CN117114313A - An AGV group scheduling method based on demand task prediction model - Google Patents

Abstract

The invention discloses an AGV group scheduling method based on a demand task prediction model, which is suitable for workshop transportation tasks. Aiming at the actual situations of large workshop transportation demand and low AGV transportation efficiency, the invention accurately and rapidly processes the distribution of the transportation tasks by introducing the transportation task demand prediction model, and schedules a plurality of AGVs to finish a plurality of transportation demand tasks. According to the AGV scheduling method based on the demand task prediction model, the prediction model of the demand task is built, a plurality of parallel task groups are built by combining the priorities of the issued plurality of demand tasks, and the distribution of the sub-priority tasks is completed through the demand task prediction database, so that the orderly, efficient and real-time parallel execution of the plurality of task groups is realized, the problems that the conventional scheduling means are prone to occurrence of task issuing delay and task congestion are effectively solved, and the efficiency optimization of the AGV scheduling system in automatic storage is realized.

Description

An AGV group scheduling method based on demand task prediction model

Technical field

The invention relates to an automated warehousing AGV resource scheduling method, and in particular to an AGV group scheduling method based on a demand task prediction model.

Background technique

Automated warehousing is an intelligent logistics warehousing management system, digital and intelligent storage management resources. AGV (automatic moving vehicle) is the most common automated warehousing execution terminal, and its management and scheduling is the most important part of automated warehousing. With the rapid development of the production and manufacturing fields, AGV groups have gradually become the mainstream direction of automated warehousing development due to the strong ability of multiple AGVs to perform tasks and quick environmental response. In recent years, against the background of the booming development of big data and artificial intelligence technology, the workload of tasks has increased dramatically, the working environment of AGV groups has become more and more complex, and traditional scheduling algorithms can no longer meet current needs.

Currently, AGV groups have two scheduling methods: centralized control and distributed control. In the centralized scheduling control method, the AGV group is uniformly scheduled through only one dispatch center; in distributed scheduling control, the AGV accepts tasks and plans by itself. route. Patents "A task allocation method based on AGV scheduling system" (CN115657616), "Multi-AGV scheduling method and device and electronic equipment based on graph convolutional neural network" (CN113253684), "A multi-AGV motion planning method, device and System" (CN112015174) and other systems enable the AGV group to accurately, safely and efficiently perform handling operations under a centralized dispatch control method, improving the performance of the AGV group's motion planning in a dynamic environment.

However, the above-mentioned scheduling methods for AGV groups lack flexibility in processing demand tasks. The processing of demand tasks generally adopts a first-come, first-process strategy, that is, when the demand terminal releases a task, an AGV is scheduled according to its content. Execution, tasks are executed only based on the release time, there is no distinction between task priorities, AGV resource utilization is not high, and work efficiency is low. On the one hand, when there are too many tasks released by the demand terminal, traditional scheduling methods are prone to task release delays and task congestion, seriously affecting work efficiency; on the other hand, when multiple urgent tasks, that is, high-priority tasks, are released, AGV The group cannot react quickly and flexibly.

Contents of the invention

Purpose of the invention: In response to the above problems, the purpose of the present invention is to provide an AGV group scheduling method based on a demand task prediction model, to solve the delay and congestion problems of demand tasks, and to achieve efficiency optimization of the AGV scheduling system in automated warehousing.

Technical solution: An AGV group scheduling method based on a demand task prediction model, including the following steps:

Step 1: Collect samples of demand tasks in the work environment and build an auxiliary demand calculation model, including:

S11: Establish a workspace model, including scene map model and AGV state model;

S12: Establish path calculation model and energy loss model;

Step 2: For the tasks released by the demand terminal, first send the task information to the auxiliary demand calculation model, and output the AGV's current status information table, the optimal AGV number and its path mode; when the AGV working terminal performs the task, the demand terminal is used Based on the historical task data, the demand task prediction model is obtained through BP neural network training, predicts the release of subsequent terminal tasks, and outputs the predicted number of subsequent terminal tasks;

