Scheduling optimization method and system based on multi-criterion variable neighborhood search and storage medium

Abstract

The invention provides a scheduling optimization method and system based on multi-criterion variable neighborhood search and a storage medium, and belongs to the field of scheduling optimization. Comprising the following steps: constructing a target function and a constraint function of the comprehensive cost of the variable-speed motor; generating an initial solution meeting a constraint function through a sorting scoring function based on a greedy random construction method; performing intelligent jitter and local search on the initial solution or the historical optimal solution through an iterative search mode based on a variable neighborhood search algorithm of a self-adaptive neighborhood selection mechanism so as to determine a local optimal solution of the target function in each iterative search; and terminating iteration and outputting a globally optimal solution of the target function under the condition that the iterative search reaches a termination condition. According to the method, the carbon emission cost and the total weighted delay time are comprehensively considered, an improved variable neighborhood search algorithm based on a self-adaptive neighborhood selection mechanism is provided, global exploration and local development are balanced through an intelligent jitter strategy, falling into a single search mode is avoided, and a high-quality solution can be quickly converged.

Description

Technical Field

[0001] This invention relates to the field of scheduling optimization technology, specifically to a scheduling optimization method, system, and storage medium based on multi-criteria variable neighborhood search. Background Technology

[0002] The newly released "Action Plan for Green and Low-Carbon Development of Manufacturing" clearly states that the industrial value chain should be restructured through a green manufacturing system. Green manufacturing minimizes environmental impact through product life cycle management (from raw material acquisition to recycling and disposal), while carbon emission control focuses on reducing greenhouse gas emissions. The two are naturally coupled in terms of sustainable development goals.

[0003] It is evident that carbon emission intensity will be incorporated as a core indicator into the high-quality development of the manufacturing industry, for example, it can be applied to the variable speed machine scheduling problem. The variable speed machine scheduling problem typically requires minimizing the total weighted delay time. It can be briefly described as: assigning multiple tasks to several variable speed machines for processing, with different processing speeds on different machines. The constraints for processing tasks on the machines are: each machine can process tasks independently and continuously, and each machine can only process one task at a time.

[0004] Currently, most existing technologies focus on single-objective optimization. Most workshop production scheduling problems do not consider carbon emission costs in their modeling. Some studies that do consider machine usage costs treat them as objective functions and solve the problem using heuristic algorithms. Furthermore, traditional variable neighborhood search relies on random strategies for initial solution generation, failing to incorporate job priorities, thus impacting solution efficiency and quality. Traditional variable neighborhood search also features a fixed neighborhood structure design and does not dynamically adjust the global and local search ratio, resulting in low efficiency for solving large-scale problems. Summary of the Invention

[0005] The purpose of this invention is to provide a scheduling optimization method, system, and storage medium based on multi-criteria variable neighborhood search. By improving the variable neighborhood search algorithm, the solution efficiency and solution quality are improved, meeting the needs of enterprises for low-carbon and high-efficiency production, achieving a balance between production efficiency and environmental protection, and realizing the goal of green manufacturing.

[0006] To achieve the above objectives, this invention provides a scheduling optimization method based on multi-criteria variable neighborhood search. The optimization method includes: constructing an objective function and its constraint function for the overall cost of the transmission, wherein the overall cost is a weighted sum of the delay time of the task and the carbon emission cost of the machine; generating an initial solution that satisfies the constraint function using a sorting and scoring function based on a greedy random construction method; using a variable neighborhood search algorithm based on an adaptive neighborhood selection mechanism to intelligently jitter and perform local search on the initial solution or historical optimal solution through iterative search to determine the local optimal solution of the objective function in each iteration search, wherein the historical optimal solution is the local optimal solution before the current iteration search; and terminating the iteration and outputting the global optimal solution of the objective function when the iterative search reaches a termination condition.

[0007] On the other hand, the present invention provides a scheduling optimization system based on multi-criteria variable neighborhood search. The optimization system includes: an objective function construction device for constructing an objective function and its constraint function for the overall cost of the transmission, wherein the overall cost is a weighted sum of the delay time of the task and the carbon emission cost of the machine; an initial solution generation device for generating an initial solution that satisfies the constraint function through a sorting and scoring function based on a greedy random construction method; a local optimal solution generation device for intelligently jittering and local searching on the initial solution or historical optimal solution through iterative search based on a variable neighborhood search algorithm with an adaptive neighborhood selection mechanism, to determine the local optimal solution of the objective function in each iterative search, wherein the historical optimal solution is the local optimal solution before the current iterative search; and a global optimal solution generation device for terminating the iteration and outputting the global optimal solution of the objective function when the iterative search reaches the termination condition.

[0008] On the other hand, the present invention provides a machine-readable storage medium storing instructions for causing a machine to execute: the scheduling optimization method based on multi-criteria variable neighborhood search as described above.

[0009] Through the above technical solutions, this invention provides a scheduling optimization method, system and storage medium based on multi-criteria variable neighborhood search. Taking into account carbon emission costs and total weighted delay time, it proposes an improved variable neighborhood search algorithm based on an adaptive neighborhood selection mechanism. By using an intelligent jitter strategy to balance global exploration and local development, it avoids falling into a single search mode and can quickly converge to a high-quality solution.

[0010] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0011] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart illustrating the scheduling optimization method based on multi-criteria variable neighborhood search provided by this invention; Figure 2 This is a schematic diagram illustrating the specific execution flow of the variable speed gear scheduling problem of the present invention; Figure 3 This is a flowchart illustrating the specific operation of the jittering process of the present invention; Figure 4 This is a flowchart of the exploratory phase operation of the present invention; Figure 5 This is a flowchart of the development process of the present invention; Figure 6 This is a flowchart of the standby operation of the present invention; Figure 7 This is a schematic diagram of the scheduling optimization system based on multi-criteria variable neighborhood search provided by the present invention. Detailed Implementation

[0012] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0013] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0014] The applicant first provides the following explanation of the application background and existing technology of this invention. Since the manufacturing industry contributes over 33% of global carbon emissions, this industrial characteristic dictates that green manufacturing must become the core carrier of carbon emission reduction. With the implementation of the "dual control of carbon emissions" system, green manufacturing will shift from single-point technological breakthroughs to system-wide innovation encompassing "design-production-service". It is estimated that green manufacturing technologies can reduce emissions in the industrial sector by 40-60%, becoming a key lever for achieving carbon neutrality.

