An AI algorithm-based power grid maintenance plan intelligent arrangement method
By constructing a knowledge graph and using a hybrid optimization algorithm, the problems of low efficiency in power grid maintenance plan scheduling and insufficient dynamic response capability were solved, and efficient and accurate automatic generation of maintenance plans was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TONGCHUAN POWER SUPPLY CO OF STATE GRID SHAANXI ELECTRIC POWER CO LTD
- Filing Date
- 2026-06-10
- Publication Date
- 2026-07-24
AI Technical Summary
The existing power grid maintenance planning relies on manual methods, which are inefficient, prone to errors, and difficult to meet the complex needs of large-scale power grids. Furthermore, existing intelligent methods are insufficient in terms of multi-objective optimization and dynamic response capabilities.
An AI-based intelligent scheduling method for power grid maintenance plans is adopted. Maintenance elements are extracted using a large language model, a knowledge graph is constructed and encapsulated as a legality judgment function, and multi-objective optimization is performed by combining an improved adaptive ant colony algorithm and an improved bat algorithm. Constraints are verified in real time to achieve global search and local fine-tuning.
It improves the efficiency and quality of maintenance plan preparation, enhances dynamic response capabilities, ensures that the preparation results meet all constraints, reduces human error, and achieves automatic generation and quality assurance of maintenance plans.
Smart Images
Figure CN122452944A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power system automation and artificial intelligence technology, specifically relating to an intelligent scheduling method for power grid maintenance plans based on AI algorithms. Background Technology
[0002] With the advancement of the construction of new power systems, the power grid is continuously expanding in scale and becoming increasingly complex in structure, resulting in a significant increase in the number of equipment such as transmission cables, transformers, circuit breakers, and switchgear. To ensure the stability of power grid operation and the reliability of power supply, it is necessary to routinely carry out preventive maintenance, fault elimination, and equipment upgrades. Maintenance tasks are becoming more batch-oriented and complex, placing higher demands on maintenance planning.
[0003] Currently, power grid maintenance planning is primarily done manually. Dispatchers use Excel spreadsheets or basic dispatching systems, combined with their own experience, to manually compile a list of equipment to be maintained. They then comprehensively consider multiple constraints, including grid operation modes, equipment outage windows, and personnel allocation, to match maintenance tasks, work teams, and work periods. This method has significant shortcomings: First, it heavily relies on the personal experience and proficiency of dispatchers, with planning taking hours or even days, resulting in low efficiency and a high risk of errors. As the power grid continues to expand, manual planning is no longer sufficient to meet development needs. Second, each planning session is essentially starting from scratch, making it difficult to effectively absorb and reuse historical maintenance experience and expert knowledge, and the quality of planning relies excessively on personal experience.
[0004] Furthermore, while some existing intelligent orchestration methods incorporate optimization algorithms or rule-based reasoning, they still have limitations. Single optimization algorithms are prone to getting trapped in local optima when faced with complex optimization problems involving high dimensions and multiple peaks, making it difficult to strike a balance between the breadth of the global search and the precision of the local search. Simultaneously, during maintenance, plans often need to be adjusted based on on-site conditions, such as schedule fluctuations, sudden weather events, or changes to important meetings. Existing methods have weak responsiveness to such dynamic changes, making it difficult to achieve flexible and rapid adjustments.
[0005] Therefore, how to improve scheduling efficiency, optimize multi-objective balance, and enhance dynamic response capability while ensuring the compliance and security of maintenance plans is a technical problem that urgently needs to be solved in the field of power grid maintenance plan scheduling. Summary of the Invention
[0006] To address the aforementioned problems, this invention provides an intelligent scheduling method for power grid maintenance plans based on AI algorithms.
[0007] To achieve the above objectives, this invention provides an intelligent scheduling method for power grid maintenance plans based on AI algorithms, comprising: Obtain maintenance tasks reported by maintenance requesting units; obtain power grid equipment models, equipment status data, maintenance work orders, dispatching procedures and historical maintenance plan data of the power grid to be maintained; construct a knowledge graph about equipment and maintenance; and encapsulate the knowledge graph into a legality judgment function.
[0008] The equipment identifier, time information, resource requirements, and priority markers of the maintenance task are extracted using a pre-trained large language model as maintenance elements. Based on the maintenance elements, the constraint relationships between the equipment in the maintenance task are translated into structured constraint data for each maintenance task. Using the structured constraint data and the preset scheduling cycle days, constraint conditions including equipment mutual exclusion constraints, equipment synchronization constraints, daily work limit constraints, and prohibited time period constraints are constructed, along with a multi-objective optimization function that minimizes the workload balance of the power grid to be maintained, minimizes the power outage loss cost, and minimizes the plan execution deviation.