Step 3: Based on the real-time task group, perform priority preprocessing on the predicted terminal tasks and send them to the auxiliary demand calculation model, and determine whether the solution is a single AGV solution or a multi-AGV solution based on task priority and content selection;

Step 4: Check the AGV's working status information table, including working status, estimated idle time, and estimated remaining power of the AGV. Evaluate whether the current scheduling plan can meet the tasks of the demand terminal in the AGV state at this stage. If satisfied, the task will be executed; if If a new demand terminal task appears and all AGVs are busy, update the priority preprocessing in step 3, update the parallel task group, and then perform subsequent scheduling processing; if the optimal AGV remaining power output by the prediction task is less than the energy consumption of the output path mode , then update the AGV working status information table in step 1, and then perform step 3 to re-distribute tasks and schedule AGVs.

This method aims at the problems of large workshop transportation demand, low AGV transportation efficiency, task release delay, and task congestion. The AGV scheduling method based on the demand task prediction model combines artificial intelligence with the automated warehousing scheduling system, and establishes a prediction model of demand tasks. , combine the priorities of multiple released demand tasks to construct multiple parallel task groups, and then complete the allocation of sub-priority tasks through the demand task prediction database to achieve orderly, efficient and real-time parallel execution of multiple task groups, effectively It improves the problems of task release delay and task congestion that are prone to occur with traditional scheduling methods, and optimizes the performance of the AGV scheduling system in automated warehousing.

Further, in step S11, to establish a scene map model, it is necessary to establish a demand task target area based on the actual working environment, including M demand terminals and N AGV working terminals, in which the demand terminal position is fixed and the AGV standby position is fixed.

Further, in step S11, an AGV status model is established. According to the actual working conditions, before dispatching the AGV, an AGV status information table is established through the AGV's own positioning system, software working status feedback flag, and software power monitoring data, including AGV number, working status, and AGV current battery level.

Further, in step S12, a path calculation model is established, and it is defined that every time T, a demand terminal generates a demand task to be executed, and the task is preprocessed and entered into the parallel task group or candidate task library; each task has k +1 path mode, that is, choose not to process or choose any numbered AGV from 1 to k to process, represented by K={0, 1,...,k+1}, and the corresponding path length is L={L ₀ , L ₁ ,···,L _k } represents.

Optimally, in step S12, at the standard driving speed, the energy loss of AGV is Expressed as:

where m is the weight carried by the AGV, g is the acceleration of gravity, L _k is the distance traveled by the AGV in the k-th path mode, v is the standard driving speed of the AGV, and P _task represents the loss incurred by the operation after the AGV reaches the designated position of the task. energy of;

Collect the historical electric power of AGVs with different numbers before and after performing tasks under different loads and modes to form an energy loss model.

Further, in step 2, training the demand task prediction model through BP neural network includes the following steps:

S1: Obtain the historical task data released by the demand terminal, including the priority of the demand task, the frequency of demand task release and the number of AGVs required, and use the historical task data as a training sample;

S2: Take 80% of the training samples to form the training set R _p , and the remaining 20% as the test set _Sp ; use the mapminmax function to normalize the data of the training set R _p and the test set _Sp ;

S3: Build BP neural network:

S4: Use the trained model to simulate the test set S _p , and output the prediction results after denormalization.

Optimally, in step S3, the construction of the BP neural network includes the following steps:

S3.1: Establish a three-layer BP neural network, which is the input layer, hidden layer and output layer. The number of nodes in the input layer is determined by the number of training samples; the number of nodes in the output layer is 1, and the prediction of the next moment is output. The number of the required task; determine the number of hidden layer nodes, use a loop to traverse the hidden layer nodes and training error conditions within the range, find the smallest error, and follow The empirical formula is calculated and rounded, where a represents the number of nodes in the input layer, b represents the number of nodes in the output layer, and c is any constant between [1,10];

S3.2: Set the number of training times to N _n , the learning rate to t, the training minimum mean square error target to s, and the training minimum performance gradient to s′; during the training process, the neuron activation functions all use the tansig function and Bayesian Regularization algorithm as training algorithm;

S3.3: Carry out BP neural network training.