[0015] Currently, existing technologies mostly focus on single-objective optimization. Most workshop production scheduling problems do not consider carbon emission costs in their modeling. Some studies that do consider machine usage costs treat machine usage costs as the objective function and solve the problem using heuristic algorithms. Furthermore, traditional variable neighborhood search relies on random strategies for initial solution generation, failing to incorporate job priorities, thus affecting solution efficiency and quality. Traditional variable neighborhood search also uses a fixed neighborhood structure design and does not dynamically adjust the global and local search ratio, resulting in low efficiency for solving large-scale problems. Specifically, the applicant has identified the following shortcomings in existing technologies: 1. Most studies focus only on optimizing a single objective, such as minimizing total weighted delay time or minimizing energy consumption, lacking comprehensive consideration of multiple practical factors, and thus failing to meet the actual needs of enterprises for low-carbon and efficient production.

[0016] 2. Although the variable neighborhood search algorithm has some applications in scheduling problems, the traditional variable neighborhood search algorithm is prone to getting trapped in local optima when dealing with large-scale problems, and it has shortcomings in the generation of initial solutions and the selection of neighborhood structures, which leads to the need to improve its solution efficiency and solution quality in complex scheduling problems.

[0017] 3. Some studies have modeled the scheduling problem in a relatively simple way, without fully considering the various complex constraints in actual production, which limits the applicability and effectiveness of the model in practical applications.

[0018] 4. Traditional dithering operations lack targeted exploration of the high-quality solution space, resulting in low search efficiency and easy repetition of ineffective dithering. When the release algorithm gets stuck in a local optimum, it only relies on simple random reset or fixed rule dithering, and cannot generate counter-attempt operations to break the deadlock.

[0019] To address this issue, this invention proposes a variable speed engine scheduling optimization method that comprehensively considers carbon emission costs and total weighted delay time. This method minimizes the sum of total weighted delay time and carbon emission costs. By improving the variable neighborhood search algorithm, it enhances solution efficiency and quality, meeting enterprises' demands for low-carbon and efficient production. The overall approach of this invention is as follows: first, the parameters and decision variables used in the model are given; then, a mixed-integer programming model is constructed for the problem; finally, an improved variable neighborhood search algorithm is designed.

[0020] Specifically, the present invention first provides a scheduling optimization method 100 based on multi-criteria variable neighborhood search, such as... Figure 1 As shown, the optimization method 100 of the present invention may include steps S110-S140.

[0021] Step S110: Construct the objective function and its constraint function for the overall cost of the transmission.

[0022] The comprehensive cost is the weighted sum of the delay time and carbon emission costs of the variable speed machine's operation. Different machines performing the same processing task typically produce significantly different carbon emissions. The core reason for this difference lies in the machine's "energy consumption efficiency," "energy type," and "production process auxiliary links," ultimately manifested through "direct carbon emissions" (generated directly by machine operation) and "indirect carbon emissions" (indirectly related to energy production, auxiliary material consumption, etc.). For example, taking the processing of aluminum alloy parts (0.5kg, including milling / drilling processes) as an example, machines of different eras and types can have significantly different carbon emissions of 3-10 times when performing the same processing task. This difference is mainly determined by three key factors: energy consumption efficiency (e.g., intelligent machine tools are 50% more energy-efficient than traditional machine tools), energy type (the carbon intensity of different energy sources such as diesel, thermal power, green electricity, or hydrogen can differ by up to 10 times), and production process auxiliary links (such as coolant use, waste disposal, etc., which are indirect emissions). These factors collectively influence the final carbon emission performance, with the proportions of direct carbon emissions (such as diesel combustion) and indirect carbon emissions (such as the electricity production chain) changing with technological iterations—for example, in the 1990s, diesel machine tools accounted for over 80% of direct carbon emissions, while in 2020, hydrogen-powered machine tools accounted for a significantly reduced proportion of indirect carbon emissions.

[0023] It is worth noting that, unlike the concept of parallel machines in other scenarios, where each machine in a parallel machine has the same processing time for all jobs, the variable speed machine studied in this invention has different processing times for different jobs.

[0024] In one embodiment, some parameters and decision variables of this embodiment can be represented as follows: Indicates the first One machine, ; Indicates the first This assignment, ; Indicate homework The weights; Indicates the weight of the delay time; Indicates the carbon emission cost weight, ; Indicate homework Completion time; Indicate homework The delay time; Indicate homework In the machine Processing time; Indicates machine Fixed carbon emission costs, including start-up carbon emission costs, refer to the economic cost of carbon emissions generated by equipment / systems during the transition from shutdown to steady-state operation. Indicates machine The unit carbon emission cost during operation.

[0025] in, .

[0026] The objective function of this invention represents the overall objective function of the model, which is to minimize the total weighted comprehensive cost. The mathematical model can be expressed as follows: (1) Where S0 is the initial solution; S is the current solution; and T(S) is the total delay time of all tasks under the current solution. .

[0027] C(S) represents the total machine usage cost, including machine carbon emission costs, for all currently solved tasks. Equation (1) serves as the objective function, with the scheduling goal of minimizing the weighted sum of task delay time and machine carbon emission costs. This is achieved by adjusting the weights... and This approach allows for a flexible balance between the relative importance of total weighted delay time and machine carbon emission costs. It can be understood as a dynamic balance between cost and time based on weighted allocation, with the objective function minimizing the weighted combination of cost and time, thus achieving a dynamic balance between production efficiency and environmental costs. By adjusting the weights, it flexibly adapts to enterprise strategies prioritizing efficiency, environmental protection, or balanced production.

[0028] It is important to note that since the delay time and cost in the objective function have different dimensions, directly adding them together may lead to inaccurate optimization. Therefore, it is advisable to use the initial solution for normalization. First, generate the initial solution, and then the objective function is the sum of the weighted service time of the current solution divided by the weighted service time of the initial solution and the cost of the current solution divided by the cost of the initial solution.

