Electric power system maintenance plan intelligent arrangement method and system based on cooperation of large and small models

Abstract

The invention provides an electric power system maintenance plan intelligent arrangement method and system based on cooperation of large and small models, and belongs to the technical field of electric power system operation and maintenance. Initial maintenance plans of all equipment units are read, plan relations are judged, and conversion arrangement information is input into an arrangement small model; performing preliminary arrangement on the maintenance plan by using a deep reinforcement learning algorithm, forming a primary maintenance plan and returning the primary maintenance plan to the maintenance large model, analyzing opinions or new requirements of departments of the balance conference on the primary maintenance plan, converting the opinions or new requirements into adjustment codes, performing secondary or multiple arrangement on the primary maintenance plan by using the adjusted deep reinforcement learning algorithm, and returning the arrangement result to the maintenance large model, and analyzing the arrangement result. According to the invention, through cooperation of the large language model and the deep reinforcement learning model, intelligent arrangement, dynamic adjustment and multi-objective optimization of the maintenance plan are realized, and the efficiency, expansibility and economy of the arrangement of the maintenance plan are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system operation and maintenance, and particularly relates to an intelligent scheduling method and system for power system maintenance plans based on the collaboration of large and small models. Background Art

[0002] As a key link to ensure the reliable operation of power equipment, the scientific and reasonable scheduling of maintenance plans can not only ensure the safe and stable operation of the system, but also improve power supply reliability, optimize the power outage scope and duration, and enhance economic benefits. However, with the continuous expansion of the power grid scale and the continuous increase in the number of equipment, the traditional maintenance plan scheduling method has been difficult to meet the development needs of the current power system.

[0003] Currently, there are mainly the following problems in the scheduling of maintenance plans: First, the scheduling of maintenance plans still mainly relies on manual scheduling. When faced with a large number of maintenance tasks and complex constraint conditions, manual scheduling is not only inefficient but also prone to omissions. Second, the maintenance plan needs to take into account multiple objectives such as workload balance, economy, and reliability at the same time, and it is difficult for traditional methods to achieve the dynamic balance and optimization of multiple objectives. Third, during the actual maintenance process, it is often necessary to adjust the plan according to the on-site situation, and the traditional method has a weak response ability to such dynamic changes. In addition, the scheduling of maintenance plans involves a large amount of professional knowledge and experience accumulation, but currently, there is a lack of effective knowledge precipitation and experience inheritance mechanism, resulting in the scheduling quality being overly dependent on personal experience. Finally, the maintenance plan involves various complex constraint conditions such as mutual dependence between equipment, resource limitations, and time windows, and it is difficult for traditional methods to fully consider these constraint requirements.

[0004] In recent years, the rapid development of artificial intelligence technology has provided new ideas for solving the above problems. Among them, large language models have shown excellent capabilities in natural language understanding, knowledge processing, and logical reasoning, and can effectively understand and transform various unstructured constraint conditions, and analyze and store dispatching rules and regulations; while deep reinforcement learning has also shown strong performance in the field of complex decision-making optimization, and is particularly suitable for dealing with the scheduling problem under multi-objective constraints. More importantly, large language models provide a natural language interaction interface, enabling non-computer professional maintenance personnel to easily participate in the scheduling and adjustment process, breaking through the limitations of professional technical barriers in traditional methods. However, there has been no research on effectively combining these two advanced technologies and applying them to the field of maintenance plan scheduling. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent scheduling method and system for power system maintenance plans based on large and small models. Through the collaborative cooperation of large language models and deep reinforcement learning models, intelligent scheduling, dynamic adjustment, and multi-objective optimization of maintenance plans are realized, improving the efficiency, scalability, and economy of maintenance plan scheduling to solve at least one of the technical problems existing in the above background technology.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] In the first aspect, the present invention provides an intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models, including:

[0008] The scheduling large model reads the initial maintenance plans of each equipment unit, judges the plan relationships, and inputs the converted scheduling information into the scheduling small model;

[0009] The scheduling small model receives the construction period and original date scheduling information of the reported plan, and uses the deep reinforcement learning algorithm to preliminarily schedule the maintenance plan to form a preliminary version of the maintenance plan and return it to the maintenance large model;

[0010] The scheduling large model receives and analyzes the opinions or new requirements of each department on the preliminary version of the maintenance plan in the balance meeting, converts them into adjustment codes, uses the adjustment codes to perform collaborative adjustment on the scheduling small model and conducts self-checking;

[0011] The scheduling small model receives the adjustment information from the large model, uses the adjusted deep reinforcement learning algorithm to schedule the preliminary version of the maintenance plan twice or multiple times, and returns the scheduling results to the maintenance large model;

[0012] The maintenance large model receives the scheduling results of the scheduling small model, analyzes the scheduling results, and provides auxiliary decision-making for the dispatching department.

[0013] As a further limitation of the first aspect of the present invention, in the deep reinforcement learning algorithm, the objective function is determined corresponding to the workload fluctuation degree and the plan execution deviation degree respectively; including:

[0014] The objective function is constructed by the formula minF = αf1 + βf2; in the formula, α and β are weight coefficients, f1 is the workload fluctuation degree, and f2 is the plan execution deviation degree;

[0015] By the formula The workload fluctuation degree is constructed; in the formula, w ij is the workload of the jth plan on the ith day, is the average workload, and T is the scheduling period;

[0016] By the formula The plan execution deviation degree is constructed; in the formula, s j is the start time of the jth plan after scheduling, s′j is the start time of the initial reporting plan, λ i is the deviation coefficient.

