Heterogeneous resource fusion scheduling method based on multi-agent collaborative optimization

Abstract

The invention discloses a heterogeneous resource fusion scheduling method based on multi-agent collaborative optimization. The method comprises the following steps: collecting heterogeneous resource state data and task demand data to construct a resource state diagram and a task demand diagram; mapping each resource unit in the heterogeneous resource map into a multi-agent generation scheme ranking list; performing sorting comparison attribution on the scheme ranking list to generate a sub-path contribution table; reversibly updating the pheromone account book based on a bidirectional pledge sorting ant colony optimization algorithm to generate a pheromone account book state set; performing phase adaptive adjustment on the search process according to the pheromone account book state set; and verifying the pheromone account book state set according to the execution feedback data, and outputting an optimal scheduling scheme. According to the method, multi-agent cooperation and a bidirectional pledge ant colony optimization algorithm are fused, heterogeneous resource adaptive scheduling is realized, and the method has the advantages of convergence robustness, load balancing and result interpretability.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling, and in particular to a heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization. Background Technology

[0002] Existing heterogeneous resource scheduling methods mostly employ heuristics or ant colony-based global search, abstracting the relationship between resources and tasks into a graph structure. They guide path selection through pheromone accumulation and evaporation, supplemented by rule constraints to generate feasible solutions. In multi-agent scenarios, a common approach is to generate candidate scheduling schemes by combining local observations with global constraints, then score and rank them using indicators such as energy consumption, time windows, and capacity matching, iteratively approximating a better solution. This type of technology has a foundation in general manufacturing, transportation, and energy coordination, but it mainly remains at the level of rewarding the entire path or elite strategies, lacking fine-grained characterization of the sources of scheme differences.

[0003] However, traditional ant colony optimization reveals several shortcomings under heterogeneous resource collaboration: pheromone updates are mainly unidirectional accumulations, which are prone to premature convergence and difficult to undo misleading paths; whole-path rewards dilute the contributions of key sub-paths and key resource combinations, resulting in insufficient interpretability; the lack of reversible accounting based on execution feedback makes it impossible to separate and recover bonuses from staking; at the resource level, there is a lack of congestion suppression and liquidity return for bottleneck nodes, making the load prone to imbalance; phase scheduling lacks measurable criteria, and the exploration and consolidation switching lags, making it difficult to form a traceable closed loop from candidate solutions to the optimal solution.

[0004] Therefore, how to provide a heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization. This invention integrates multi-agent cooperation with bidirectional ant colony optimization algorithm to achieve adaptive scheduling of heterogeneous resources, and has the advantages of convergence robustness, load balancing and interpretable results.

[0006] A heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization according to an embodiment of the present invention includes the following steps: Collect heterogeneous resource status data and task requirement data, perform format unification and correlation modeling, construct resource status graph and task requirement graph, and obtain heterogeneous resource map and candidate task set; Each resource unit in the heterogeneous resource map is mapped to a multi-agent system, and a set of candidate scheduling schemes is generated based on local observations and global constraints to form a scheme ranking list. Perform sorting, comparison, and attribution on the solution ranking list, identify key sub-paths and key resource combinations, and generate a sub-path contribution table; Based on the bidirectional staking sorting ant colony optimization algorithm, the pheromone ledger is reversibly updated according to the sub-path contribution table. The pheromone of the effective path is recorded as the bonus share, and the path to be verified is recorded as the staking share. During the iteration, the pheromone is dynamically recycled or solidified according to the verification results, and a pheromone ledger state set is generated. Based on the pheromone ledger state set, the search process is phase-adaptive adjusted, switching between the exploration phase and the consolidation phase. Overflow tax is applied to bottleneck nodes to balance resource load, and scheduling execution instructions and execution feedback data are generated. The pheromone ledger state set is verified based on the execution feedback data. Staking shares that have not contributed are reclaimed, and bonus shares of stable sub-paths are solidified. The optimal scheduling scheme is then output.

[0007] Optionally, the generation of the heterogeneous resource map and candidate task set specifically includes: Collect heterogeneous resource status data from computing nodes, energy devices, transportation equipment and production equipment, and task requirement data from work orders, process plans and service requests. Perform timestamp unification, spatial coordinate unification and field definition unification operations on the two types of data to obtain a set of data to be standardized. The time base alignment is performed on the dataset to be normalized, and the range normalization is performed on the numerical fields to obtain intermediate results of time and numerical normalization. Spatial registration is performed on data containing spatial location fields in the dataset to be normalized, and then merged with intermediate results of temporal and numerical normalization to obtain a spatiotemporal normalization result set; Based on the spatiotemporal normalization result set, a resource capability representation is constructed for each resource unit, and a task requirement representation is constructed for each task instance, forming a resource capability representation set and a task requirement representation set; Based on the resource capability representation set and the task requirement representation set, a feasible matching determination is performed, and the resource and task pairs that meet the conditions are formed into a feasible relationship set. Based on the set of feasible relationships, a resource state graph is constructed, with resource units as nodes and substitutability, proximity and mutual exclusion relationships between resources as edges. The edge weights are determined by a weighted combination of state similarity, energy consumption difference and spatial distance, and the resource state graph is output. Based on the set of feasible relationships, a task requirement graph is constructed, with task instances as nodes and task dependencies, same-location merging, and time window conflicts as edges. The edge weights are determined by a weighted combination of the strength of temporal dependencies, the strength of same-location merging, and the strength of time window conflicts, and the task requirement graph is output. By jointly modeling the resource state graph, task requirement graph, and feasible relationship set, resource nodes and task nodes are uniformly encoded and cross-domain connections are established to form a heterogeneous resource graph that represents resource topology, task constraints, and resource-task mapping relationships. From this graph, task nodes with a correlation degree higher than a preset threshold are extracted to generate a candidate task set.