[0029] The constraint function can be specifically expressed as follows: (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Equation (2) ensures that each job is processed only once on all machines; Equation (3) is used to initialize the machines, ensuring that each machine processes at most one job initially; Equation (4) represents the flow conservation formula, ensuring that the number of predecessor and successor jobs is balanced on each machine, guaranteeing a reasonable processing order of jobs on the machines; Equations (5) and (6) determine the job completion time constraints, where Equation (7) is a sufficiently large number to ensure that the logical relationship holds true; Equation (7) calculates the usage cost of all machines, which consists of two parts: the fixed cost of machine usage and the variable cost of machines including carbon emission costs; Equation (8) ensures that the completion time of the operation is non-negative; Equation (9) ensures that the delay time of the operation is non-negative; Equation (10) indicates that the decision variable is a 0-1 decision variable.

[0030] The hybrid integer programming model constructed in this invention features model innovation and flexibility, achieving a breakthrough combination of carbon emission quantification and dynamic optimization in the field of variable speed engine production scheduling. The constructed objective function comprehensively considers total weighted delay time and carbon emission cost, making the weighted combined objective function, which includes both total weighted delay time and carbon emission cost, closer to the low-carbon and efficient scheduling requirements in actual production, thus achieving a dynamic balance based on dual criteria. Furthermore, this invention can directly convert process energy consumption data into carbon emissions, enabling traditional scheduling algorithms to visualize their carbon footprint for the first time. For example, the additional energy consumption caused by delays in heat treatment processes can be displayed in real time as a 15% increase in carbon emissions. This carbon cost explicitness mechanism, combined with the dynamic adjustment function of the subsequent variable neighborhood search algorithm, can identify changes in carbon emission intensity or generate low-carbon solutions within seconds when urgent orders arrive during production scheduling.

[0031] In addition, this invention also introduces a weighting coefficient. and This model flexibly balances the relative importance of total weighted delay time and carbon emission costs, breaking through the limitations of traditional single-objective optimization. It achieves a quantitative balance between production efficiency and environmental costs, and the weights can be flexibly adjusted according to enterprise needs, providing enterprises with a more comprehensive basis for production decisions. Thus, this invention creatively incorporates production efficiency and carbon emissions into a unified scheduling framework, achieving a precise balance between the two through a dynamic weight adjustment mechanism. The model overcomes the limitations of traditional single-objective optimization, empowering enterprises to flexibly adjust the priorities of environmental protection and efficiency according to actual needs. This technology provides a quantifiable implementation path for green transformation in the industrial sector and has significant industry promotion value under dual-carbon objectives.

[0032] Step S120: Generate an initial solution that satisfies the constraint function by using a sorting scoring function based on a greedy random construction method.

[0033] To ensure the quality of the initial solution, this invention designs a greedy random construction method based on the Greedy Randomized Adaptive Search Procedure (GRASP) to generate the initial solution for the aforementioned objective function. GRASP can be understood as an improvement on traditional variable neighborhood search methods for generating initial solutions, aiming to generate higher-quality initial solutions. For example, see [reference needed]. Figure 2 The flowchart of the improved multi-criteria variable neighborhood search algorithm in the variable speed machine scheduling problem shows that step S120 may include the following execution steps: Step S121: Obtain the set of unscheduled job tasks and the machine load list for the transmission.

[0034] Specifically, the list of unscheduled jobs needs to be initialized first. Scheduling schemes and machine load lists.

[0035] Step S122: Determine the priority threshold based on the priority of each job task in the unscheduled job task set, and add job tasks with a priority threshold to the restricted candidate list (RCL).

[0036] You can cycle through the list of unscheduled jobs. Calculate the priority of candidate jobs, for example, using the following formula:

[0037] It is evident that jobs with shorter delivery times and longer processing times have higher priority. Then, candidate jobs can be sorted in descending order of priority to generate a candidate list. Based on the maximum and minimum priority values, a priority threshold is calculated, and jobs with priorities greater than or equal to the priority threshold are added to the restricted candidate list.

[0038] Step S123: Randomly select a job task from the restricted candidate list and add it to the machine with the lowest load in the machine load list.

[0039] That is, a job can be randomly selected from the RCL and added to the list of unscheduled jobs. Remove from the list. Then find the machine with the lowest current load, assign the selected job to that machine, and update the load on that machine.

[0040] Step S124: Update the unscheduled job task set and machine load list, and iterate the above steps until the unscheduled job task set is empty to generate an initial solution.

[0041] Record the selected jobs in the scheduling scheme list, and when the unscheduled jobs list... If the list is empty, return the scheduling scheme as the final initial solution.

[0042] Through the above steps, this invention combines the GRASP greedy random strategy to generate initial solutions, and comprehensively considers delivery time and processing time to adjust job priorities, thereby improving the quality of initial solutions. This allows for further algorithm optimization and efficiency improvement, effectively enhancing both the algorithm's solution efficiency and solution quality.

[0043] Step S130: The variable neighborhood search algorithm based on the adaptive neighborhood selection mechanism intelligently jitters and performs local search on the initial solution or historical optimal solution through iterative search to determine the local optimal solution of the objective function in each iterative search.

[0044] It should be noted that the historical optimal solution is the local optimal solution before the current iteration search.

[0045] The aforementioned variable neighborhood search algorithm can be used to intelligently jitter the initial solution or historical best solution based on the iteration stage of the iterative search. Therefore, an adaptive neighborhood selection mechanism can be used to dynamically adjust neighborhood priorities based on the iteration stage and the historical validity of neighborhoods to balance global exploration and local development. In other words, after generating the initial solution, it is necessary to judge the iteration stage and select the neighborhood structure to use based on the iteration stage.

[0046] In this invention, intelligent jitter can be executed based on a deep learning-driven jitter model, which can also be dynamically updated through online learning. The implementation process of intelligent jitter follows a three-stage framework, dynamically adjusting the jitter strategy by combining deep learning to form a closed-loop optimization mechanism: exploration phase – global exploration, development phase – local refinement, and stagnation phase – breakthrough restart. For details, please refer to [reference needed]. Figure 3 The process flow chart for the shaking operation.