[0017] As a further limitation of the first aspect of the present invention, the defined constraint conditions specifically include:

[0018] Define the daily work upper limit constraint through the formula ; where ω max is the daily maximum work upper limit value, and T is the scheduling period;

[0019] Define the equipment mutual exclusion constraint through the formula ; where μ is the set of mutual exclusion plan pairs, and s m , s n are the maintenance start times of plans m and n respectively, and d m , d n are the maintenance durations of plans m and n respectively;

[0020] Define the synchronous maintenance constraint through the formula Δ pq = s p - s q = 0, ; where φ is the set of synchronous maintenance plan pairs, and s p , s q are the start times of the p-th and q-th plans respectively;

[0021] Define the non-maintenance period constraint for major projects through the formula s k + d k ≤ {a k} ∨ s k ≥ {a k}, ; where s k is the start time after scheduling of plan k, dk is the duration, and {ak} is the set of non-maintenance dates;

[0022] Define the time-fixed constraint through the formula s r - s′ r = 0, ; where G is the set of plans that need to fix the time, s r is the start time after scheduling of plan r, and s′ r is the start time of the initial reporting plan.

[0023] As a further limitation of the first aspect of the present invention, obtain the state space through the formula S = {N, j, d j , s' j , {wij}, φ, μ}; where N is the total number of plans, j is the current plan number, d j is the duration of the current plan, and s' jis the start time of the initial reporting plan, {w ij} is the daily workload array, φ is the set of synchronous maintenance plan pairs, and μ is the set of mutually exclusive plan pairs;

[0024] The action space is obtained through formula A = {s|s ∈ {1, 2,..., T}}; where T is the scheduling period and s is the start time of maintenance;

[0025] Through formula the reward function is obtained; where λ1, λ1 are penalty weight coefficients, and ε k is the weight coefficient of the k-th type of constraint penalty term, and P k is the k-th type of constraint violation penalty. The penalty terms include the daily work upper limit constraint penalty, the equipment mutual exclusion constraint penalty, the synchronous maintenance constraint penalty, the non-maintenance period constraint penalty for major projects, and the time-fixed constraint penalty.

[0026] As a further limitation of the first aspect of the present invention, after the scheduling large model reads the initial maintenance plan information, through natural language processing capabilities and topological information, it judges the relationships of the maintenance plans, identifies the maintenance plan combinations that need to be merged and the mutually exclusive maintenance plan combinations; subsequently, it converts the judgment results into standardized scheduling information.

[0027] As a further limitation of the first aspect of the present invention, the initial maintenance plan information includes: maintenance plan number, total number of plans, equipment type, equipment index number, affiliated substation, original reported start maintenance time, maintenance duration, and remarks.

[0028] In the second aspect, the present invention provides a power system maintenance plan intelligent scheduling system based on the cooperation of large and small models, including:

[0029] A reading module for the scheduling large model to read the initial maintenance plans of each equipment unit, judge the plan relationships, and input the converted scheduling information into the scheduling small model;

[0030] A preliminary scheduling module for the scheduling small model to receive the construction period and the original scheduled date scheduling information of the reported plan, and use the deep reinforcement learning algorithm to preliminarily schedule the maintenance plan to form a preliminary version of the maintenance plan and return it to the maintenance large model;

[0031] A conversion module for the scheduling large model to receive and analyze the opinions or new requirements of each department in the balance meeting on the preliminary version of the maintenance plan, convert them into adjustment codes, use the adjustment codes to perform collaborative adjustment on the scheduling small model and perform self-checking;

[0032] A second scheduling module for the scheduling small model to receive the adjustment information of the large model, use the adjusted deep reinforcement learning algorithm to perform secondary or multiple scheduling on the preliminary version of the maintenance plan, and return the scheduling result to the maintenance large model;

[0033] An analysis module for checking and receiving the scheduling results of the small models arranged by the large model, analyzing the scheduling results, and providing auxiliary decision-making for the scheduling department.

[0034] In a third aspect, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models as described in the first aspect.

[0035] In a fourth aspect, the present invention provides a computer device including a memory and a processor, where the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models as described in the first aspect.

[0036] In a fifth aspect, the present invention provides an electronic device including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory so that the electronic device executes instructions to implement the intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models as described in the first aspect.

[0037] The beneficial effects of the present invention: Innovatively construct a collaborative architecture of a large language model and a deep reinforcement learning model: the large language model is responsible for understanding and processing various types of unstructured information, including judgment of maintenance plan relationships, analysis of meeting opinions, and result interpretation, etc.; the deep reinforcement learning model focuses on optimizing and solving plans under multiple constraints to achieve intelligent scheduling of maintenance plans. The two models achieve effective collaboration through a specific information conversion mechanism, which not only ensures the intelligent level of the system but also provides a friendly human-computer interaction interface. The present invention can not only automatically complete the preliminary scheduling of maintenance plans but also dynamically adjust according to actual needs, and at the same time considers multiple optimization goals such as workload balance, economy, and reliability.