[0008] Optionally, the generation of the scheme ranking list specifically includes: A multi-agent mapping list is established based on the resource nodes in the heterogeneous resource graph. Each resource node corresponds to an agent entry, and each agent entry is associated with the resource capability representation in the resource capability representation set and the adjacency relationship in the resource state graph, thus forming the agent mapping list. For each task instance in the candidate task set, retrieve the feasible relationship set to obtain the matching resource and task pairs, and combine the task requirement representation in the task requirement representation set with the resource capability representation in the resource capability representation set to generate a task ready list. Based on the agent mapping list and the task ready list, according to the edge weights of the resource state graph and the task requirement graph, the local observation elements and system-level constraint elements of each agent entry are summarized to obtain the local observation set and the global constraint set. Under the constraints of the local observation set and the global constraint set, resource and task assignment combinations are generated around the candidate task set to form a candidate scheduling scheme set with task instances as units and resource assignment and execution order as content, and the resource consumption, energy consumption estimate and time window occupation corresponding to each candidate scheduling scheme are recorded. Calculate the capability satisfaction, energy consumption estimate, and time cost of each candidate scheduling scheme in the candidate scheduling scheme set. According to the preset weights, the capability satisfaction is used as a bonus item, and the energy consumption estimate and time cost are used as deduction items. The weighted summation is used to obtain the scheme score list. The candidate scheduling schemes are sorted according to the scheme scoring list to generate a scheme ranking list.

[0009] Optionally, the generation of the sub-path contribution table specifically includes: Extract the resource allocation and execution order corresponding to each candidate scheduling scheme from the scheme ranking list to form a set of comparison schemes; Based on the cross-domain connections in the heterogeneous resource graph, each candidate scheduling scheme in the comparison scheme set is integrated into a path sequence consisting of alternating resource nodes and task nodes according to the execution order of task instances and the corresponding resource assignment, thus obtaining a scheme sub-path sequence set; The sub-path sequence set of the scheme is compared pairwise according to the adjacent order in the scheme ranking list. The path segments that differ in resource assignment or execution order are located. The difference sub-paths composed of continuous cross-domain connections and the difference resource combinations composed of the resource node sets involved in the difference sub-paths are extracted to generate a set of difference segments. For each differential sub-path in the differential segment set, calculate the difference between its capability satisfaction, energy consumption estimate and time cost value with the adjacent sequential schemes, and perform weighted synthesis according to preset weights to obtain the differential sub-path attribution score. Then, perform weighted summation according to the occurrence frequency of candidate scheduling schemes and the execution order of task instances to obtain the sub-path attribution score table. Based on the sub-path attribution score table, a threshold is set, and sub-paths and resource node sets with attribution scores not lower than the preset threshold are selected as key sub-paths and key resource combinations, which are then merged to form a sub-path contribution table.

[0010] Optionally, the generation of the pheromone ledger state set specifically includes: Receive the sub-path contribution table, determine the target sub-path and its corresponding key resource combination according to the sub-path attribution score from high to low, and generate a list to be updated. Based on the bidirectional staking sorting ant colony optimization algorithm, new entries are created in the pheromone ledger according to the list to be updated. The pheromone ledger is a structured record set that records the pheromone bonus and pheromone staking of sub-paths and their involved resource nodes. The new entries are initialized with zero bonus and zero staking to obtain the initial view of the pheromone ledger. Based on the sub-path attribution score table, perform allocation updates on the initial view of the pheromone ledger. Write the target sub-paths and their key resource combinations in the list to be updated into the pheromone bonus increment, and write the pheromone quality pledge increment into the sub-paths with attribution scores below the preset threshold, thus forming the allocation update record for this round. By combining the order and frequency of occurrence of target sub-paths in the scheme ranking list, the allocation update records of this round are processed by order weighting and frequency weighting to ensure that the increment of pheromone bonus and pheromone quality bet is consistent with the scheme ranking list, and the weighted update result is obtained. Reversible labeling and status labeling are performed on the weighted update results. Entries written to pheromone bonuses are labeled as solidifiable candidates, and entries written to pheromone quality scores are labeled as candidates to be verified, generating a list of reversible labels. The reversible marker list is merged into the pheromone ledger to generate the pheromone ledger state set.

[0011] Optionally, the generation of the scheduling execution instructions and execution feedback data specifically includes: Read the pheromone ledger state set, the scheme ranking list, the candidate scheduling scheme set, and the heterogeneous resource map to form the phase evaluation input set; The phase judgment index is calculated based on the phase evaluation input set. The phase judgment index is a weighted composite of the ranking change rate of the scheme ranking list in the sliding window, the difference in resource allocation and path segment between schemes in the candidate scheduling scheme set, and the proportion of pheromone quality stake in the pheromone ledger status set according to a preset weight ratio. Phase states are generated based on phase determination indicators. If the phase determination indicators are higher than the preset threshold, they are set to the exploration phase; otherwise, they are set to the consolidation phase. The pheromone staking ratio and pheromone bonus release rate are dynamically adjusted according to the phase states to obtain the pheromone ledger state set after phase adjustment. Based on the adjacency relationship between the phase-adjusted pheromone ledger state set and the resource state graph in the heterogeneous resource graph, the cumulative pheromone bonus of each resource node's incoming and outgoing edges and the time window occupancy of the corresponding candidate scheduling scheme are calculated. Resource nodes that exceed the preset congestion threshold are marked as bottleneck nodes, and a bottleneck node list is generated. The overflow tax is returned to the bottleneck node list. The tax pool is formed by deducting from the incremental pheromone bonus at the inbound and outbound edges of the bottleneck node according to the preset tax rate. According to the resource and task pairs in the set of substitutable and feasible relationships in the resource state diagram, the tax pool is returned to the non-saturated edge of the same node according to the competition edge ratio. The liquidity adjustment record and the pheromone ledger state set after liquidity adjustment are output. Based on the liquidity-adjusted pheromone ledger state set and the scheme ranking list, candidate scheduling schemes that meet the constraints of the resource state diagram and task requirement diagram are selected according to the ranking, and scheduling execution instructions are generated, including resource assignment, task start and end time periods, execution order and synchronization constraints. During the execution of scheduling instructions, the task completion status, resource usage, energy consumption indicators, and deviations relative to the task start and end times are collected to form execution feedback data.