[0047] The variable neighborhood search algorithm in step S130 may include the following steps S131-S133: Step S131: In the case that the iteration phase is the exploration phase, perform global jitter on the initial solution or the historical best solution.

[0048] The neighborhood structure of the algorithm can be adjusted using dithering operations. Dithering operations include inserting and deleting nodes in a single task set or multiple task sets. Specifically, this can involve moving any task node forward or backward, swapping any two task nodes, or flipping part or all of the task node sequence.

[0049] Global jitter requires operations such as insertion, deletion, and swapping of individual job tasks or subsequences to be performed across multiple machines, i.e., multiple tasks in a concentrated manner. Examples include the following: : : Randomly select two different machines, randomly select one task node from each of their task sequences, and swap the positions of these two nodes.

[0050] : Randomly select two different machines, randomly select a task node from the task sequence of one machine, and randomly insert the task node into the task sequence of the other machine.

[0051] : Randomly select two different machines and swap the task nodes with the highest weights between them.

[0052] In one embodiment, performing global jitter during the exploration period may specifically include: 1) outputting the predicted optimal global jitter direction based on the pre-trained model; 2) performing jitter with a first probability using the optimal global jitter direction, and performing jitter randomly with a second probability. The first probability should be greater than 60%, for example, 80%.

[0053] Specifically, such as Figure 4 As shown, the exploration phase can be performed by following these steps: Step S1311: Initialization and data preprocessing.

[0054] Input initial solution Calculate the eigenvectors of the initial solution. This includes global features, machine-level features, and job-level features. An empty dataset is created. It is used to store jitter experience.

[0055] Step S1312: Perform global jitter operation.

[0056] Randomly select two machines to execute the job sequence. Any operation within the process controls the jitter amplitude, modifying a certain proportion of machine job allocation with each jitter to ensure a broad search range. The jitter solution is then obtained. Calculate the objective function value .

[0057] Step S1313: Model pre-training.

[0058] Shaking experience ( Store in dataset In the middle. Among them, The feature vector representing the historical optimal solution includes global features, machine-level features, and job-level features. Indicates the operation number to be performed. This indicates the jitter result, including whether the objective function improves after jitter and the improved value. The dataset is trained using a Long Short-Term Memory (LSTM) network, and the input is the dataset... The output is the probability distribution of the optimal jitter type.

[0059] Step S1314: Differentiated execution of optimal jitter.

[0060] For example, the optimal jitter direction predicted by the model is used with an 80% probability (first probability), and random jitter is used with a 20% probability (second probability), thereby achieving a balance between exploration and utilization.

[0061] It is evident that during the exploration phase, a collaborative approach of global jitter and jitter direction pre-training can be adopted. Global jitter expands the search range, while historical data is used to train a deep learning model, thereby predicting effective jitter directions.

[0062] Step S132: When the iteration phase is the development phase, perform local jittering on the historical best solution. Local jittering requires performing the following operations on a single machine, i.e., a single task set: insertion, swapping, flipping subsequences, etc., for example including the following: : : Randomly select a task node from the task sequence on a machine and swap its position with any task node that was processed before it but has a lower weight.

[0063] : Delete a task node from the task sequence of the same machine and insert it into a random, feasible position in the task sequence.

[0064] Reverse all task sequences on a machine.

[0065] : Randomly select three task nodes from the task sequence of the same machine, divide the original task sequence into three segments, and combine them in a random manner to output the arrangement with the shortest total weighted comprehensive cost.

[0066] Select the processing task with the highest weight in a machine and move its processing order forward by 1 to 3 positions.

[0067] Based on the current task sequence, calculate the contribution of each task to the total weighted composite cost (based on the current completion time × weight), select the task with the highest contribution, and move it forward 1 to 3 places.

[0068] In one embodiment, performing local jitter during development may include: 1) outputting the predicted optimal local jitter direction based on a pre-trained model that is updated iteratively a set number of times; and 2) performing jitter using the optimal local jitter direction.

[0069] Specifically, such as Figure 5 As shown, the development phase can be performed by following these steps: Step S1321: Perform local fine jitter operation.

[0070] Randomly select a machine and execute the job sequence for that machine. Each operation in the process modifies only a portion of the job position on a single machine for each jitter, ensuring fine-grained optimization. This yields the jitter solution. Calculate the objective function value .

[0071] Step S1322: Online update of deep learning model.

[0072] After each jitter, if the solution quality improves, the current state and operation ( The data is stored in an online buffer. For example, every 10 iterations, the model parameters are fine-tuned using the data in the online buffer, and the jitter direction prediction capability is updated.

[0073] Step S1323: Prediction based on model.

[0074] The type of local jitter that has the greatest impact on the objective function can be prioritized. In another embodiment, random jitter can also be triggered when the confidence of the optimal operation predicted by the model is lower than a threshold to increase diversity and avoid over-reliance on model prediction.

[0075] Step S1324: Perform optimal jitter.

[0076] It is evident that during the development phase, a collaborative approach of local fine-tuning and online learning was adopted. Deep learning models guided local optimization, while new data was used to dynamically update the model, thereby improving the accuracy of jittering.

[0077] Step S133: When the iteration phase is in a stagnation period, perform adversarial jittering or forced jittering on the historical best solution.

[0078] Specifically, such as Figure 6 As shown, the following steps can be performed during the standby period: Step S1331: Generate adversarial jitter.

[0079] Generative adversarial networks are used to generate counterintuitive jitter directions, given the current solution. The system outputs abnormal jitter operations. It determines whether the jitter is valid and guides the generator to explore unconventional directions. The generator and discriminator are trained alternately, and the jitter generated by the adversarial network is applied to the current solution to obtain... .

[0080] Step S1332: Forced shaking.

[0081] Retain the job assignment with the shortest completion time from the historical best solution, randomly select machines to assign high-latency jobs, completely reset their speed and job sequence, and partially shuffle the global job assignments. If the solution after shuffling is worse than the historical best solution, retain the original solution.