Further, in step 3, when the single AGV solution is selected, the demand terminal releases a single high-priority or sub-priority demand task. At this time, the demand task is preprocessed and the task is processed directly without the need for a database. match.

Further, in step 3, when selecting the multi-AGV solution, the demand terminal releases two or more demand tasks with the same priority or different priorities. At this time, after the demand tasks are preprocessed, multiple parallel high-priority tasks are first constructed. group, and send the demand tasks into the auxiliary demand calculation model, and then complete the allocation of sub-priority tasks based on the historical completion time of the demand tasks, historical energy consumption, and AGV status information tables to form a complete parallel task group.

Optimally, in step 4, information is exchanged between the terminal and the AGV through a communication module. The communication module includes a wireless access point and a wireless serial server 1. The AGV has a built-in wireless serial server 2. The wireless access point is installed in the work space. within, thereby ensuring the reliability of communication within the entire dispatching system.

Beneficial effects: Compared with the existing technology, the advantages of the present invention are: by introducing a transportation task demand prediction model to predict demand tasks, considering the AGV working status table, and constructing multiple parallel tasks based on the priorities of multiple released demand tasks. groups, and then complete the allocation of sub-priority tasks through the demand task prediction database, thereby achieving orderly, efficient and real-time parallel execution of multiple task groups, which can accurately and quickly handle the allocation of transportation tasks and schedule multiple AGVs to complete multiple tasks. Transportation requirements tasks. By predicting and assigning tasks to designated AGVs in advance, the handling efficiency of AGVs is improved. When a single batch of tasks is large, it effectively improves the problems of task release delay and task congestion that traditional scheduling methods are prone to, and achieves efficiency optimization of the AGV scheduling system in automated warehousing.

Description of drawings

Figure 1 is a schematic diagram of the method steps of the present invention;

Figure 2 is a flow chart of AGV scheduling based on task prediction results in the present invention;

Figure 3 is an example of a smart warehousing scenario.

Detailed ways

The present invention will be further clarified below with reference to the accompanying drawings and specific examples. It should be understood that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

The present invention proposes an AGV group scheduling method based on a demand task prediction model, which is suitable for automated workshop transportation tasks. AGV serves as a transportation task execution terminal and realizes efficiency optimization of the AGV scheduling system in automated warehousing by establishing a prediction model of demand tasks. As shown in Figure 1, multiple released demand tasks are first preprocessed, that is, the demand tasks are prioritized. Tasks with high priority are constructed into multiple parallel task groups, and those with lower priority are put into the candidate task library. Then, through the demand task prediction model database, the allocation of sub-priority tasks is completed based on the demand task historical completion time, historical energy consumption, and AGV status information table to achieve orderly, efficient, and real-time parallel execution of multiple task groups, improving It improves the utilization rate of the AGV group; at the same time, based on the tasks and time training released by the historical demand terminal, a demand prediction model is obtained to predict the release of subsequent terminal tasks, which improves transportation efficiency. The AGV scheduling flow chart based on task prediction results in the present invention is shown in Figure 2. First, the work space model, path calculation model and energy loss model are modeled to build an auxiliary demand calculation model; secondly, the tasks and time released by the historical demand terminal are After training, the demand prediction model is obtained and the predicted demand tasks at the next moment are output; then, the predicted demand tasks are preprocessed, and an AGV pre-scheduling plan is formed based on the energy consumption of the tasks and the AGV status; then, when the real demand tasks arrive and Compare the predicted demand tasks. If the prediction is correct, enter the AGV detection stage according to the predicted scheduling plan. If the prediction is wrong, re-execute the task scheduling and obtain a new scheduling plan. Finally, check the AGV's working status information table and evaluate the current scheduling plan for the AGV at this stage. Whether the task of the demand terminal can be satisfied in the state, if so, it enters the task execution stage, and the demand terminal controls the AGV to execute the task through the communication module; otherwise, the task scheduling is re-executed to obtain a new scheduling plan.

Information interaction between the terminal and the AGV is required through the communication module. The communication module includes a wireless access point and a wireless serial port server. The AGV has a built-in wireless serial port server 2. The wireless access point is installed in the work space to ensure the communication of the entire dispatching system. reliability.