[0012] Optionally, the generation of the optimal scheduling scheme specifically includes: Read the pheromone ledger state set and execution feedback data, and associate them with the scheme ranking list and sub-path contribution table to form a set of verification inputs; Based on the verification input set, the actual performance of each sub-path in the corresponding candidate scheduling scheme is statistically analyzed, and the actual capacity satisfaction, actual energy consumption value and actual time deviation value are summarized to form a list of sub-path actual performance. The actual performance list of sub-paths is compared with the capability satisfaction, energy consumption estimate and time cost item by item. The verification result of each sub-path is determined according to the preset consistency criterion, and the sub-path verification result list is output. Based on the sub-path verification result list, a recycling and solidification list is generated. For the pheromone ledger entries corresponding to the verified sub-paths, the bonus share solidification increment is recorded. For the pheromone ledger entries corresponding to the unverified sub-paths, the pledge share recycling increment is recorded, thus forming the recycling and solidification list. The recycling and solidification list is applied to the pheromone ledger state set to update the cumulative bonus and pledge values ​​of each sub-path and its key resource combination, resulting in the verified pheromone ledger state set. By combining the verified pheromone ledger state set with the bottleneck node list, a freeze mark is applied to entries that involve bottleneck nodes and are marked as failing sub-paths for multiple consecutive rounds. The freeze mark is written into the pheromone ledger to limit its participation time in the subsequent candidate scheduling scheme set and generate a freeze mark record. Based on the verified pheromone ledger state set and the scheme ranking list, the candidate scheduling scheme that satisfies the resource state diagram and task requirement diagram constraints and has the best ranking in the current round is selected as the optimal scheduling scheme, including resource assignment, task start and end time periods, execution order and synchronization constraints.

[0013] The beneficial effects of this invention are: This invention achieves a unified expression of resource status, task requirements, and cross-domain relationships by introducing multi-agent mapping and joint modeling of heterogeneous resource graphs, maintaining spatiotemporal consistency and scheduling interpretability in complex heterogeneous environments. By constructing a candidate scheduling scheme set based on local observations and global constraints, and generating a sub-path contribution table through a ranking and comparison attribution mechanism, the key sources of performance differences between different schemes are effectively distinguished, transforming the optimization process from overall scoring to path-level attribution, significantly improving the accuracy of resource allocation and the transparency of scheduling decisions.

[0014] Furthermore, the bidirectional staking-based ant colony optimization algorithm proposed in this invention combines the dual-channel design of pheromone ledger rewards and staking, enabling the pheromone update process to possess reversibility and self-correction capabilities. This allows for the dynamic recycling of inefficient exploration paths and the consolidation of stable optimal solutions. Through phase adaptive adjustment and an overflow tax return mechanism, balanced control is achieved between the exploration and consolidation phases, suppressing bottleneck node congestion and maintaining resource load balancing. Finally, by combining execution feedback, pheromone verification, freezing inefficient sub-paths, and outputting a contribution graph are completed. This not only achieves a fully closed-loop optimization from data input to the optimal scheduling scheme but also significantly improves the system's stability, interpretability, and global convergence performance. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0016] Figure 1This is a flowchart of a heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization proposed in this invention; Figure 2 This is a flowchart illustrating the generation process of the pheromone ledger state set for a heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-2 A heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization includes the following steps: Collect heterogeneous resource status data and task requirement data, perform format unification and correlation modeling, construct resource status graph and task requirement graph, and obtain heterogeneous resource map and candidate task set; Each resource unit in the heterogeneous resource map is mapped to a multi-agent system, and a set of candidate scheduling schemes is generated based on local observations and global constraints to form a scheme ranking list. Perform sorting, comparison, and attribution on the solution ranking list, identify key sub-paths and key resource combinations, and generate a sub-path contribution table; Based on the bidirectional staking sorting ant colony optimization algorithm, the pheromone ledger is reversibly updated according to the sub-path contribution table. The pheromone of the effective path is recorded as the bonus share, and the path to be verified is recorded as the staking share. During the iteration, the pheromone is dynamically recycled or solidified according to the verification results, and a pheromone ledger state set is generated. Based on the pheromone ledger state set, the search process is phase-adaptive adjusted, switching between the exploration phase and the consolidation phase. Overflow tax is applied to bottleneck nodes to balance resource load, and scheduling execution instructions and execution feedback data are generated. The pheromone ledger state set is verified based on the execution feedback data. Staking shares that have not contributed are reclaimed, and bonus shares of stable sub-paths are solidified. The optimal scheduling scheme is then output.