[0038] The advantages of the additional aspects of the present invention will be more clearly given in the following description part or learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0040] Figure 1Flow chart of the intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models according to the embodiments of the present invention.

[0041] Figure 2 Interactive schematic diagram of the overhaul large model analyzing the initial scheduling result according to the embodiments of the present invention.

[0042] Figure 3 Interactive schematic diagram of the scheduling large model modifying the code of the small model and checking through the self-check module according to the embodiments of the present invention.

[0043] Figure 4 Interactive schematic diagram of the overhaul large model analyzing the final scheduling result according to the embodiments of the present invention.

[0044] Figure 5 Interactive schematic diagram of the large model comprehensively analyzing and evaluating the scheduling result according to the embodiments of the present invention.

[0045] Figure 6 Comparison chart of the original result, initial scheduling result and final scheduling result according to the embodiments of the present invention. Detailed implementation manners

[0046] The following details the implementation manners of the present invention. The examples of the implementation manners are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The implementation manners described through the drawings are exemplary and are only used to explain the present invention, and cannot be construed as a limitation of the present invention.

[0047] Those skilled in the art of the present technology can understand that unless otherwise defined, all terms (including technical terms and scientific terms) used here have the same meaning as the general understanding of those of ordinary skill in the art in the field to which the present invention belongs.

[0048] It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless defined as here.

[0049] Those skilled in the art of the present technology can understand that unless specifically stated, the singular forms "a", "an", "the" and "said" used here may also include the plural forms. It should be further understood that the term "comprising" used in the description of the present invention means the presence of the described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements and / or groups thereof.

[0050] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0051] To facilitate the understanding of the present invention, the following will further explain the present invention with specific embodiments in conjunction with the accompanying drawings, and the specific embodiments do not constitute a limitation to the embodiments of the present invention.

[0052] Those skilled in the art should understand that the drawings are only schematic diagrams of the embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0053] The present invention proposes an intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models. This method constructs a collaborative architecture of a large language model and a deep reinforcement learning model, gives full play to the advantages of the two models, and realizes the intelligent scheduling, dynamic adjustment, and multi-objective optimization of maintenance plans. Among them, the large language model is responsible for understanding and processing various types of unstructured information and converting it into standardized scheduling instructions; the deep reinforcement learning model then solves the optimal maintenance plan scheme based on these instructions while meeting various constraint conditions. Through this collaborative mechanism, not only can the efficiency and quality of maintenance plan scheduling be significantly improved, but also more intelligent and user-friendly technical support can be provided for power system maintenance management.

[0054] Embodiment 1

[0055] In this Embodiment 1, first, an intelligent scheduling system for power system maintenance plans based on the cooperation of large and small models is provided, including: a reading module, which is used to schedule the large model to read the initial maintenance plans of each equipment unit, judge the plan relationships, and input the converted scheduling information into the scheduling small model; a preliminary scheduling module, which is used to schedule the small model to receive the duration and original scheduled date scheduling information of the reported plans, and use the deep reinforcement learning algorithm to preliminarily schedule the maintenance plans to form a preliminary version of the maintenance plan and return it to the maintenance large model; a conversion module, which is used to schedule the large model to receive and analyze the opinions or new requirements of each department in the balance meeting on the preliminary version of the maintenance plan, convert them into adjustment codes, use the adjustment codes to perform collaborative adjustment on the scheduling small model and perform self-checking; a second scheduling module, which is used to schedule the small model to receive the adjustment information of the large model, use the adjusted deep reinforcement learning algorithm to perform secondary or multiple scheduling on the preliminary version of the maintenance plan, and return the scheduling result to the maintenance large model; an analysis module, which is used to schedule the large model to receive the scheduling result of the scheduling small model, analyze the scheduling result, and provide auxiliary decision-making for the dispatching department.

[0056] In this embodiment, based on the above system, an intelligent scheduling method for power system maintenance plans based on the cooperation of large and small models is realized, including: Step S1: Schedule the large model to read the initial maintenance plans of each equipment unit, judge the plan relationships, and input the converted scheduling information into the scheduling small model; Step S2: Schedule the small model to receive the duration, original scheduled date and other scheduling information of the reported plans, and use the deep reinforcement learning algorithm to preliminarily schedule the maintenance plans to form a preliminary version of the maintenance plan and return it to the maintenance large model; Step S3: Schedule the large model to receive and analyze the opinions or new requirements of each department in the balance meeting on the preliminary version of the maintenance plan, convert them into adjustment codes, use the adjustment codes to perform collaborative adjustment on the scheduling small model and perform self-checking; Step S4: Schedule the small model to receive the adjustment information of the large model, use the adjusted deep reinforcement learning algorithm to perform secondary or multiple scheduling on the preliminary version of the maintenance plan, and return the scheduling result to the maintenance large model; Step S5: Schedule the large model to receive the scheduling result of the scheduling small model, analyze the scheduling result, and provide auxiliary decision-making for the dispatching department.

[0057] The initial maintenance plan information read in Step S1 includes:

[0058] Maintenance plan number, total number of plans, equipment type, equipment index number, affiliated substation, original reported start maintenance time, maintenance duration, remarks; After the large model reads the above information, through natural language processing capabilities and topological information, it judges the relationships of the maintenance plans, and identifies the maintenance plan combinations that need to be merged and the mutually exclusive maintenance plan combinations. Subsequently, the large model converts these judgment results into standardized scheduling information and inputs it to the scheduling small model for subsequent processing.