[0082] Step S1333: Result feedback.

[0083] If the objective function improves after adversarial or forced jittering, return to the development phase; otherwise, terminate the algorithm and output the optimal solution.

[0084] It is evident that during the stagnation period, a combination of forced jitter and adversarial jitter generation is required. Adversarial deep learning is used to generate counterintuitive jitter directions, thereby breaking the stagnation of the solution.

[0085] In summary, this invention introduces an intelligent jittering mechanism into the jittering process of the variable neighborhood search algorithm. By combining a multi-stage dynamic jittering strategy with deep learning models to predict jittering directions, it enhances the algorithm's ability to conduct directional searches in the solution space and improves its efficiency in escaping local optima. During the exploration phase, an LSTM model is used to store the mapping relationship between jittering operations and changes in the objective function, which helps improve the effectiveness of jittering. During the development phase, the model is dynamically updated through online learning, prioritizing local operations that have the greatest impact on the objective function. During the stagnation phase, adversarial jittering is employed to help break local optimum deadlocks.

[0086] This invention specifically divides the jittering process into three stages, using this three-stage design to gradually narrow the search range and improve accuracy. In the exploration phase, a collaborative approach of global jittering and jitter direction pre-training is adopted. Global jittering expands the search range, while historical data is used to train a deep learning model to predict effective jittering directions. In the development phase, a collaborative approach of local fine-tuning and online learning is adopted. A deep learning model guides local optimization, while new data is used to dynamically update the model, improving jittering accuracy. In the stagnation phase, a forced jittering and adversarial jittering generation approach is adopted. Adversarial deep learning generates counterintuitive jittering directions, breaking the stagnation of solutions.

[0087] Additionally, in step S130, the local search may also include the following steps S134-S137: Step S134: Intelligent jittering is performed on the historical optimal solution according to the iteration stage to generate a jittered solution.

[0088] Specifically, the historical optimal solution can be intelligently jittered through steps S131-S133 above, depending on the different iteration stages. Specifically, for the input historical optimal solution... and the generated jitter solution You can first determine the number of iterations based on the current iteration count. and stagnation count The number of times is compared with the corresponding threshold to determine the current stage. If (Threshold number of times during the exploration period) and (If the first stagnation threshold number is 5, for example), enter the exploration phase and perform global jitter; if (Threshold number of times during development) and (The second stagnation threshold number, for example, is 20), then local jitter is executed during the development phase; if or continuous If, for example, there is no improvement after 30 consecutive attempts, an adversarial jitter is implemented during the stagnation period.

[0089] To improve the effectiveness of the algorithm, this invention designs an adaptive neighborhood selection mechanism based on the traditional variable neighborhood search, and designs a dynamic selection strategy to adjust the neighborhood priority according to the iteration stage and the historical effectiveness of the neighborhood, so as to balance global exploration and local development.

[0090] For example, the neighborhood selection strategy based on the iterative stage can be designed using the following formula:

[0091] This represents the selection of a neighborhood probability vector, where α and β are probability adjustment coefficients. It can be between 0.1 and 0.3, for example, 0.2. The value can be between 0.05 and 0.2, for example, 0.08. In the early stages of the algorithm, i.e., the exploration phase, the focus is on inter-machine perturbations. Once N7-N9 are selected, their probability increases, allowing for broader exploration and the discovery of new solutions. Conversely, when intra-machine perturbations N1-N3 are selected, their probability decreases, ultimately achieving the goal of rapidly exploring the global solution space during the exploration phase. This enables the discovery of new solutions based on the global neighborhood structure, avoiding local optima. The selection probability of neighborhoods is dynamically adjusted to increase the global neighborhood selection probability and decrease the local neighborhood selection probability. Refer to the following formula:

[0092] in, This represents the effectiveness of each neighborhood, recording the number of times that neighborhood generates a better solution. In the later stages of the algorithm, i.e., the development phase, the current solution is finely optimized using the local neighborhood structure. When the number of iterations exceeds 70% of the total number of iterations or there are no consecutive iterations of improvement... If the maximum number of attempts is reached, then based on The selection process employs a roulette wheel selection method, prioritizing neighborhoods with high historical validity.

[0093] Step S135: Perform local optimization on the jitter solution to determine the current optimized solution of the objective function.

[0094] Through local optimization, the current optimized solution of the objective function, which is initially the jitter solution, can be calculated; this can also be called the new solution s. new And calculate its objective function value f(s) new ).

[0095] Step S136: If the current optimized solution is better than the historical best solution, the current optimized solution is determined as the local optimal solution.

[0096] Compare the current optimal solution with the historical best solution. If f(s) new ) <f(s best If so, then update. That is, the current optimal solution is determined as a local optimum.

[0097] Step S137: If the current optimal solution is worse than the historical optimal solution, the historical optimal solution is determined as a local optimal solution, and the number of stagnations is increased by one.

[0098] If f(s) new )>f(s best If the original historical best solution is not found, then the original best solution is retained as a local best solution, and an additional stagnation count is added.

[0099] Additionally, if the maximum number of iterations is reached... or continuous If, for example, no improvement is achieved in 50 consecutive iterations, the algorithm needs to be terminated; otherwise, it will return to the exploration phase and re-enter the next iteration.

[0100] In another embodiment, the iterative search in step S130 of this invention, the dynamic update design of the neighborhood effectiveness in the variable neighborhood search algorithm also includes a stage switching mode. This invention designs two stage switching modes: automatic switching based on iteration progress and forced switching based on solution stagnation. Since the algorithm needs to extensively explore the solution space in the early stages and improve solution quality through local optimization in the middle and later stages, it is set that when the number of iterations exceeds 70% of the total number of iterations, the algorithm switches from the exploration phase to the development phase. To avoid the algorithm getting trapped in local optima, a limit is set on the number of consecutive unimproved iterations. When the limit is exceeded, the algorithm is forcibly switched from the development phase to the exploration phase. Specifically, this includes the following steps S138-S139: Step S138: If the iteration progress is greater than the first set value, switch the iteration phase to the development phase.