[0019] In this embodiment, the generation of the heterogeneous resource map and candidate task set specifically includes: Heterogeneous resource status data from computing nodes, energy devices, transportation equipment, and production equipment, and task requirement data from work orders, process plans, and service requests are collected. The two types of data are subjected to unified timestamps, unified spatial coordinates, and unified field definitions to obtain a set of data to be standardized. The set of data to be standardized includes resource identifiers, available time periods, capacity indicators, energy consumption indicators, and status tags, as well as task identifiers, start and end time periods, required capacity, service level, and location tags. The time base alignment is performed on the dataset to be normalized, and the range normalization is performed on the numerical fields to obtain intermediate results of time and numerical normalization. Spatial registration is performed on data containing spatial location fields in the dataset to be normalized, and then merged with intermediate results of temporal and numerical normalization to obtain a spatiotemporal normalization result set; Based on the spatiotemporal normalization result set, a resource capability representation is constructed for each resource unit, and a task requirement representation is constructed for each task instance, forming a resource capability representation set and a task requirement representation set. The resource capability representation consists of the available time period, capability index, energy consumption index, status label and spatial location of the resource, and the task requirement representation consists of the start and end time period of the task, required capability, service level and location label. Based on the resource capability representation set and the task requirement representation set, a feasible matching determination is performed, and the resource and task pairs that meet the conditions are formed into a feasible relationship set. The feasible matching criteria include: the available capability level corresponding to the resource's capability index and status label is not lower than the required capability of the task; the available time period of the resource and the start and end time period of the task have a non-empty intersection; and the spatial distance between the spatial location of the resource and the location corresponding to the location label of the task does not exceed a preset reachable threshold. Based on the set of feasible relationships, a resource state graph is constructed, with resource units as nodes and substitutability, proximity and mutual exclusion relationships between resources as edges. The edge weights are determined by a weighted combination of state similarity, energy consumption difference and spatial distance, and the resource state graph is output. Based on the set of feasible relationships, a task requirement graph is constructed, with task instances as nodes and task dependencies, same-location merging, and time window conflicts as edges. The edge weights are determined by a weighted combination of the strength of temporal dependencies, the strength of same-location merging, and the strength of time window conflicts, and the task requirement graph is output. The resource status graph, task requirement graph, and feasible relationship set are jointly modeled. Resource nodes and task nodes are uniformly encoded and cross-domain connections are established. Cross-domain association weights are determined based on the resource's capability indicators, energy consumption indicators, and status labels, as well as the task's required capabilities, service level, and start and end time periods. This forms a heterogeneous resource graph that represents the resource topology, task constraints, and resource-task mapping relationship. Task nodes with a correlation degree higher than a preset threshold are extracted from the graph to generate a candidate task set.

[0020] In this embodiment, the generation of the scheme ranking list specifically includes: A multi-agent mapping list is established based on the resource nodes in the heterogeneous resource graph. Each resource node corresponds to an agent entry, and each agent entry is associated with the resource capability representation in the resource capability representation set and the adjacency relationship in the resource state graph, thus forming the agent mapping list. For each task instance in the candidate task set, retrieve the feasible relationship set to obtain the matching resource and task pairs, and combine the task requirement representation in the task requirement representation set with the resource capability representation in the resource capability representation set to generate a task ready list. Based on the agent mapping list and the task ready list, according to the edge weights of the resource state graph and the task requirement graph, the local observation elements and system-level constraint elements of each agent entry are summarized to obtain the local observation set and the global constraint set. The local observation elements include available time periods, capacity indicators, energy consumption indicators, and status labels in the resource capacity representation; the system-level constraint elements include adjacency relationships in the resource status graph and temporal dependencies, co-location merging, and time window conflict information in the task requirement graph. Under the constraints of the local observation set and the global constraint set, resource and task assignment combinations are generated around the candidate task set to form a candidate scheduling scheme set with task instances as units and resource assignment and execution order as content, and the resource consumption, energy consumption estimate and time window occupation corresponding to each candidate scheduling scheme are recorded. Calculate the capability satisfaction, energy consumption estimate, and time cost of each candidate scheduling scheme in the candidate scheduling scheme set. According to the preset weights, the capability satisfaction is used as a bonus item, and the energy consumption estimate and time cost are used as deduction items. The weighted summation is used to obtain the scheme score list. The generation of the capability satisfaction, energy consumption estimate, and time cost value specifically includes: Based on the capability indicators and status labels in the resource capability representation, compare them item by item with the required capabilities in the task requirement representation, and calculate the satisfaction ratio as the capability satisfaction degree. Calculate the energy consumption estimate by summing the energy consumption indicators in the resource capability representation and the task execution time. Calculate the time cost based on the overlap ratio between the available resource time period and the task start and end time period. The candidate scheduling schemes are sorted according to the scheme scoring list to generate a scheme ranking list.

[0021] In this embodiment, the generation of the sub-path contribution table specifically includes: Extract the resource allocation and execution order corresponding to each candidate scheduling scheme from the scheme ranking list to form a set of comparison schemes; Based on the cross-domain connections in the heterogeneous resource graph, each candidate scheduling scheme in the comparison scheme set is integrated into a path sequence consisting of alternating resource nodes and task nodes according to the execution order of task instances and the corresponding resource assignment, thus obtaining a scheme sub-path sequence set; The sub-path sequence set of the scheme is compared pairwise according to the adjacent order in the scheme ranking list. The path segments that differ in resource assignment or execution order are located. The difference sub-paths composed of continuous cross-domain connections and the difference resource combinations composed of the resource node sets involved in the difference sub-paths are extracted to generate a set of difference segments. For each differential sub-path in the differential segment set, calculate the difference between its capability satisfaction, energy consumption estimate and time cost value with the adjacent sequential schemes, and perform weighted synthesis according to preset weights to obtain the differential sub-path attribution score. Then, perform weighted summation according to the occurrence frequency of candidate scheduling schemes and the execution order of task instances to obtain the sub-path attribution score table. Based on the sub-path attribution score table, a threshold is set, and sub-paths and resource node sets with attribution scores not lower than the preset threshold are selected as key sub-paths and key resource combinations, which are then merged to form a sub-path contribution table.