[0059] In steps S2 and S4, design is carried out based on two aspects: resource allocation and practicability, corresponding to the workload fluctuation degree and the planned execution deviation degree respectively to determine the objective function;

[0060] including:

[0061] Construct the objective function through the formula: minF = αf1 + βf2; in the formula, α and β are weight coefficients, f1 is the workload fluctuation degree, and f2 is the planned execution deviation degree.

[0062] Through the formula: Construct the workload fluctuation degree; in the formula, w ij is the workload of the j-th plan on the i-th day, is the average workload, and T is the scheduling period.

[0063] Through the formula: Construct the planned execution deviation degree; in the formula, s j is the start time of the j-th plan after scheduling, s j is the start time of the initially reported plan, and λ i is the deviation coefficient.

[0064] The defined constraint conditions in steps S2 and S4 specifically include:

[0065] Through the formula Define the daily work upper limit constraint; in the formula, ω max is the daily maximum work upper limit value, and T is the scheduling period.

[0066] Through the formula Define the equipment mutual exclusion constraint; in the formula, μ is the set of mutually exclusive plan pairs, and s m , s n are the maintenance start times of plans m and n respectively, and d m , d n are the maintenance durations of plans m and n respectively.

[0067] Through the formula Δ pq = s p - s q = 0, Define the synchronous maintenance constraint; in the formula, φ is the set of synchronous maintenance plan pairs, and s p , s q are the start times of the p-th and q-th plans respectively.

[0068] Through the formula s k + d k ≤ {a k} ∨ s k ≥ {a k}, Major project non-maintenance time period constraint; in the formula, sk The start time after scheduling for plan k, d k is the construction period, {a k} is the set of non-maintainable dates.

[0069] Through formula s r -s′ r = 0, Define the time-fixed constraint; where G is the set of plans that need to fix the time, s r is the start time after scheduling for plan r, s′ r is the start time of the initial reported plan.

[0070] The specific definition of the deep reinforcement learning model in steps S2 and S4 includes:

[0071] Through formula S = {N, j, d j , s' j , {wij}, φ, μ}, obtain the state space; where N is the total number of plans, j is the current plan number, d j is the construction period of the current plan, s' j is the start time of the initial reported plan, {w ij} is the daily workload array, φ is the set of synchronous maintenance plan pairs, and μ is the set of mutually exclusive plan pairs.

[0072] Through formula A = {s|s ∈ {1, 2,..., T}}, obtain the action space; where T is the scheduling period and s is the maintenance start time.

[0073] Through formula obtain the reward function; where λ1, λ1 are penalty weight coefficients, ε k is the weight coefficient of the k-th type of constraint penalty term, P k is the k-th type of constraint violation penalty. The penalty terms include the daily work upper limit constraint penalty, equipment mutual exclusion constraint penalty, synchronous maintenance constraint penalty, major project non-maintainable period constraint penalty, and time-fixed constraint penalty respectively.

[0074] The model saving part of step S2 includes: saving the deep reinforcement learning model after the initial training as a pth file to accelerate the process of re-training and save the experience learned by the deep reinforcement learning model.

[0075] Step S3 specifically includes: the scheduling large model converts the feedback opinions of each department in the balance meeting into standardized adjustment instructions, updates the small model parameters and environment code, and ensures the security and reliability of the output content through the self-check module.

[0076] The self - checking module in step S3 includes the large model comparing and arranging the parameters of the small model and the code of the arrangement environment before and after modification. If it is reasonable, it continues to the next step; otherwise, it retries. If it fails after multiple retries, it returns a failure prompt and stops.

[0077] The output module in step S3 specifically includes: according to the previous conversation interacting with the large model, the output content of the large model is converted into the json file format through the output module, the content output by the large model is saved, and it is converted into Python code by the conversion module.

[0078] The reading module in step S4 additionally includes: the arranged small model reads the Python code converted from the updated parameters and environment output by the large model, retrains using the experience of the initial arrangement, and returns the arrangement result to the large model.

[0079] The retraining module in step S4 additionally includes: the arrangement plan arranges to read the training network and parameters of the initial arrangement model, interacts with the environment training after reading the pth file, and adapts to the relevant requirements put forward by the balance meeting through rapid training and interaction with the new environment according to the experience of the previous deep reinforcement learning model.

[0080] The analysis of the large model for the arrangement result in step S5 also includes: through the reward function part of the small model training, analyzing whether the relevant constraint rules are violated, and through the comparative analysis of the initial arrangement plan and the final arrangement plan, comparing with the requirements of the balance meeting to judge whether the meeting requirements are reasonable.

[0081] Embodiment 2

[0082] In this embodiment, an intelligent arrangement method for the maintenance plan of the power system based on the cooperation of large and small models is provided. The technical solution in this embodiment includes the following concepts:

[0083] (1) Large model (large language model): An artificial intelligence model based on deep learning, with powerful natural language understanding and generation capabilities. In the present invention, it is mainly used to understand the maintenance plan constraint conditions, analyze the feedback from departments, and convert the information into arrangement instructions.

