve a black-box hybrid approach of multiple heuristic algorithms and rely heavily on manually set hyperparameters, leading to unstable algorithm performance and difficulty in adapting to environmental changes. To address this issue, this invention proposes an adaptive ant colony algorithm whose core parameters can be dynamically adjusted based on the real-time state of the system, thereby enhancing the algorithm's adaptability and robustness. The performance of the ant colony algorithm primarily depends on the calculation of the state transition probability, which is jointly determined by pheromone concentration and heuristic information. This invention not only utilizes the dynamic heuristic information defined in S2 but also improves the pheromone update mechanism and its influence in state transitions.
[0079] It should be further explained that an optimized scheduling scheme must not only consider its own execution efficiency but also assess its impact on the overall system stability. Therefore, this invention introduces consideration of the scheduling scheme's stability in the pheromone update stage of the ant colony algorithm, so that pheromone accumulation not only rewards high-performance schemes but also stable schemes that cause less disturbance to the system.
[0080] Specifically, an adaptive ant colony algorithm is used to generate a cooperative scheduling scheme, including:
[0081] Initialize the ant colony, where each ant represents an explorer of a potential scheduling scheme. For example, the number of ants is 100.
[0082] In each iteration, each ant selects from the clusters of tasks to be assigned for all RGVs in the system in turn; when an ant needs to select a task cluster from its set of candidate task clusters for an RGV, it calculates the state transition probability.
[0083] The state transition probability satisfies the expression:
[0084] ;
[0085] In the formula, This represents the probability that an ant will choose the m-th task cluster for the k-th RGV; This represents the pheromone concentration indicating the historical success of the k-th RGV executing the path corresponding to the m-th task cluster. This represents the dynamic heuristic information for the k-th RGV to select the m-th task cluster; Let represent the set of candidate task clusters for the k-th RGV; , An adaptive influence factor representing pheromones and heuristic information.
[0086] It should be noted that, , The system is dynamically adjusted based on its current total task load. For example, Set as the total system workload A monotonically decreasing function, while Set as A monotonically increasing function, where This can be quantified as the number of tasks to be assigned; when the system load... At lower levels, The value is relatively high. When the value is low, the state transition probability is more influenced by historical pheromones. Due to the influence of [various factors], the algorithm tends to use the optimal path validated in historical experience to ensure efficiency and stability; when the system load [is high / low], [the algorithm will continue to operate]. When the number of new tasks is high or there is a large influx of new tasks, The value is relatively low. When the value is higher, the state transition probability is more influenced by dynamic heuristics. Due to the influence of the current system state, the algorithm tends to use potential better solutions to deal with sudden and complex situations.
[0087] After all ants have constructed a complete scheduling scheme, the scheme is evaluated based on the comprehensive performance evaluation function, and the pheromones are updated.
[0088] It should be noted that traditional pheromone increment is only related to path length or objective function value, while the pheromone increment of this invention is a comprehensive index that not only rewards high-performing solutions but also those that contribute significantly to system stability.
[0089] The pheromone update satisfies the expression:
[0090] ;
[0091] ;
[0092] In the formula, Indicates in At time k, the pheromone concentration of the path corresponding to the k-th RGV and the m-th task cluster; This represents the pheromone evaporation coefficient, which is a constant. For example, It is 0.1; Indicates the pheromone increment; It is a plan The overall performance score; It is a plan Compared to the old plan The system disturbance quantity.
[0093] In the formula, The total completion time is represented by a weighted average; the larger the value, the smaller the pheromone increment. The smaller the value, the more stable the scheme; the smaller the value, the greater the pheromone increment. A scheme that performs well and is stable will receive more pheromone rewards.
[0094] After repeating the process a preset number of times, the historically best scheduling scheme is selected as the output of this round of scheduling.
[0095] Preferably, based on system disturbance events, the dynamic replanning of the scheme is triggered, including:
[0096] The system background continuously monitors the historical event logs in S1 for newly added system disturbance events; when a disturbance event is detected, a replanning is immediately triggered; when the adaptive ant colony algorithm is re-executed, the system disturbance amount in the pheromone update formula is adjusted. The dynamic adjustment of weights makes the algorithm more inclined to make local adjustments while maintaining system stability when dealing with disturbances. It should be noted that this dynamic adjustment is achieved through a weight coefficient. Implementation: During normal operation, the weighting coefficients Set to a lower default value, for example =1; when a disturbance event is detected and replanning is triggered, the weight coefficient is... It is immediately switched to a preset higher value, for example =10; After replanning is completed, in the next regular planning cycle, the weight coefficients will be... Restore it to its default value. This simple weight switching allows the algorithm to prioritize maintaining system stability when dealing with disturbances.