[0022] In this embodiment, the generation of the pheromone ledger state set specifically includes: Receive the sub-path contribution table, determine the target sub-path and its corresponding key resource combination according to the sub-path attribution score from high to low, and generate a list to be updated. Based on the bidirectional staking sorting ant colony optimization algorithm, new entries are created in the pheromone ledger according to the list to be updated. The pheromone ledger is a structured record set that records the pheromone bonus and pheromone staking of sub-paths and their involved resource nodes. The new entries are initialized with zero bonus and zero staking to obtain the initial view of the pheromone ledger. Based on the sub-path attribution score table, perform allocation updates on the initial view of the pheromone ledger. Write the target sub-paths and their key resource combinations in the list to be updated into the pheromone bonus increment, and write the pheromone quality pledge increment into the sub-paths with attribution scores below the preset threshold, thus forming the allocation update record for this round. By combining the order and frequency of occurrence of target sub-paths in the scheme ranking list, the allocation update records of this round are processed by order weighting and frequency weighting to ensure that the increment of pheromone bonus and pheromone quality bet is consistent with the scheme ranking list, and the weighted update result is obtained. Reversible labeling and status labeling are performed on the weighted update results. Entries written to pheromone bonuses are labeled as solidifiable candidates, and entries written to pheromone quality scores are labeled as candidates to be verified, generating a list of reversible labels. The reversible marker list is merged into the pheromone ledger to generate the pheromone ledger state set.

[0023] In this embodiment, the generation of the scheduling execution instruction and execution feedback data specifically includes: Read the pheromone ledger state set, the scheme ranking list, the candidate scheduling scheme set, and the heterogeneous resource map to form the phase evaluation input set; The phase judgment index is calculated based on the phase evaluation input set. The phase judgment index is a weighted composite of the ranking change rate of the scheme ranking list in the sliding window, the difference in resource allocation and path segment between schemes in the candidate scheduling scheme set, and the proportion of pheromone quality stake in the pheromone ledger status set according to a preset weight ratio. Phase states are generated based on phase determination indicators. If the phase determination indicators are higher than the preset threshold, they are set to the exploration phase; otherwise, they are set to the consolidation phase. The pheromone staking ratio and pheromone bonus release rate are dynamically adjusted according to the phase states to obtain the pheromone ledger state set after phase adjustment. Based on the adjacency relationship between the phase-adjusted pheromone ledger state set and the resource state graph in the heterogeneous resource graph, the cumulative pheromone bonus of each resource node's incoming and outgoing edges and the time window occupancy of the corresponding candidate scheduling scheme are calculated. Resource nodes that exceed the preset congestion threshold are marked as bottleneck nodes, and a bottleneck node list is generated. The overflow tax is returned to the bottleneck node list. The tax pool is formed by deducting from the incremental pheromone bonus at the inbound and outbound edges of the bottleneck node according to the preset tax rate. According to the resource and task pairs in the set of substitutable and feasible relationships in the resource state diagram, the tax pool is returned to the non-saturated edge of the same node according to the competition edge ratio. The liquidity adjustment record and the pheromone ledger state set after liquidity adjustment are output. Based on the liquidity-adjusted pheromone ledger state set and the scheme ranking list, candidate scheduling schemes that meet the constraints of the resource state diagram and task requirement diagram are selected according to the ranking, and scheduling execution instructions are generated, including resource assignment, task start and end time periods, execution order and synchronization constraints. During the execution of scheduling instructions, the task completion status, resource usage, energy consumption indicators, and deviations relative to the task start and end times are collected to form execution feedback data.

[0024] In this embodiment, the generation of the optimal scheduling scheme specifically includes: Read the pheromone ledger state set and execution feedback data, and associate them with the scheme ranking list and sub-path contribution table to form a set of verification inputs; Based on the verification input set, the actual performance of each sub-path in the corresponding candidate scheduling scheme is statistically analyzed, and the actual capacity satisfaction, actual energy consumption value and actual time deviation value are summarized to form a list of sub-path actual performance. Among them, the actual capability satisfaction is obtained by the degree of conformity between the task completion status and resource usage and the resource capability representation and the task requirement representation; the actual energy consumption value is obtained by accumulating the energy consumption indicators in the execution feedback data according to the execution duration; and the actual time deviation value is obtained by accumulating the deviation in the execution feedback data relative to the start and end time of the task. The actual performance list of sub-paths is compared with the capability satisfaction, energy consumption estimate and time cost item by item. The verification result of each sub-path is determined according to the preset consistency criterion, and the sub-path verification result list is output. Based on the sub-path verification result list, a recycling and solidification list is generated. For the pheromone ledger entries corresponding to the verified sub-paths, the bonus share solidification increment is recorded. For the pheromone ledger entries corresponding to the unverified sub-paths, the pledge share recycling increment is recorded, thus forming the recycling and solidification list. The recycling and solidification list is applied to the pheromone ledger state set to update the cumulative bonus and pledge values ​​of each sub-path and its key resource combination, resulting in the verified pheromone ledger state set. By combining the verified pheromone ledger state set with the bottleneck node list, a freeze mark is applied to entries that involve bottleneck nodes and are marked as failing sub-paths for multiple consecutive rounds. The freeze mark is written into the pheromone ledger to limit its participation time in the subsequent candidate scheduling scheme set and generate a freeze mark record. Based on the verified pheromone ledger state set and the scheme ranking list, the candidate scheduling scheme that satisfies the resource state diagram and task requirement diagram constraints and has the best ranking in the current round is selected as the optimal scheduling scheme, including resource assignment, task start and end time periods, execution order and synchronization constraints.

[0025] Example 1: To verify the feasibility of this invention in practice, it was applied to a cross-domain collaborative scheduling scenario in a coastal manufacturing park. This park contains heterogeneous resources of various types, including computing nodes, energy devices, transportation equipment, and production equipment. In daily operation, it faces problems such as large differences in energy consumption, inconsistent equipment status updates, and uneven distribution of task demands, often leading to resource allocation conflicts, operation delays, and energy waste. Especially during peak periods, overlapping time windows between different tasks and excessive equipment load often make it difficult for manual scheduling strategies to balance real-time performance and optimization. After deployment in this environment, this invention aims to solve the problems of information fragmentation between heterogeneous resources, delayed scheduling decisions, and low resource utilization.