The specific steps of the scheduling method are as follows:

Step 1: Collect samples of demand tasks in the work environment and build an auxiliary demand calculation model, specifically:

Collect samples of demand tasks in the work environment, that is, daily behavior data of demand terminals, and establish auxiliary demand calculation models, including work space models, path calculation models and energy loss models;

(1) The daily behavior data collected in the present invention are shown in the table below.

Task number Task priority Task release frequency Task release time Task end time AGV quantity

(2) Establish a workspace model, including scene map model and AGV status model

Map model:

According to the actual working environment, the demand task work target area is established, including M demand terminals and N AGV working terminals, in which the demand terminal position is fixed and the AGV standby position is fixed;

AGV state model:

According to the actual working conditions, before dispatching the AGV, use the AGV's own positioning system, software working status feedback flag, and software power monitoring data to establish an AGV status information table, including AGV number, working status (idle or not), and AGV's current power level. .

(3) Establish path calculation model and energy loss model

Path calculation model:

It is defined that every time T, a demand terminal generates a demand task to be executed. After preprocessing, the task enters the parallel task group or candidate task library; each task has k+1 path modes, that is, choose not to process or choose 1~ k Any numbered AGV processing is represented by K = {0, 1, ···, k+1}, and the corresponding path length is represented by L = {L ₀ , L ₁ , ···, L _k }.

Energy loss model:

During the entire task execution process, under different path modes, the AGV travels on different paths and consumes different energy; under the same path mode, the AGV carries different loads and consumes different energy; at standard driving speed, the energy of the AGV The loss is Expressed as

where m is the weight carried by the AGV, g is the acceleration of gravity, which is about 9.81m/s ² , L _k is the distance traveled by the AGV in the kth path mode, v is the standard driving speed of the AGV, and P _task indicates the arrival of the AGV The energy consumed by performing operations after the task specifies the location;

Collect the historical power of AGVs with different numbers before and after performing tasks under different loads and modes to form an energy loss model;

Step 2: Based on the tasks and time released by historical demand terminals, a demand prediction model is obtained to predict the release of subsequent terminal tasks;

For tasks issued by the demand terminal, the task information is first sent to the auxiliary demand calculation model, and the current status information table of the AGV, the optimal AGV number and its path mode are output; when the AGV working terminal performs the task, the historical tasks of the demand terminal are used Data, the demand task prediction model is obtained through BP neural network training, predicts the release of subsequent terminal tasks, and outputs the predicted number of subsequent terminal tasks;

The specific method of training the demand task model through BP neural network is:

S1, obtain the historical task data released by the demand terminal, including the demand task priority, the demand task release frequency and the required number of AGVs, and use the historical task data as a training sample;

S2, take 80% of the training samples to form the training set R _p , and the remaining 20% as the test set _Sp ; use the mapminmax function to normalize the data of the training set R _p and the test set _Sp ;

S3, build BP neural network

S3.1, BP neural network is divided into three layers, namely input layer, hidden layer and output layer. The number of nodes in the input layer is determined by the number of training samples; the number of nodes in the output layer is 1, and the number of the demand task predicted at the next moment is output; the number of nodes in the hidden layer is determined by using a loop to traverse the range. Hidden layer nodes and training error, so as to find the smallest error; according to The empirical formula is calculated and rounded, where a represents the number of nodes in the input layer, b represents the number of nodes in the output layer, and c is any constant between [1,10];

S3.2, set the training times to N _n , the learning rate to t, the training minimum mean square error target to s, and the training minimum performance gradient to s′; during the training process, the neuron activation functions all use the tansig function and Bayesian Regularization algorithm as training algorithm;

S3.3, perform BP neural network training;

S4, use the trained model to simulate the test set S _p , and output the prediction results after denormalization.