[0009] An improved adaptive ant colony algorithm is used to solve a multi-objective optimization function, which enhances global search in the early stage, accelerates convergence in the middle stage, and avoids local optima in the later stage by improving the pheromone volatility coefficient. Each ant selects a start time for each maintenance task to construct a solution path, and a validity determination function is called to prune the solution path in real time. When a preset switching condition is met, the optimal solution output by the ant colony algorithm is used as the initial population, and an improved bat algorithm, which correlates the search frequency of bats with the pheromone concentration of the ant colony, is used to locally optimize the optimal solution. After each position update by the bat algorithm, a validity determination function is called to constrain and verify the updated position, thus obtaining the final solution.
[0010] The validity determination function is called to perform a full constraint check on the final solution and output the optimal maintenance scheduling execution plan that satisfies all constraints.
[0011] Preferably, the knowledge graph regarding equipment and maintenance includes an equipment topology subgraph, a maintenance constraint subgraph, and a maintenance resource carrying capacity subgraph. The equipment topology subgraph uses transformers, circuit breakers, lines, switchgear, busbars, and substations in the power grid equipment model as nodes, and electrical connections as edges. The maintenance constraint subgraph is based on scheduling procedures, maintenance work orders, and the constraints between equipment in the maintenance tasks, using the equipment to be maintained as nodes, and mutual exclusion constraints, synchronization constraints, topology constraints, power supply constraints, and time window constraints as edges. The maintenance resource carrying capacity subgraph is based on scheduling procedures and historical maintenance plan data, using work teams as nodes, and constraints on the number of work teams, number of personnel, average daily maximum workload, tools, vehicles, and operating radius as edges.
[0012] Preferably, the bat's search frequency is correlated with the pheromone concentration of the ant colony. When the average pheromone concentration of the ant colony is higher than a threshold, the bat's search frequency increases to conduct a fine search near the high-quality solution; when the average pheromone concentration of the ant colony is lower than the threshold, the bat's search frequency decreases to expand the search range.
[0013] Preferably, in the multi-objective optimization function, the workload balance is the minimization of the deviation between the daily workload and the average daily workload within the scheduling cycle, the power outage loss cost is the minimization of the power outage loss corresponding to the start date of each maintenance task, and the plan execution deviation is the minimization of the deviation between the actual start time and the initial reporting time of each maintenance task; the weights of each objective are dynamically adjusted by the large language model according to the business priority.
[0014] Preferably, the full constraint verification also includes an automatic repair strategy, specifically including: when repairing mutual exclusion conflicts, the device that starts later will be postponed until the device that starts earlier finishes; when repairing synchronization conflicts, the start time will be unified to a later time; when repairing resource overload, the low-priority device on the overload day will be shifted backward and the legality judgment function will be called again for verification; when repairing the prohibited repair period, the entire maintenance window will be shifted out of the prohibited repair interval.
[0015] Preferably, the method further includes inputting the optimal maintenance scheduling execution plan into a large language model to generate a natural language explanation that includes information on changes in workload balance, changes in power outage loss costs, conflict adjustment status, and power supply protection period avoidance.
[0016] Preferably, the translation rules for natural language constraints into structured constraint data include: two devices needing to be repaired simultaneously are translated into synchronization constraints; two devices not being repaired simultaneously are translated into mutual exclusion constraints; one device must complete its repair before another device is translated into timing constraints; devices not being allowed to be repaired during a specified period are translated into prohibited time period constraints; and a maximum of a preset number of devices being repaired simultaneously per day are translated into resource capacity constraints.
[0017] This invention also provides an intelligent scheduling system for power grid maintenance plans based on AI algorithms, comprising: The data acquisition module is used to acquire maintenance tasks reported by maintenance requesting units; acquire power grid equipment models, equipment status data, maintenance work orders, dispatching procedures and historical maintenance plan data of the power grid to be maintained; construct a knowledge graph about equipment and maintenance; and encapsulate the knowledge graph into a legality judgment function.
[0018] The function construction module is used to extract equipment identifiers, time information, resource requirements, and priority markers of the maintenance task as maintenance elements using a pre-trained large language model; based on the maintenance elements, the constraint relationships between the equipment in the maintenance task are translated into structured constraint data for each maintenance task; using the structured constraint data and the preset scheduling cycle days, constraints including equipment mutual exclusion constraints, equipment synchronization constraints, daily work limit constraints, and prohibited time period constraints are constructed, along with a multi-objective optimization function that minimizes the workload balance of the power grid to be maintained, minimizes the power outage loss cost, and minimizes the plan execution deviation.