[0026] In the application process, multi-source operational data from within the park is first collected, including task execution logs of computing nodes, power status of energy devices, path and time distribution of transportation equipment, and operating load and status labels of production equipment. By unifying time and spatial benchmarks, a standardized set of heterogeneous resource status data and task requirement data is generated. Subsequently, using the heterogeneous resource map construction method of this invention, resource capabilities, energy consumption characteristics, and task constraints are uniformly encoded to form a spatiotemporal network structure that reflects resource topology and task dependencies. Based on this, the system automatically identifies feasible matching relationships between resources and tasks, laying the data foundation for subsequent multi-agent mapping.

[0027] After the heterogeneous resource map is generated, this invention maps the resource units within it to multiple agents, generating multiple sets of candidate scheduling schemes through local observation and global constraints. These candidate schemes are weighted, scored, and ranked to form a ranking list, used to identify the execution potential of each scheme. Subsequently, based on a ranking comparison and attribution mechanism, path-level analysis is performed on the key differences between schemes, extracting the performance contribution of sub-paths and key resource combinations, thereby generating a sub-path contribution table to provide a basis for subsequent optimization directions.

[0028] During the optimization phase, this invention introduces a bidirectional staking-based ant colony optimization algorithm. Through a pheromone ledger mechanism, pheromone from effective paths is recorded as bonus shares, and paths to be verified are recorded as stake shares. These are dynamically recycled and solidified over multiple iterations, making the optimization process reversible and self-correcting. Combined with a phase adaptive adjustment mechanism, when the ranking change rate of a solution is high, it automatically switches to an exploration phase to expand the search space; when a solution tends to stabilize, it enters a consolidation phase to strengthen the optimal solution. To address resource congestion, this invention also designs an overflow tax return strategy, which reclaims a portion of the pheromone bonuses from nodes with excessive resource load and redistributes them to unsaturated resources, thereby achieving load balancing and improved global liquidity.

[0029] In actual scheduling and execution, scheduling instructions are issued to each resource control unit through an interface with the park control center. The agents coordinate their work according to the assigned task start and end times and execution order. During execution, task completion status, energy consumption changes, and resource usage are collected in real time, forming execution feedback data. This feedback data is used to verify the distribution of bonuses and pledges in the pheromone ledger; paths that do not contribute effectively are reclaimed, while stable paths are solidified and prioritized. After multiple rounds of feedback and correction, the final optimal scheduling scheme shows significant improvements in task allocation rationality, energy consumption control, and execution consistency.

[0030] Through actual operation in this manufacturing park, this invention has achieved unified scheduling and dynamic optimization of multi-source heterogeneous resources, significantly reducing inter-task conflicts and resource idleness, resulting in more balanced energy consumption and more stable task completion timing. The results show that the multi-agent collaborative optimization and bidirectional pledge mechanism of this invention not only effectively improves resource utilization and scheduling efficiency, but also achieves a complete closed loop from data input, path attribution, pheromone updates to execution feedback, providing an efficient, reliable, and interpretable optimization method for the fusion scheduling of heterogeneous resources in complex industrial scenarios.

[0031] Table 1. Performance comparison between the heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization and traditional methods.

[0032] From the three core results of on-time completion rate, average energy consumption index, and average waiting time, this invention achieves significant improvements and reductions compared to Rank-based ACO, respectively, manifested in higher task fulfillment, lower energy consumption, and shorter queuing times. The direct reason for the improvement is that after the solution is generated, it does not evenly reinforce the entire path, but instead assigns bonuses only to key sub-paths and key resource combinations that cause ranking differences through ranking comparison and attribution, reducing the accumulation of ineffective substructures from the source; at the same time, after the execution feedback enters the pheromone ledger, the pledged shares can be recovered and the bonus shares can be solidified, so that the structure of "high capability satisfaction and low time cost" is continuously consolidated, avoiding premature lock-in caused by traditional one-way accumulation.

[0033] Regarding network congestion and load balancing, the frequency of bottleneck node congestion and load balancing have been significantly improved. The fundamental reason is the combined effect of phase adaptation and overflow tax recirculation: when scheme diversity decreases or dividends are excessively concentrated locally, phase adaptation is used to increase the staking write ratio and reduce the dividend release rate to prevent unilateral solidification; once a bottleneck node is detected, overflow tax recirculation directs dividends from saturated edges to unsaturated edges according to the competing edges, promoting liquidity redistribution at the resource level, thereby shortening queuing, reducing waiting time, and smoothing node utilization. Unlike the approach of simply adjusting parameters based on evaporation rate, this "tax-recirculation" forms a controllable load guiding channel on the graph.

[0034] Regarding convergence efficiency and stability, both the number of iterative convergence rounds and the number of rounds required for the optimal solution to stabilize are significantly reduced, indicating that this invention reaches and maintains the optimal solution faster under the same time budget. The mechanism lies in the bidirectional reversible design of the pheromone ledger: for sub-paths below a threshold and repeatedly failing verification, the staked portion is dynamically recovered and a freeze flag is triggered, significantly reducing the proportion of subsequent invalid explorations; for stable sub-paths, the gradual release and solidification of bonuses are accelerated, forming a steady-state convergence where "the better becomes better." This mechanism reduces unnecessary iterations on poor paths and avoids the noise caused by the traditional elitist strategy's one-size-fits-all reward application across the entire path.