[0084] (2) Small model (deep reinforcement learning model): An intelligent optimization model based on deep learning and reinforcement learning, which learns the optimal strategy by interacting with the environment. In the present invention, it is used to receive arrangement instructions and perform specific arrangement optimization of the maintenance plan.

[0085] (3) Monthly maintenance plan: The equipment maintenance plan formulated monthly in the power system, including maintenance items, time, methods, personnel arrangements, etc.

[0086] (4) Balance meeting: A meeting mechanism for departments to discuss and coordinate the maintenance plan, which is an important link for dynamic adjustment of the plan.

[0087] See Figure 1 , this embodiment provides an intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models, including the following steps: Step S1: The scheduling large model reads the initial maintenance plans of each equipment unit, judges the plan relationships, and inputs the converted scheduling information into the scheduling small model; Step S2: The scheduling small model receives scheduling information such as the number, duration, and original date of the reported plans, and uses the reinforcement learning algorithm to preliminarily schedule the maintenance plan to form a preliminary version of the maintenance plan and return it to the maintenance large model; Step S3: The scheduling large model receives and analyzes the opinions or new requirements of each department on the preliminary version of the maintenance plan in the balance meeting, converts them into adjustment codes, and uses the adjustment codes to collaboratively adjust the scheduling small model and perform self-checking; Step S4: The scheduling small model receives the adjustment information from the large model, uses the adjusted deep reinforcement learning algorithm to schedule the preliminary version of the maintenance plan twice or multiple times, and returns the scheduling results to the maintenance large model; Step S5: The maintenance large model receives the scheduling results from the scheduling small model, analyzes the scheduling results, and provides auxiliary decision-making for the dispatching department.

[0088] The initial maintenance plan information read in Step S1 includes: maintenance plan number, total number of plans, equipment type, equipment index number, affiliated substation, original reported start maintenance time, maintenance duration, remarks; after the large model reads the above information, through natural language processing capabilities, it judges the relationships of the maintenance plans, identifies the maintenance plan combinations that need to be merged and the mutually exclusive maintenance plan combinations. Subsequently, the large model converts these judgment results into standardized scheduling information and inputs it to the scheduling small model for subsequent processing.

[0089] In this embodiment, taking the content of simulating the initial maintenance as an example: the requirements include: Plan 3 must be maintained on the original start date (the 3rd), Plan 2 cannot be maintained on the 14th, Plan 7 needs to be maintained on the 29th in the subsequent arrangements for the initial scheduling, and at the same time, Plan Sets 4 and 5 are synchronous maintenance plan sets, and Plan Sets 10 and 11 and Plan Sets 14 and 15 are mutually exclusive plan sets for equipment.

[0090] Step S2 specifically includes the following execution process:

[0091] Step S21: The scheduling small model receives scheduling information such as the number of maintenance plans given by the large model, the duration of each plan, the original maintenance date, the set of synchronous maintenance plan combinations, the set of mutually exclusive maintenance plan combinations, the set of non-maintenance date constraints, and the daily maximum workload limit, and performs preliminary scheduling.

[0092] Step S22: Design according to two aspects of resource allocation and practicality, and determine the objective function corresponding to the workload fluctuation degree and the plan execution deviation degree respectively. It includes: minF = αf1 + βf2, where α and β are weight coefficients, f1 is the workload fluctuation degree, and f2 is the plan execution deviation degree. Among them, the workload fluctuation degree is represented by the fluctuation degree of the monthly workload: where w ij is the workload of the jth plan on the ith day, is the average workload, and T is the scheduling cycle.

[0093] For N plans, the plan execution deviation degree is: where s j is the start time of the jth plan after scheduling, s′ j is the start time of the initial reported plan, and λ i is the deviation coefficient.

[0094] Step S23: Construct constraint conditions according to the characteristics of the maintenance plan scheduling problem, mainly including the daily work upper limit constraint, equipment mutual exclusion constraint, synchronous maintenance constraint, major project non-maintenance time period constraint, and time fixed constraint.

[0095] The daily work upper limit constraint is: where ω max is the daily maximum work upper limit value, and T is the scheduling cycle.

[0096] The equipment mutual exclusion constraint is: where μ is the set of mutually exclusive plan pairs, s m , s n are the maintenance start times of plans m and n respectively, d m , d n are the maintenance durations of plans m and n respectively.

[0097] Synchronous maintenance constraint: Δ pq = s p - s q = 0, where φ is the set of synchronous maintenance plan pairs, s p , s q are the start times of the pth and qth plans respectively.

[0098] Major project non-maintenance time period constraint: s k + d k ≤ {a k} ∨ s k ≥ {a k}, where sk is the start time of plan k after scheduling, dk is the duration, and {ak} is the set of non-maintenance dates.

[0099] Time-fixed constraint: s r -s' r = 0, where G is the set of plans that need to fix the time, s r is the start time after the scheduling of plan r, and s' r is the start time of the initially reported plan.