[0101] This is because the algorithm needs to extensively explore the solution space in the early stages, while in the middle and later stages, it needs to optimize locally to improve the solution quality. Therefore, a setting can be made to switch the iteration phase from the exploration phase to the development phase when the number of iterations exceeds a first predetermined value (usually greater than 60%, such as 70%). Understandably, the iteration phase can automatically switch when the number of iterations or stalls reaches the phase switching condition. This strategy avoids over-exploration in the early stages that could lead to missing the global optimal solution, and allows for a focus on local optimization in the later stages, improving solution quality and convergence speed.

[0102] Step S139: If the number of stalls exceeds the second set value, switch the iteration phase to the exploration phase.

[0103] Understandably, this forced switching based on solution stagnation is a strategy that forces the algorithm to jump out of its current state and restart the global search when it gets stuck in a local optimum, thus preventing the algorithm from getting trapped in local optima. For example, a limit can be set on the number of consecutive unimproved iterations (a second set value, such as 20-50 times). When the number of consecutive unimproved iterations exceeds 50, if a new solution improves the solution, the number of consecutive unimproved iterations can be reset to 0, and the jittering strategy during the development phase can continue. If the new solution still does not improve the solution, then during the stagnation period, a global jittering strategy is needed to break the deadlock and return to the exploration phase, performing random jittering or adversarial jittering. This strategy can prevent the algorithm from prematurely converging due to local optima and maintain search diversity by forcing global jittering.

[0104] In summary, this invention introduces two modes: automatic switching based on iteration progress and forced switching based on unstagnation. By combining these two strategies, the search behavior is dynamically adjusted, balancing global exploration and local development, and timely adjustments to the search strategy to avoid getting trapped in local optima.

[0105] In addition, you can continue to refer to Figure 2 In step S130 of the present invention, the execution steps of the variable neighborhood search algorithm based on the adaptive neighborhood selection mechanism in the overall optimization process are as follows: 1) Define a neighborhood set; local neighborhoods can be selected. As a single-machine operation, it is used for fine-grained optimization; the global neighborhood can be selected. As a multi-machine operation, it is used to expand the search scope.

[0106] 2) Set an effective counter for each neighborhood. This records the number of times the neighborhood generates a better solution; for example, it can be used to initialize the selection probability vector. [ Initially, all probabilities are the same; set an iteration phase marker. , As an exploratory phase, During the development phase, the initial value is 0; define a counter. , is used to record the number of consecutive unimproved iterations, and is initially set to 0.

[0107] 3) According to and Update selection probability .when , Updated to the probability distribution during the exploration period; when , Updated to a probabilistic roulette wheel during development, prioritizing neighborhoods with high historical validity.

[0108] 4) By probability Select a neighborhood structure, and perform a jittering process and a local search process based on the selected neighborhood structure, updating the solution based on its quality. and .

[0109] 5) If the number of iterations exceeds 70% of the total number of iterations, the algorithm switches from the exploration phase to the development phase. If the number of consecutive unimproved iterations exceeds a threshold, the algorithm switches back to the exploration phase and randomly jitters the current solution.

[0110] 6) The algorithm terminates when the running time exceeds the time limit or the number of iterations reaches the limit.

[0111] In summary, step S130 is the core and key step of this invention, which can achieve algorithm optimization and efficiency improvement. It includes: designing an adaptive neighborhood selection mechanism to dynamically adjust the neighborhood priority based on the iteration stage and the effectiveness of the neighborhood history; introducing two modes, automatic switching based on iteration progress and forced switching based on solution stagnation, to adjust the search strategy in a timely manner and avoid getting trapped in local optima; and enhancing the global search capability and stability of the algorithm, thereby effectively improving the solution efficiency and solution quality of the algorithm.

[0112] Furthermore, the deep learning-driven jitter strategy proposed in this invention is groundbreaking in the field of intelligent control systems. Traditional jitter algorithms often rely on fixed thresholds or simple probabilistic models, making them difficult to adapt to complex dynamic environments. This technology constructs a temporal memory module through an LSTM network, establishing an operation-effect mapping database during the exploration phase, enabling the system to predict the potential impact of different jitter strategies on the objective function. In the development phase, the system employs an incremental learning mechanism to continuously optimize the decision path, gradually reducing the proportion of blind random searches by analyzing the effectiveness distribution of historical operations. To address the common local optima dilemma during optimization, an adversarial generation mechanism is innovatively introduced, using targeted perturbations to help the system escape suboptimal states. This hierarchical and progressive intelligent jitter architecture demonstrates significant advantages in various practical application scenarios, achieving a convergence speed improvement of over 40% under ideal conditions compared to traditional methods.

[0113] Step S140: If the iterative search reaches the termination condition, terminate the iteration and output the global optimal solution of the objective function.

[0114] Termination conditions may include: the running time of the iterative search exceeds the time limit; and / or the number of iterations of the iterative search reaches the limit.

[0115] Through the above steps, this invention provides enterprises with flexible production scheduling strategy options. It can quickly generate corresponding scheduling schemes by adjusting weighting coefficients based on the objectives of different production stages, meeting the diverse needs of enterprises. By optimizing scheduling, it reduces operational delays, improves machine utilization, lowers energy consumption and carbon emissions, and achieves efficient resource utilization and sustainable environmental development, resulting in significant economic and environmental benefits.

[0116] In summary, this invention comprehensively considers carbon emission costs and total weighted delay time in the scheduling problem, achieving a balance between production efficiency and environmental protection requirements through an improved variable neighborhood search algorithm. Compared with traditional methods, this invention exhibits significant advantages in several aspects, with key points and beneficial technical effects as follows: 1. Multi-criteria weighted balance optimization enables model innovation and flexibility. The mixed-integer programming model constructed in this invention comprehensively considers carbon emission costs and total weighted delay time, more closely aligning with the low-carbon and efficient scheduling requirements of actual production. It introduces delay time weights and carbon emission cost weights to construct a weighted combined objective function that includes total weighted delay time and carbon emission costs. This allows for flexible balancing of the relative importance of total weighted delay time and carbon emission costs, achieving a dynamic balance between production efficiency and environmental costs. Therefore, it can be flexibly adapted to enterprise needs by adjusting the weights to suit efficiency-first, environmental-first, or balanced production strategies, providing enterprises with a more comprehensive basis for production decision-making.