[0035] The significant increase in the attributable percentage of critical sub-paths stems from the integrated output of the sub-path contribution table and contribution graph. Since each ranking change corresponds to a clearly defined differentiated sub-path, and execution feedback can link these sub-paths to the final optimal solution, managers can trace the reason for the given scheduling recommendation along the "resource-sub-path-task" link. This interpretability transforms the trade-offs between energy consumption optimization, time window coordination, and resource allocation from a "black box" to "auditable," facilitating rapid verification and alignment of operational goals in actual park-level operations. In summary, the simultaneous improvement of various indicators in the table originates from the organic coupling of multi-agent collaboration, ranking comparison attribution, bidirectional pledge accounting, and phase-backflow adjustment in this invention. This not only solves the problems of premature convergence, congestion, and unexplainability inherent in traditional ant colony optimization but also creates a closed-loop advantage in engineering availability and deployment controllability.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization, characterized in that, Includes the following steps: Collect heterogeneous resource status data and task requirement data, perform format unification and correlation modeling, construct resource status graph and task requirement graph, and obtain heterogeneous resource map and candidate task set; Each resource unit in the heterogeneous resource map is mapped to a multi-agent system, and a set of candidate scheduling schemes is generated based on local observations and global constraints to form a scheme ranking list. Perform sorting, comparison, and attribution on the solution ranking list, identify key sub-paths and key resource combinations, and generate a sub-path contribution table; Based on the bidirectional staking sorting ant colony optimization algorithm, the pheromone ledger is reversibly updated according to the sub-path contribution table. The pheromone of the effective path is recorded as the bonus share, and the path to be verified is recorded as the staking share. During the iteration, the pheromone is dynamically recycled or solidified according to the verification results, and a pheromone ledger state set is generated. Based on the pheromone ledger state set, the search process is phase-adaptive adjusted, switching between the exploration phase and the consolidation phase. Overflow tax is applied to bottleneck nodes to balance resource load, and scheduling execution instructions and execution feedback data are generated. The pheromone ledger state set is verified based on the execution feedback data. Staking shares that have not contributed are reclaimed, and bonus shares of stable sub-paths are solidified. The optimal scheduling scheme is then output.

2. The heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization according to claim 1, characterized in that, The generation of the heterogeneous resource map and candidate task set specifically includes: Collect heterogeneous resource status data from computing nodes, energy devices, transportation equipment and production equipment, and task requirement data from work orders, process plans and service requests. Perform timestamp unification, spatial coordinate unification and field definition unification operations on the two types of data to obtain a set of data to be standardized. The time base alignment is performed on the dataset to be normalized, and the range normalization is performed on the numerical fields to obtain intermediate results of time and numerical normalization. Spatial registration is performed on data containing spatial location fields in the dataset to be normalized, and then merged with intermediate results of temporal and numerical normalization to obtain a spatiotemporal normalization result set; Based on the spatiotemporal normalization result set, a resource capability representation is constructed for each resource unit, and a task requirement representation is constructed for each task instance, forming a resource capability representation set and a task requirement representation set; Based on the resource capability representation set and the task requirement representation set, a feasible matching determination is performed, and the resource and task pairs that meet the conditions are formed into a feasible relationship set. Based on the set of feasible relationships, a resource state graph is constructed, with resource units as nodes and substitutability, proximity and mutual exclusion relationships between resources as edges. The edge weights are determined by a weighted combination of state similarity, energy consumption difference and spatial distance, and the resource state graph is output. Based on the set of feasible relationships, a task requirement graph is constructed, with task instances as nodes and task dependencies, same-location merging, and time window conflicts as edges. The edge weights are determined by a weighted combination of the strength of temporal dependencies, the strength of same-location merging, and the strength of time window conflicts, and the task requirement graph is output. By jointly modeling the resource state graph, task requirement graph, and feasible relationship set, resource nodes and task nodes are uniformly encoded and cross-domain connections are established to form a heterogeneous resource graph that represents resource topology, task constraints, and resource-task mapping relationships. From this graph, task nodes with a correlation degree higher than a preset threshold are extracted to generate a candidate task set.

3. The heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization according to claim 1, characterized in that, The generation of the ranking list of schemes specifically includes: A multi-agent mapping list is established based on the resource nodes in the heterogeneous resource graph. Each resource node corresponds to an agent entry, and each agent entry is associated with the resource capability representation in the resource capability representation set and the adjacency relationship in the resource state graph, thus forming the agent mapping list. For each task instance in the candidate task set, retrieve the feasible relationship set to obtain the matching resource and task pairs, and combine the task requirement representation in the task requirement representation set with the resource capability representation in the resource capability representation set to generate a task ready list. Based on the agent mapping list and the task ready list, according to the edge weights of the resource state graph and the task requirement graph, the local observation elements and system-level constraint elements of each agent entry are summarized to obtain the local observation set and the global constraint set. Under the constraints of the local observation set and the global constraint set, resource and task assignment combinations are generated around the candidate task set to form a candidate scheduling scheme set with task instances as units and resource assignment and execution order as content, and the resource consumption, energy consumption estimate and time window occupation corresponding to each candidate scheduling scheme are recorded. Calculate the capability satisfaction, energy consumption estimate, and time cost of each candidate scheduling scheme in the candidate scheduling scheme set. According to the preset weights, the capability satisfaction is used as a bonus item, and the energy consumption estimate and time cost are used as deduction items. The weighted summation is used to obtain the scheme score list. The candidate scheduling schemes are sorted according to the scheme scoring list to generate a scheme ranking list.

4. The heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization according to claim 1, characterized in that, The generation of the sub-path contribution table specifically includes: Extract the resource allocation and execution order corresponding to each candidate scheduling scheme from the scheme ranking list to form a set of comparison schemes; Based on the cross-domain connections in the heterogeneous resource graph, each candidate scheduling scheme in the comparison scheme set is integrated into a path sequence consisting of alternating resource nodes and task nodes according to the execution order of task instances and the corresponding resource assignment, thus obtaining a scheme sub-path sequence set; The sub-path sequence set of the scheme is compared pairwise according to the adjacent order in the scheme ranking list. The path segments that differ in resource assignment or execution order are located. The difference sub-paths composed of continuous cross-domain connections and the difference resource combinations composed of the resource node sets involved in the difference sub-paths are extracted to generate a set of difference segments. For each differential sub-path in the differential segment set, calculate the difference between its capability satisfaction, energy consumption estimate and time cost value with the adjacent sequential schemes, and perform weighted synthesis according to preset weights to obtain the differential sub-path attribution score. Then, perform weighted summation according to the occurrence frequency of candidate scheduling schemes and the execution order of task instances to obtain the sub-path attribution score table. Based on the sub-path attribution score table, a threshold is set, and sub-paths and resource node sets with attribution scores not lower than the preset threshold are selected as key sub-paths and key resource combinations, which are then merged to form a sub-path contribution table.