[0100] Step S24: Convert the maintenance plan problem into a Markov decision process, and construct a deep reinforcement learning model, including the design of the state space, action space, and reward function. The state space contains the key information of the maintenance plan, the action space represents the set of selectable decisions, and the reward function evaluates the decision quality through immediate rewards and delayed rewards. The specific design is as follows:

[0101] The state space is specifically as follows: S = {N, j, d j , s' j , {wij}, φ, μ}, where N is the total number of plans, j is the current plan number, d j is the current planned duration, s' j is the start time of the initially reported plan, {w ij} is the daily workload array, φ is the set of synchronized maintenance plan pairs, and μ is the set of mutually exclusive plan pairs.

[0102] The action space is specifically as follows: A = {s|s ∈ {1, 2,..., T}}, where T is the scheduling period and s is the maintenance start time.

[0103] The reward function is as follows: where λ1, λ1 are penalty weight coefficients, ε k is the weight coefficient of the k-th type of constraint penalty term, and P k is the k-th type of constraint violation penalty. The penalty terms include the daily work upper limit constraint penalty, equipment mutual exclusion constraint penalty, synchronized maintenance constraint penalty, non-maintenance period constraint penalty for major projects, and time-fixed constraint penalty.

[0104] Step S25: Use the deep reinforcement learning algorithm for solution and output the initial version of the maintenance plan, including the scheduled time of each plan, the workload fluctuation degree, and the plan execution deviation degree. The interactive schematic diagram of the output of the initial version of the scheduling result by the large model for analysis is as Figure 2 shown.

[0105] Step S3 specifically includes the following process:

[0106] Step S31: The scheduling large model receives and analyzes the feedback opinions of each department in the balance meeting, including: maintenance time adjustment requirements, requirements for non-changeable plans for newly added major projects, suggestions for adjusting workload distribution, etc.

[0107] This process is as follows Figure 3 shown. In this embodiment, the content of the simulated balance meeting is taken as an example:

[0108] In this balance meeting, there are 4 points that need attention and need to be modified:

[0109] 1. The construction department stated that due to work requirements, Plan 3 can only start maintenance on the 25th

[0110] 2. The planning department stated that Plan 2 cannot be maintained on the 19th and the original 14th due to major power supply guarantee events

[0111] 3. The marketing department stated that Plan 15 can only be maintained on the 12th due to work requirements

[0112] 4. The dispatching department stated that due to the large fluctuations in the recent workload, the workload needs to be made more uniform and the variance smaller.

[0113] Step S32: The large model converts various feedback opinions into standardized adjustment instructions, updates the small model parameters and environment codes, and ensures the security and reliability of the output content through the self-check module. The self-check module is for the large model to compare and arrange the small model parameters and arrangement environment codes before and after modification. If it is reasonable, it continues to the next step; otherwise, it retries. If it fails after multiple retries, it returns a failure prompt and stops.

[0114] Step S33: The large model outputs the modified small model code to the arranged small model.

[0115] Step S4 specifically includes the following execution process:

[0116] The arranged small model receives the adjustment information output by the large model, including the modified small model code. The small model updates the deep reinforcement learning environment and parameters according to the new code settings, and uses the updated deep reinforcement learning algorithm to optimize the initial maintenance plan in multiple rounds. Finally, it returns the optimized arrangement result to the maintenance large model, including: the adjusted maintenance plan time arrangement.

[0117] Through the above process, the dynamic optimization arrangement of the maintenance plan by the small model is realized to ensure that the new requirements of each department are met.

[0118] Step S5 specifically includes the following execution process:

[0119] Step S51: As Figure 4 shown, the maintenance large model receives the optimized result returned by the arranged small model, including: the adjusted maintenance plan time arrangement.

[0120] Step S52: As Figure 5As shown in the figure, the large model comprehensively analyzes and evaluates the scheduling results, including examining the uniformity and reasonableness of the workload fluctuation of the inspection and maintenance plan, and evaluating the satisfaction of various constraint conditions.

[0121] The comparison chart of the initial maintenance plan and the maintenance results before and after the adjustment of the large model is as Figure 6 shown. From the comparative analysis, it can be seen that compared with the unarranged original results, the initial arrangement results meet the initial arrangement regulations that the initial plan 3 must be maintained on the starting date (the 3rd), plan 2 cannot be maintained on the 14th, and plan 7 must be changed to be maintained on the 29th. At the same time, it meets the constraints of synchronous maintenance of plan sets 4 and 5 and the relevant constraints of equipment mutual exclusion between plan sets 10, 11 and plan sets 14, 15.

[0122] After the balance meeting, four relevant requirements were added, namely:

[0123] 1. The construction department stated that due to work needs, plan 3 was changed to start maintenance on the 25th.

[0124] 2. The planning department stated that due to major power supply guarantee events, plan 2 cannot be maintained on the 19th and the original 14th.

[0125] 3. The marketing department stated that due to work needs, plan 15 can only be maintained on the 12th.

[0126] 4. The dispatching department stated that due to the large workload fluctuation recently, the workload should be made more uniform and the variance smaller.

[0127] By comparing the final scheduling results with the initial scheduling results, it can be known that plan 3 was changed to start maintenance on the 25th, plan 2 avoided maintenance on the 14th and 19th, and plan 15 was selected to start maintenance on the 12th. At the same time, it also meets various constraints and the initial relevant requirements, and makes the workload fluctuation smaller and the scheduling more reasonable.