Step 3: Based on the real-time task group, preprocess the predicted terminal tasks, and then select a single AGV solution or a multi-AGV solution according to the task priority and content;

When the AGV executes a task, based on the real-time task group, the predicted terminal tasks are prioritized and sent to the auxiliary demand calculation model, and the status information table of the AGV under the predicted task, the optimal AGV number and its path mode are output; Among them, the single AGV solution or the multi-AGV solution is selected according to the task priority and content;

In the single AGV solution, the demand terminal only releases a single high-priority or sub-priority demand task. At this time, the demand task is preprocessed and processed directly;

In the multi-AGV scheme, the demand terminal releases two or more demand tasks with the same priority or different priorities. At this time, after the demand tasks are preprocessed, multiple parallel high-priority task groups are first constructed and the demand tasks are sent to step 1. In the established auxiliary demand calculation model, the allocation of sub-priority tasks is completed based on the historical completion time of demand tasks, historical energy consumption, and AGV status information tables to form a complete parallel task group.

Step 4: Check the AGV's working status information table, including working status, estimated idle time, and estimated remaining power of the AGV. Evaluate whether the current scheduling plan can meet the tasks of the demand terminal in the AGV state at this stage. If not, retry the task. Distribution and scheduling of AGVs.

If a new demand terminal task appears and all AGVs are busy, update the priority preprocessing in step 3, update the parallel task group, and then perform subsequent scheduling processing; if the optimal AGV remaining power output by the prediction task is less than the output path mode energy consumption, update the AGV working status information table in step 1, and then perform step 3 to re-allocate tasks and schedule AGVs.

In the following embodiments, the application scenarios and simulation results of the present invention are described.

The AGV group scheduling method of the present invention was tested in the intelligent warehousing scenario as shown in Figure 3, and was implemented based on Anylogic software. This scene includes AGV parking area 1, storage shelf area 2, and container unloading area 3. Area 1 has several AGVs, and the demand terminal issues multiple cargo handling tasks, which need to be moved from container unloading area 3 to storage shelf area 2.

The result is that when the demand terminal releases 100 high-priority and 200 sub-priority tasks at a lower frequency, the average task completion time of the traditional strategy is 9.18 minutes, while the average task completion time of the AGV group scheduling method of the present invention is 7.77 minutes. When terminals are required to perform the same amount of tasks at a higher frequency, the average task completion time of the traditional strategy is 7.47 minutes, while the average task completion time of the AGV group scheduling method of the present invention is 6.85 minutes.

The above are only some embodiments of the present invention. It should be pointed out that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention. These improvements and modifications can also be made. should be regarded as the protection scope of the present invention.

Claims (10)