5. The heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization according to claim 1, characterized in that, The generation of the pheromone ledger state set specifically includes: Receive the sub-path contribution table, determine the target sub-path and its corresponding key resource combination according to the sub-path attribution score from high to low, and generate a list to be updated. Based on the bidirectional staking sorting ant colony optimization algorithm, new entries are created in the pheromone ledger according to the list to be updated. The pheromone ledger is a structured record set that records the pheromone bonus and pheromone staking of sub-paths and their involved resource nodes. The new entries are initialized with zero bonus and zero staking to obtain the initial view of the pheromone ledger. Based on the sub-path attribution score table, perform allocation updates on the initial view of the pheromone ledger. Write the target sub-paths and their key resource combinations in the list to be updated into the pheromone bonus increment, and write the pheromone quality pledge increment into the sub-paths with attribution scores below the preset threshold, thus forming the allocation update record for this round. By combining the order and frequency of occurrence of target sub-paths in the scheme ranking list, the allocation update records of this round are processed by order weighting and frequency weighting to ensure that the increment of pheromone bonus and pheromone quality bet is consistent with the scheme ranking list, and the weighted update result is obtained. Reversible labeling and status labeling are performed on the weighted update results. Entries written to pheromone bonuses are labeled as solidifiable candidates, and entries written to pheromone quality scores are labeled as candidates to be verified, generating a list of reversible labels. The reversible marker list is merged into the pheromone ledger to generate the pheromone ledger state set.

6. The heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization according to claim 1, characterized in that, The generation of the scheduling execution instructions and execution feedback data specifically includes: Read the pheromone ledger state set, the scheme ranking list, the candidate scheduling scheme set, and the heterogeneous resource map to form the phase evaluation input set; The phase judgment index is calculated based on the phase evaluation input set. The phase judgment index is a weighted composite of the ranking change rate of the scheme ranking list in the sliding window, the difference in resource allocation and path segment between schemes in the candidate scheduling scheme set, and the proportion of pheromone quality stake in the pheromone ledger status set according to a preset weight ratio. Phase states are generated based on phase determination indicators. If the phase determination indicators are higher than the preset threshold, they are set to the exploration phase; otherwise, they are set to the consolidation phase. The pheromone staking ratio and pheromone bonus release rate are dynamically adjusted according to the phase states to obtain the pheromone ledger state set after phase adjustment. Based on the adjacency relationship between the phase-adjusted pheromone ledger state set and the resource state graph in the heterogeneous resource graph, the cumulative pheromone bonus of each resource node's incoming and outgoing edges and the time window occupancy of the corresponding candidate scheduling scheme are calculated. Resource nodes that exceed the preset congestion threshold are marked as bottleneck nodes, and a bottleneck node list is generated. The overflow tax is returned to the bottleneck node list. The tax pool is formed by deducting from the incremental pheromone bonus at the inbound and outbound edges of the bottleneck node according to the preset tax rate. According to the resource and task pairs in the set of substitutable and feasible relationships in the resource state diagram, the tax pool is returned to the non-saturated edge of the same node according to the competition edge ratio. The liquidity adjustment record and the pheromone ledger state set after liquidity adjustment are output. Based on the liquidity-adjusted pheromone ledger state set and the scheme ranking list, candidate scheduling schemes that meet the constraints of the resource state diagram and task requirement diagram are selected according to the ranking, and scheduling execution instructions are generated, including resource assignment, task start and end time periods, execution order and synchronization constraints. During the execution of scheduling instructions, the task completion status, resource usage, energy consumption indicators, and deviations relative to the task start and end times are collected to form execution feedback data.

7. The heterogeneous resource fusion scheduling method based on multi-agent cooperative optimization according to claim 1, characterized in that, The generation of the optimal scheduling scheme specifically includes: Read the pheromone ledger state set and execution feedback data, and associate them with the scheme ranking list and sub-path contribution table to form a set of verification inputs; Based on the verification input set, the actual performance of each sub-path in the corresponding candidate scheduling scheme is statistically analyzed, and the actual capacity satisfaction, actual energy consumption value and actual time deviation value are summarized to form a list of sub-path actual performance. The actual performance list of sub-paths is compared with the capability satisfaction, energy consumption estimate and time cost item by item. The verification result of each sub-path is determined according to the preset consistency criterion, and the sub-path verification result list is output. Based on the sub-path verification result list, a recycling and solidification list is generated. For the pheromone ledger entries corresponding to the verified sub-paths, the bonus share solidification increment is recorded. For the pheromone ledger entries corresponding to the unverified sub-paths, the pledge share recycling increment is recorded, thus forming the recycling and solidification list. The recycling and solidification list is applied to the pheromone ledger state set, and the cumulative bonus and pledge values ​​of each sub-path and its key resource combination are updated to obtain the verified pheromone ledger state set. By combining the verified pheromone ledger state set with the bottleneck node list, a freeze mark is applied to entries that involve bottleneck nodes and are marked as failing sub-paths for multiple consecutive rounds. The freeze mark is written into the pheromone ledger to limit its participation time in the subsequent candidate scheduling scheme set and generate a freeze mark record. Based on the verified pheromone ledger state set and the scheme ranking list, the candidate scheduling scheme that satisfies the resource state diagram and task requirement diagram constraints and has the best ranking in the current round is selected as the optimal scheduling scheme, including resource assignment, task start and end time periods, execution order and synchronization constraints.

Images