[0097] Thus, an adaptive, stable, and dynamic-responsive collaborative scheduling scheme has been obtained.
[0098] S4: Plan conflict-free paths for scheduling schemes in the spatiotemporal resource network and perform online model correction based on the execution deviation vector.
[0099] It should be noted that even if the scheduling scheme is optimal in terms of task order, conflicts and congestion can still occur if the path planning ignores the occupancy of spatiotemporal resources. Path planning through spatiotemporal resource networks can avoid conflicts in theoretical planning. Timeouts may be due to improper path planning, inaccurate energy consumption models requiring additional charging of the RGV midway, or the joint prediction model underestimating congestion in specific areas. Therefore, to achieve accurate model correction, this invention proposes the concept of an execution deviation vector. This vector decomposes the execution deviation from multiple dimensions, enabling the system to locate the source of error and update the corresponding upstream models.
[0100] Specifically, planning conflict-free paths for collaborative scheduling schemes in the spatiotemporal resource network includes:
[0101] Based on the RGV and task cluster sequence determined by the collaborative scheduling scheme, an execution path is planned for each RGV in turn. For the task cluster sequence of each RGV, the time-dependent A* algorithm is used to find the lowest-cost path in the spatiotemporal resource network. Once the path of the RGV is determined, the occupancy status of the corresponding spatiotemporal resource node in the spatiotemporal resource network is immediately updated, making it unavailable for subsequent path planning. It should be noted that the A* algorithm is existing technology and will not be elaborated upon here.
[0102] The complete scheduling instructions, including conflict-free paths, are issued to the RGV fleet for execution.
[0103] Preferably, the joint prediction model is corrected online based on the execution deviation vector of the scheduling scheme, including:
[0104] Monitor the execution status of tasks in real time and record the actual completion time of each task cluster. Actual energy consumption And the actual driving trajectory.
[0105] It should be noted that the execution bias vector aims to decompose the total execution error into different dimensions in order to trace the source of the error. The total time bias is a macroscopic result, while the congestion risk prediction bias and energy consumption prediction bias are direct tests of the model's predictive ability.
[0106] The execution deviation vector satisfies the expression:
[0107] ;
[0108] In the formula, This represents the execution deviation vector; Indicates the total time deviation, expressed in terms of actual completion time. Completion time as planned Obtained by subtraction; This represents the congestion risk prediction bias, which is equal to the portion of the total time bias that exceeds the congestion risk predicted by the joint prediction model. The residual time deviation; This indicates the deviation in energy consumption prediction, expressed through actual energy consumption. Energy consumption predicted by the joint prediction model Obtained by doing a difference.
[0109] Using the execution bias vector as the key label for new training samples, the parameters of the joint prediction model are incrementally updated online through an online learning algorithm. For example, when When the primary bias term is 'congestion risk,' the learning algorithm will prioritize adjusting the model weights in the joint prediction model that are relevant to congestion risk prediction. The online learning algorithm mentioned, such as stochastic gradient descent with a target decomposition loss function, is existing technology and will not be elaborated upon here.
[0110] This completes the closed loop from decision execution to model feedback, ensuring the continuous learning and self-improvement of the entire scheduling optimization system.
[0111] This invention also discloses an RGV shuttle multi-task scheduling optimization system, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, an RGV shuttle multi-task scheduling optimization method according to the present invention is implemented.
[0112] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.
[0113] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.