[0128] Through the above process, the cooperation between the large language model and the deep reinforcement learning model is realized, the intelligent scheduling, dynamic adjustment and multi-objective optimization of the maintenance plan are realized, which can significantly improve the efficiency, scalability and economy of the maintenance plan scheduling, and provide a new technical solution for the maintenance plan management of the power system.

[0129] Embodiment 3

[0130] This Embodiment 3 provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the intelligent scheduling method of the power system maintenance plan based on the cooperation of large and small models as described above is realized. The method includes:

[0131] The orchestration large model reads the initial maintenance plans of each equipment unit, determines the plan relationships, and inputs the converted orchestration information into the orchestration small model;

[0132] The orchestration small model receives the duration and original date orchestration information of the reported plans, and uses the deep reinforcement learning algorithm to preliminarily orchestrate the maintenance plans, forming a preliminary version of the maintenance plan and returning it to the maintenance large model;

[0133] The orchestration large model receives and analyzes the opinions or new requirements of each department in the balance meeting on the preliminary version of the maintenance plan, converts them into adjustment codes, and uses the adjustment codes to perform collaborative adjustment on the orchestration small model and conduct self-checking;

[0134] The orchestration small model receives the adjustment information from the large model, uses the adjusted deep reinforcement learning algorithm to perform secondary or multiple orchestrations on the preliminary version of the maintenance plan, and returns the orchestration results to the maintenance large model;

[0135] The maintenance large model receives the orchestration results from the orchestration small model, analyzes the results, and provides auxiliary decision-making for the dispatching department.

[0136] Embodiment 4

[0137] Embodiment 4 of the present invention provides a computer device, including a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the intelligent orchestration method for power system maintenance plans based on the collaboration of large and small models as described above. The method includes:

[0138] The orchestration large model reads the initial maintenance plans of each equipment unit, determines the plan relationships, and inputs the converted orchestration information into the orchestration small model;

[0139] The orchestration small model receives the duration and original date orchestration information of the reported plans, and uses the deep reinforcement learning algorithm to preliminarily orchestrate the maintenance plans, forming a preliminary version of the maintenance plan and returning it to the maintenance large model;

[0140] The orchestration large model receives and analyzes the opinions or new requirements of each department in the balance meeting on the preliminary version of the maintenance plan, converts them into adjustment codes, and uses the adjustment codes to perform collaborative adjustment on the orchestration small model and conduct self-checking;

[0141] The orchestration small model receives the adjustment information from the large model, uses the adjusted deep reinforcement learning algorithm to perform secondary or multiple orchestrations on the preliminary version of the maintenance plan, and returns the orchestration results to the maintenance large model;

[0142] The maintenance large model receives the orchestration results from the orchestration small model, analyzes the results, and provides auxiliary decision-making for the dispatching department.

[0143] Embodiment 5

[0144] Embodiment 5 of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory, so that the electronic device executes instructions for implementing the intelligent scheduling method of the power system maintenance plan based on the cooperation of large and small models as described above. The method includes:

[0145] The large scheduling model reads the initial maintenance plans of each equipment unit, judges the plan relationships, and inputs the converted scheduling information into the small scheduling model;

[0146] The small scheduling model receives the construction period and the original date scheduling information of the reported plan, and uses the deep reinforcement learning algorithm to preliminarily schedule the maintenance plan, forms a preliminary version of the maintenance plan and returns it to the large maintenance model;

[0147] The large scheduling model receives and analyzes the opinions or new requirements of each department in the balance meeting on the preliminary version of the maintenance plan, converts them into adjustment codes, uses the adjustment codes to perform collaborative adjustment on the small scheduling model and performs self-checking;

[0148] The small scheduling model receives the adjustment information from the large model, uses the adjusted deep reinforcement learning algorithm to perform secondary or multiple scheduling on the preliminary version of the maintenance plan, and returns the scheduling result to the large maintenance model;

[0149] The large maintenance model receives the scheduling result from the small scheduling model, analyzes the scheduling result, and provides auxiliary decision-making for the dispatching department.

[0150] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0151] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing in the process Figure 1 one process or multiple processes and / or blocksFigure 1 means for the functions specified in one or more boxes.

[0152] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction means that implements the functions specified in one Figure 1 process or processes and / or boxes Figure 1 or more boxes.

[0153] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to perform a series of operational steps on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one Figure 1 process or processes and / or boxes Figure 1 or more boxes.

[0154] Although the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative efforts should be covered within the scope of protection of the present invention.

Claims

1. An intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models, characterized in that, It includes: The scheduling large model reads the initial maintenance plans of each equipment unit, judges the plan relationships, and inputs the converted scheduling information into the scheduling small model; The scheduling small model receives the construction period and the original date scheduling information of the reported plans, and uses the deep reinforcement learning algorithm to preliminarily schedule the maintenance plans, forms the initial version of the maintenance plan and returns it to the maintenance large model; The scheduling large model receives and analyzes the opinions or new requirements of each department in the balance meeting on the initial version of the maintenance plan, converts them into adjustment codes, uses the adjustment codes to perform collaborative adjustment on the scheduling small model and conducts self-checking; The scheduling small model receives the adjustment information of the large model, uses the adjusted deep reinforcement learning algorithm to schedule the initial version of the maintenance plan twice or multiple times, and returns the scheduling result to the maintenance large model; The maintenance large model receives the scheduling result of the scheduling small model, analyzes the scheduling result, and provides auxiliary decision-making for the dispatching department.