[0117] 2. Initial solution generation optimization This invention proposes to generate initial solutions using a GRASP greedy random strategy, comprehensively considering delivery time and processing time to adjust job priorities and improve the quality of initial solutions. Simultaneously, it combines weighted adjustments to job priorities to enhance the diversity and quality of solutions.

[0118] 3. Adaptive neighborhood search enables algorithm optimization and efficiency improvement. An adaptive neighborhood selection mechanism is designed to dynamically adjust the neighborhood search strategy based on the iteration stage and the weight of the historical validity of the neighborhood. Through a roulette wheel selection based on historical validity, it quickly converges to a high-quality solution, balancing global exploration and local development. Furthermore, this invention introduces two modes: automatic switching based on iteration progress and forced switching based on solution stagnation. Intelligent jittering dynamically switches between exploration, development, and stagnation stages to balance global exploration and local development, avoiding getting trapped in a single search mode and promptly adjusting the search strategy to avoid getting stuck in local optima.

[0119] 4. Intelligent jitter mechanism based on deep learning This invention introduces a deep learning-driven jittering strategy. During the exploration phase, an LSTM model is used to store the mapping relationship between jittering operations and changes in the objective function, predicting efficient jittering directions and improving jittering effectiveness. During the development phase, the model is dynamically updated through online learning, and the search path is optimized by combining historical jittering experience to reduce the proportion of random jittering and prioritize local operations that have the greatest impact on the objective function. During the stagnation phase, adversarial jittering is used to break the deadlock of local optima.

[0120] In addition, the present invention also provides a scheduling optimization system 200 based on multi-criteria variable neighborhood search, such as Figure 6 As shown, the optimization system 200 may include: The objective function construction device 210 is used to construct the objective function and its constraint function for the overall cost of the transmission, wherein the overall cost is a weighted sum of the delay time of the operation task and the carbon emission cost of the machine; The initial solution generation device 220 is used to generate an initial solution that satisfies the constraint function through a sorting scoring function based on a greedy random construction method; The local optimum generation device 230 is used in a variable neighborhood search algorithm based on an adaptive neighborhood selection mechanism. It performs intelligent jittering and local search on the initial solution or historical optimum through iterative search to determine the local optimum of the objective function in each iteration. The historical optimum is the local optimum before the current iteration. The global optimal solution generation device 240 is used to terminate the iteration and output the global optimal solution of the objective function when the iterative search reaches the termination condition.

[0121] In one embodiment, the objective function and its constraint functions are represented by the following equation: (1)

[0122] (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) in, Indicates the first One machine, ; Indicates the first This assignment, ; Indicate homework The weights; Indicates the weight of the delay time; Indicates the carbon emission cost weight, ; Indicate homework Completion time; Indicate homework The delay time; Indicate homework In the machine Processing time; Indicates machine The fixed carbon emission costs, including start-up carbon emission costs; Indicates machine The unit carbon emission cost during runtime; S0 is the initial solution; S is the current solution; T(S) is the total delay time of all tasks under the current solution; C(S) is the total machine carbon emission cost of all tasks under the current solution.

[0123] In one embodiment, the initial solution generation device is further configured to perform the following functions: obtain an unscheduled job task set and a machine load list for the transmission; determine a priority threshold based on the priority of each job task in the unscheduled job task set, and add job tasks with priorities greater than the priority threshold to a restricted candidate list; randomly select a job task from the restricted candidate list to add to the machine load list of the machine with the lowest load; and update the unscheduled job task set and the machine load list, iterating the above steps until the unscheduled job task set is empty, so as to generate an initial solution.

[0124] In one embodiment, the intelligent jittering is performed based on a deep learning-driven jittering model, which is dynamically updated through online learning. The local search performed by the local optimum generation device includes the following functions: intelligently jittering the historical optimum based on the iterative stage of the iterative search to generate a jittered solution; locally optimizing the jittered solution through local search to determine the current optimized solution of the objective function; if the current optimized solution is better than the historical optimum, determining the current optimized solution as the local optimum; if the current optimized solution is worse than the historical optimum, determining the historical optimum as the local optimum and adding one stagnation count.

[0125] In one embodiment, during the iterative search performed by the local optimal solution generation device, the variable neighborhood search algorithm further includes: switching the iterative phase to the development phase when the iteration progress is greater than a first preset value; or switching the iterative phase to the exploration phase when the number of stalls is greater than a second preset value.

[0126] In one embodiment, the variable neighborhood search algorithm executed by the local optimal solution generation device is used to perform: intelligent jittering of the initial solution or the historical optimal solution according to the iterative stage of the iterative search, including at least one of the following: when the iterative stage is an exploration period, performing global jittering of the initial solution or the historical optimal solution; when the iterative stage is a development period, performing local jittering of the historical optimal solution; when the iterative stage is a stagnation period, performing adversarial jittering or forced jittering of the historical optimal solution.

[0127] In one embodiment, the global jitter is to perform the following operations across multiple machines: inserting or swapping a single job or subsequence, wherein performing the global jitter during the exploration period includes: outputting a predicted optimal global jitter direction based on a pre-trained model; and performing jitter with a first probability using the optimal global jitter direction and performing jitter randomly with a second probability, the first probability being greater than 60%.

[0128] In one embodiment, the local jitter is performed on a single machine by inserting, swapping, or flipping subsequences, wherein performing the local jitter during the development phase includes: outputting a predicted optimal local jitter direction based on a pre-trained model that is updated iteratively a set number of times; and performing jitter using the optimal local jitter direction.

[0129] In one embodiment, the pre-trained model includes: incorporating jitter experience ( Store in dataset In, among them, The feature vector representing the historical optimal solution includes global features, machine-level features, and job-level features. Indicates the operation number to be performed. The results of the jittering include whether the objective function has improved after jittering and the improved value; and the dataset... As input, the model is trained using the LSTM network training dataset to construct the first pre-trained model and output the optimal jitter type probability distribution.