1. An AGV group scheduling method based on a demand task prediction model, which is characterized by including the following steps: Step 1: Collect samples of demand tasks in the work environment and build an auxiliary demand calculation model, including: S11: Establish a workspace model, including scene map model and AGV state model; S12: Establish path calculation model and energy loss model; Step 2: For the tasks released by the demand terminal, first send the task information to the auxiliary demand calculation model, and output the AGV's current status information table, the optimal AGV number and its path mode; when the AGV working terminal performs the task, the demand terminal is used Based on the historical task data, the demand task prediction model is obtained through BP neural network training, predicts the release of subsequent terminal tasks, and outputs the predicted number of subsequent terminal tasks; Step 3: Based on the real-time task group, perform priority preprocessing on the predicted terminal tasks and send them to the auxiliary demand calculation model, and determine whether the solution is a single AGV solution or a multi-AGV solution based on task priority and content selection; Step 4: Check the AGV's working status information table, including working status, estimated idle time, and estimated remaining power of the AGV. Evaluate whether the current scheduling plan can meet the tasks of the demand terminal in the AGV state at this stage. If satisfied, the task will be executed; if If a new demand terminal task appears and all AGVs are busy, update the priority preprocessing in step 3, update the parallel task group, and then perform subsequent scheduling processing; if the optimal AGV remaining power output by the prediction task is less than the energy consumption of the output path mode , then update the AGV working status information table in step 1, and then perform step 3 to re-distribute tasks and schedule AGVs. 2. An AGV group scheduling method based on a demand task prediction model according to claim 1, characterized in that in step S11, to establish a scene map model, it is necessary to establish a demand task work target area according to the actual working environment, including There are M demand terminals and N AGV working terminals, among which the demand terminal position is fixed and the AGV standby position is fixed. 3. An AGV group scheduling method based on a demand task prediction model according to claim 1, characterized in that, in step S11, an AGV state model is established, and according to the actual working conditions, before scheduling the AGV, the AGV comes with The positioning system, software working status feedback flag, and software power monitoring data are used to establish an AGV status information table, including AGV number, working status, and AGV current power. 4. An AGV group scheduling method based on a demand task prediction model according to claim 1, characterized in that: in step S12, a path calculation model is established to define that every time T, a demand terminal generates a task to be executed. Demand tasks, after task preprocessing, enter the parallel task group or candidate task library; each task has k+1 path modes, that is, choose not to process or choose any numbered AGV from 1 to k to process, use K = {0, 1,···,k+1}, and the corresponding path length is represented by L={L ₀ , L ₁ ,···,L _k }. 5. A kind of AGV group scheduling method based on demand task prediction model according to claim 4, characterized in that, in step S12, under the standard driving speed, the energy loss of AGV is Expressed as: where m is the weight carried by the AGV, g is the acceleration of gravity, L _k is the distance traveled by the AGV in the k-th path mode, v is the standard driving speed of the AGV, and P _task represents the loss incurred by the operation after the AGV reaches the designated position of the task. energy of; Collect the historical electric power of AGVs with different numbers before and after performing tasks under different loads and modes to form an energy loss model. 6. A kind of AGV group scheduling method based on demand task prediction model according to claim 1, characterized in that, in step 2, training the demand task prediction model through BP neural network includes the following steps: S1: Obtain the historical task data released by the demand terminal, including the priority of the demand task, the frequency of demand task release and the number of AGVs required, and use the historical task data as a training sample; S2: Take 80% of the training samples to form the training set R _p , and the remaining 20% as the test set _Sp ; use the mapminmax function to normalize the data of the training set R _p and the test set _Sp ; S3: Build BP neural network: S4: Use the trained model to simulate the test set S _p , and output the prediction results after denormalization. 7. A kind of AGV group scheduling method based on demand task prediction model according to claim 6, characterized in that, in step S3, the construction of BP neural network includes the following steps: S3.1: Establish a three-layer BP neural network, which is the input layer, hidden layer and output layer. The number of nodes in the input layer is determined by the number of training samples; the number of nodes in the output layer is 1, and the prediction of the next moment is output. The number of the required task; determine the number of hidden layer nodes, use a loop to traverse the hidden layer nodes and training error conditions within the range, find the smallest error, and follow The empirical formula is calculated and rounded, where a represents the number of nodes in the input layer, b represents the number of nodes in the output layer, and c is any constant between [1, 10]; S3.2: Set the number of training times to N _n , the learning rate to t, the training minimum mean square error target to s, and the training minimum performance gradient to s′; during the training process, the neuron activation functions all use the tansig function and Bayesian Regularization algorithm as training algorithm; S3.3: Carry out BP neural network training. 8. An AGV group scheduling method based on a demand task prediction model according to claim 1, characterized in that in step 3, when a single AGV solution is selected, the demand terminal issues a single high-priority or sub-priority demand. Task, at this time, the task is required to be preprocessed and processed directly without database matching. 9. An AGV group scheduling method based on a demand task prediction model according to claim 1, characterized in that in step 3, when selecting a multi-AGV solution, the demand terminal releases two or more with the same priority or different priorities. level demand tasks. At this time, after the demand tasks are preprocessed, multiple parallel high-priority task groups are first constructed, and the demand tasks are sent to the auxiliary demand calculation model, and then based on the demand task historical completion time, historical energy consumption, AGV The status information table completes the allocation of sub-priority tasks and forms a complete parallel task group. 10. An AGV group scheduling method based on a demand task prediction model according to claim 1, characterized in that, in step 4, information interaction is performed between the demand terminal and the AGV through a communication module, and the communication module includes wireless access Point and wireless serial server 1, AGV built-in wireless serial server 2, wireless access point installed in the work space.