Claims
1. A multi-task scheduling optimization method for RGV shuttles, characterized in that, include: Historical operational data is acquired and a joint prediction model is trained. The joint prediction model is used to predict the congestion risk and energy consumption cost of the candidate scheduling scheme in response to the candidate scheduling scheme. Dynamic heuristic information is constructed for the ant colony algorithm. The dynamic heuristic information is positively correlated with the expected business benefits of the task cluster and negatively correlated with the congestion risk predicted by the joint prediction model and the energy cost with added energy consumption aversion coefficient. The energy consumption aversion coefficient is inversely proportional to the ratio of the product of the current RGV power and the full power. An ant colony algorithm incorporating the dynamic heuristic information is used to generate and execute a scheduling scheme; the execution deviation vector between the actual execution result and the predicted result of the scheduling scheme is calculated, and the joint prediction model is corrected online using the execution deviation vector; Dynamic heuristic information satisfies the expression: In the formula, This represents the dynamic heuristic information for the k-th RGV to select the m-th task cluster; This represents the number of tasks in the m-th task cluster; In the m-th task cluster, the first task is the first task. The static priority of each task; This represents the dynamic value of the j-th task over time, which is a normalized value of the reciprocal of the remaining completion time of the j-th task. , This represents the normalized value of the congestion risk and energy cost generated by the k-th RGV executing the m-th task cluster, as predicted by the joint prediction model. Let represent the energy aversion coefficient of the k-th RGV at time t; The energy aversion coefficient satisfies the following expression: In the formula, Let represent the energy aversion coefficient of the k-th RGV at time t; This is the adjustment constant; This represents the current remaining power of the k-th RGV; This represents the full charge level of the RGV. This represents the natural exponential function.
2. The RGV shuttle multi-task scheduling optimization method according to claim 1, characterized in that, Training the joint prediction model includes: extracting completed task clusters and their corresponding RGV, actual path, actual completion time, actual energy consumption, and system resource occupancy density during execution from historical operation data to form a training dataset; using the feature vectors of the task clusters and the feature vectors of the planned paths as model inputs, using the difference between the actual completion time and the theoretical completion time as a congestion risk label, and using the actual energy consumption as an energy cost label, and training the model using a multi-objective regression algorithm.
3. The RGV shuttle multi-task scheduling optimization method according to claim 1, characterized in that, It also includes constructing a spatiotemporal resource network and a task association graph; the spatiotemporal resource network is used to plan conflict-free paths when executing the scheduling scheme; the task association graph is used to divide the tasks to be assigned into multiple task clusters.
4. The RGV shuttle multi-task scheduling optimization method according to claim 3, characterized in that, The association strength of the task association graph satisfies the expression: ； In the formula, This indicates the strength of the association between task i and task j; This indicates an order association indicator function; if task i and task j have the same order number, then... =1, otherwise =0; , Indicates the start and end points of task i. , Indicates the start and end points of task j; , This represents the physical path distance from the starting point of task i to the ending point of task j, and from the starting point of task j to the ending point of task i. This represents the maximum path distance within the warehouse; , This represents the dynamic urgency of tasks i and j at the current time t, where the dynamic urgency is the reciprocal of the remaining time. Represents the natural exponential function; Represents the maximum value function; It is a local minimum.
5. The RGV shuttle multi-task scheduling optimization method according to claim 3, characterized in that, The edge weights of the spatiotemporal resource network satisfy the expression: ； In the formula, Represents spatiotemporal resource nodes arrive Directed edge The weights; Indicates that RGV is through directed edges The required time cost; Indicates that RGV is through directed edges Required energy costs; This indicates the preset baseline energy consumption value.
6. The RGV shuttle multi-task scheduling optimization method according to claim 1, characterized in that, The pheromone update of the ant colony algorithm also includes a correction term based on the system disturbance amount, which is positively correlated with the difference between the current scheduling scheme and the previous round's optimal scheduling scheme, and negatively correlated with the correction term.
7. The RGV shuttle multi-task scheduling optimization method according to claim 1, characterized in that, The execution bias vector includes the following components: The execution deviation vector includes total time deviation, congestion risk prediction deviation, and energy consumption prediction deviation; The total time deviation is obtained by subtracting the planned completion time from the actual completion time; the congestion risk prediction deviation is equal to the residual time deviation in the total time deviation that exceeds the congestion risk predicted by the joint prediction model; the energy consumption prediction deviation is obtained by subtracting the energy consumption predicted by the joint prediction model from the actual energy consumption.
8. A multi-task scheduling optimization system for RGV shuttles, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a multi-task scheduling optimization method for RGV shuttles according to any one of claims 1-7.