2. The intelligent scheduling method for the maintenance plan of the power system based on the cooperation of large and small models according to claim 1, characterized in that, In the deep reinforcement learning algorithm, the objective function is determined corresponding to the workload fluctuation degree and the plan execution deviation degree respectively; It includes: The objective function is constructed by the formula minF = αf1 + βf2; where α and β are weight coefficients, f1 is the workload fluctuation degree, and f2 is the plan execution deviation degree; Pass-through Construct the workload fluctuation degree; where w ij is the workload of the j-th plan on the i-th day, is the average workload, and T is the scheduling period; Pass-through Construct the execution deviation degree of the plan; where s j is the start time after the j-th plan arrangement, s′ j is the start time of the initially reported plan, and λ i is the deviation coefficient.

3. The intelligent scheduling method for power system maintenance plan based on the collaboration of large and small models according to claim 2, characterized in that, Defining the constraint conditions specifically includes: Pass-through Define the daily work upper limit constraint; where ω max is the daily maximum work upper limit value, and T is the scheduling period; Pass-through Define the mutual exclusion constraints of devices; where μ is the set of mutual exclusion plans, s m 、s n are the start times of the overhauls of plans m and n respectively, and d m 、d n are the overhaul durations of plans m and n respectively; Pass-through Define the synchronous maintenance constraint; where φ is the set of synchronous maintenance plans, and s p , s q are the start times of the p-th and q-th plans respectively; Pass-through Indicates the constraint of the non-maintainable period of major projects; where s k is the start time after the schedule arrangement of plan k, d k is the construction period, {a k} is the set of non-maintainable dates; Pass-through Define the time-fixed constraint; where G is the set of plans that need to fix the time, and s r is the start time after scheduling plan r, and s' r is the start time of the initial reported plan.

4. The intelligent scheduling method for power system maintenance plan based on the collaboration of large and small models according to claim 3, characterized in that Pass - through type S = {N, j, d j , s' j , {w ij}, φ, μ} to obtain the state space; where N is the total number of plans, j is the current plan number, d j is the current planned duration, s' j is the start time of the initial reported plan, {w ij} is the daily workload array, φ is the set of synchronous maintenance plans, and μ is the set of mutually exclusive plan pairs; The action space is obtained by the formula A = {s|s ∈ {1, 2,..., T}}; where T is the scheduling period and s is the maintenance start time; Pass-through Obtain the reward function; where λ1 and λ1 are penalty weight coefficients, and ε k is the weight coefficient of the k-th type of constraint penalty term, and P k is the k-th type of constraint violation penalty. The penalty terms respectively include the daily work upper limit constraint penalty, the equipment mutual exclusion constraint penalty, the synchronous maintenance constraint penalty, the non-maintenance period constraint penalty for major projects, and the time-fixed constraint penalty.

5. The intelligent scheduling method for the maintenance plan of the power system based on the collaboration of large and small models according to claim 1, characterized in that After the scheduling large model reads the initial maintenance plan information, through natural language processing capabilities and topological information, it judges the relationships of the maintenance plans, and identifies the maintenance plan combinations that need to be merged and the mutually exclusive maintenance plan combinations; Subsequently, the judgment result is converted into standardized scheduling information.

6. The intelligent scheduling method for power system maintenance plan based on the collaboration of large and small models according to claim 5, characterized in that, The initial maintenance plan information includes: maintenance plan number, total number of plans, equipment type, equipment index number, affiliated substation, original reported start maintenance time, maintenance construction period, and remarks.

7. An intelligent scheduling system for power system maintenance plans based on the collaboration of large and small models, characterized in that, It includes: A reading module, used for the scheduling large model to read the initial maintenance plans of each equipment unit, judge the plan relationships, and input the converted scheduling information into the scheduling small model; A preliminary scheduling module, used for the scheduling small model to receive the construction period and the original date scheduling information of the reported plans, and use the deep reinforcement learning algorithm to preliminarily schedule the maintenance plans, form the initial version of the maintenance plan and return it to the maintenance large model; A conversion module, used for the scheduling large model to receive and analyze the opinions or new requirements of each department in the balance meeting on the initial version of the maintenance plan, convert them into adjustment codes, and use the adjustment codes to perform collaborative adjustment on the scheduling small model and conduct self-checking; A second scheduling module, the scheduling small model receives the adjustment information of the large model, uses the adjusted deep reinforcement learning algorithm to schedule the initial version of the maintenance plan twice or multiple times, and returns the scheduling result to the maintenance large model; An analysis module, used for the maintenance large model to receive the scheduling result of the scheduling small model, analyze the scheduling result, and provide auxiliary decision-making for the dispatching department.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the intelligent scheduling method for power system maintenance plans based on the collaboration of large and small models as described in any one of claims 1-6 is implemented.

9. A computer device, characterized in that, It includes a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the intelligent scheduling method for the power system maintenance plan based on the cooperation of large and small models according to any one of claims 1-6.

10. An electronic device, characterized in that, It includes: A processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory so that the electronic device executes the instructions for implementing the intelligent scheduling method for the power system maintenance plan based on the cooperation of large and small models according to any one of claims 1-6.

Images