[0130] The beneficial effects of the scheduling optimization system based on multi-criteria variable neighborhood search of the present invention can be found in the above description of the variable speed machine scheduling optimization method, and will not be repeated here.

[0131] In addition, embodiments of this application also provide a machine-readable storage medium storing instructions for causing a machine to execute the above-described scheduling optimization method based on multi-criteria variable neighborhood search.

[0132] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0133] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0134] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0135] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps that specify the function are in one or more boxes.

[0136] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0137] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0138] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0139] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0140] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A scheduling optimization method based on multi-criteria variable neighborhood search, used for variable speed machines, characterized in that, The optimization method includes: Construct an objective function and its constraint function for the overall cost of the transmission, wherein the overall cost is a weighted sum of the delay time of the task and the carbon emission cost of the machine; An initial solution satisfying the constraint function is generated by using a sorting and scoring function based on a greedy random construction method; A variable neighborhood search algorithm based on an adaptive neighborhood selection mechanism intelligently jitters and performs local search on the initial solution or historical optimal solution through iterative search to determine the local optimal solution of the objective function in each iteration search, wherein the historical optimal solution is the local optimal solution before the current iteration search; and If the iterative search reaches the termination condition, the iteration terminates and the global optimal solution of the objective function is output.

2. The optimization method according to claim 1, characterized in that, The objective function and its constraint functions are expressed by the following equation: ， ， ， ， ， ， ， ， ， ， ， ， in, Indicates the first One machine, ; Indicates the first This assignment, ; Indicate homework The weights; Indicate homework Completion time; Indicate homework The delay time; Indicate homework In the machine Processing time; Indicates machine The fixed carbon emission costs, including start-up carbon emission costs; Indicates machine Unit carbon emission cost during operation; Indicates the weight of the delay time; Indicates the carbon emission cost weight, S0 is the initial solution; S is the current solution; T(S) is the total delay time of all tasks under the current solution; C(S) is the total machine carbon emission cost of all tasks under the current solution.

3. The optimization method according to claim 1, characterized in that, The step of generating an initial solution that satisfies the constraint function through a sorting and scoring function based on a greedy random construction method includes: Obtain the set of unscheduled job tasks and the machine load list for the transmission; Based on the priority of each task in the unscheduled task set, a priority threshold is determined, and tasks with a priority higher than the priority threshold are added to the restricted candidate list. Randomly select a job task from the restricted candidate list and add it to the machine with the lowest load in the machine load list; and Update the unscheduled job task set and the machine load list, and iterate the above steps until the unscheduled job task set is empty to generate an initial solution.

4. The optimization method according to claim 1, characterized in that, The intelligent jittering is performed based on a deep learning-driven jittering model, which is dynamically updated through online learning. The local search includes: The historical optimal solution is intelligently jittered according to the iterative stage of the iterative search to generate a jittered solution; Local optimization is performed on the jittered solution to determine the current optimized solution of the objective function; If the current optimal solution is better than the historical optimal solution, the current optimal solution is determined as the local optimal solution; If the current optimal solution is inferior to the historical optimal solution, the historical optimal solution is determined as the local optimal solution, and the number of stagnations is increased by one.

5. The optimization method according to claim 4, characterized in that, The intelligent jittering of the historical optimal solution based on the iterative stage of the iterative search includes at least one of the following: If the iteration phase is an exploratory phase, global jitter is performed on the initial solution or the historical best solution; If the iteration phase is the development phase, local jitter is performed on the historical optimal solution; If the iteration phase is a standstill period, adversarial jitter or forced jitter is performed on the historical best solution.

6. The optimization method according to claim 5, characterized in that, In the iterative search, the variable neighborhood search algorithm further includes: If the iteration progress exceeds a first set value, the iteration phase will be switched to the development phase; or If the number of pauses exceeds a second set value, the iteration phase will be switched to the exploration phase.

7. The optimization method according to claim 5, characterized in that, The global jitter refers to performing the following operations across multiple machines: inserting or swapping individual jobs or subsequences. The global jittering during the exploration period includes: Based on the pre-trained model, the predicted optimal global jitter direction is output; and Jitter is performed with a first probability using the optimal global jitter direction, and jitter is performed randomly with a second probability, wherein the first probability is greater than 60%.

8. The optimization method according to claim 5, characterized in that, The local jitter refers to performing the following operations between single machines: inserting, swapping, and flipping subsequences. The local jittering performed during the development phase includes: Based on a pre-trained model that is updated iteratively a set number of times, the predicted optimal local jitter direction is output; and Jitter is performed using the optimal local jitter direction.

9. The optimization method according to claim 7 or 8, characterized in that, The pre-trained model includes: Shaking experience ( Store in dataset In, among them, The feature vector representing the historical optimal solution includes global features, machine-level features, and job-level features. Indicates the operation number to be performed. The results of the dithering are indicated, including whether the objective function has improved after dithering and the improved value; and The dataset As input, the model is trained using the LSTM network training dataset to construct the first pre-trained model and output the optimal jitter type probability distribution.

10. A scheduling optimization system based on multi-criteria variable neighborhood search, used for a variable speed engine, characterized in that, The optimization system includes: An objective function construction device is used to construct an objective function and its constraint function for the overall cost of the transmission, wherein the overall cost is a weighted sum of the delay time of the operation task and the carbon emission cost of the machine; An initial solution generation device is used to generate an initial solution that satisfies the constraint function by using a sorting scoring function based on a greedy random construction method; A local optimum generation device is used in a variable neighborhood search algorithm based on an adaptive neighborhood selection mechanism. It performs intelligent jittering and local search on the initial solution or historical optimum through iterative search to determine the local optimum of the objective function in each iteration. The historical optimum is the local optimum before the current iteration. A global optimal solution generation device is used to terminate the iteration and output the global optimal solution of the objective function when the iterative search reaches the termination condition.

11. A machine-readable storage medium storing instructions for causing a machine to execute: the scheduling optimization method based on multi-criteria variable neighborhood search according to any one of claims 1-9.

Images