[0019] The orchestration module employs an improved adaptive ant colony algorithm, which enhances global search in the early stages, accelerates convergence in the mid-stage, and avoids local optima in the later stages, using the pheromone evaporation coefficient to solve a multi-objective optimization function. Each ant selects a start time for each maintenance task to construct a solution path and calls a validity determination function to prune the solution path in real time. When preset switching conditions are met, the optimal solution output by the ant colony algorithm is used as the initial population, and an improved bat algorithm, which correlates the search frequency of bats with the pheromone concentration of the ant colony, is used to locally optimize the optimal solution. After each position update by the bat algorithm, a validity determination function is called to perform constraint verification on the updated position to obtain the final solution. Finally, a full constraint verification is performed on the final solution, and the optimal maintenance orchestration execution plan that satisfies all constraints is output.
[0020] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in the AI-based intelligent scheduling method for power grid maintenance plans.
[0021] The present invention also provides a computer-readable storage medium storing a computer program, which, when loaded by a processor, is capable of executing any of the steps in the AI-based intelligent scheduling method for power grid maintenance plans.
[0022] The intelligent scheduling method for power grid maintenance plans based on AI algorithms provided by this invention has the following beneficial effects: By constructing a knowledge graph and encapsulating it into a legality judgment function, the topology of power grid equipment, maintenance procedures, and historical experience are precipitated into a reusable rule engine, solving the problems of difficult knowledge accumulation and error-prone manual scheduling; by using a large language model to extract maintenance elements from natural language application texts and translating them into structured constraints, the problems of difficult natural language requirement understanding and dynamic response are solved; a weighted optimization function is constructed with workload balancing, power outage losses, and plan deviations as objectives, and an improved adaptive ant colony algorithm with periodic changes in pheromone evaporation coefficient is used for global search, and an improved bat algorithm with search frequency correlated with pheromone concentration is used for local fine-tuning after the switching conditions are met. The two algorithms complement each other, solving the problems of single algorithms being prone to getting trapped in local optima and difficulty in balancing multiple objectives; at the same time, the legality judgment function runs through the entire process of ant colony pruning, bat verification, and final full verification, ensuring that the output plan meets all constraints, greatly improving scheduling efficiency and compliance, and realizing automatic generation and quality assurance of maintenance plans. Attached Figure Description
[0023] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 This is a flowchart illustrating an intelligent scheduling method for power grid maintenance plans based on AI algorithms, according to an embodiment of the present invention. Figure 2 This is a system architecture diagram of an embodiment of the present invention; Figure 3 This is a flowchart illustrating the execution of the hybrid algorithm according to an embodiment of the present invention. Figure 4 This is a flowchart of constraint verification and automatic repair according to an embodiment of the present invention. Detailed Implementation
[0025] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.
[0026] like Figure 2As shown, this invention includes the following core modules: Large Language Model Layer: Completes natural language requirement understanding, constraint translation, dynamic adjustment of multi-objective weights, and result generation. Maintenance Knowledge Graph Layer: Completes modeling and reasoning of equipment topology, maintenance constraints, conflict rules, and resource carrying capacity. Hybrid Optimization Algorithm Layer: Utilizes the Improved Adaptive Ant Colony Algorithm (I-ACO) for global maintenance orchestration and the Improved Bat Algorithm (I-BA) for local fine-tuning of the solution. Constraint Verification and Iteration Layer: Utilizes the graph for real-time verification, completing conflict repair and multi-round rearrangement.
[0027] This invention provides an intelligent scheduling method for power grid maintenance plans based on AI algorithms, specifically an intelligent scheduling method integrating a Large Language Model (LLM), a Knowledge Graph (KG), and an Improved Ant Colony-Bat Hybrid Optimization Algorithm (I-ACO-BA). The method constructs a power grid maintenance knowledge graph as a dynamic constraint engine, utilizes the strong global search capability of the Improved Ant Colony Algorithm (I-ACO) for primary optimization, and introduces the Improved Bat Algorithm (I-BA) for local fine-tuning of high-quality solutions, achieving complementary advantages. During algorithm iteration, the knowledge graph is called in real time for dynamic determination of topology and rule conflicts, pruning illegal solutions. Simultaneously, the Large Language Model is used to translate natural language requirements into structured constraint code, and the algorithm switching logic, multi-objective weights, and search parameters are dynamically adjusted based on optimization feedback. This achieves intelligent scheduling throughout the entire process of semantic understanding, constraint translation, optimization control, and result interpretation, effectively solving problems such as insufficient optimization capabilities, lagging constraint verification, and weak human-computer interaction in existing technologies. Figure 1 As shown, it includes the following steps:
[0028] S1. Constructing a knowledge graph for power grid maintenance Acquire data such as power grid equipment models, graphics, equipment status, measurement data, maintenance work orders, dispatching procedures, and historical maintenance plans to construct the following three types of knowledge graphs: S101, Equipment Topology Diagram A topology map of equipment is constructed using transformers, circuit breakers, lines, switchgear, busbars, and substations as nodes, and electrical connections, hierarchical relationships, and power supply relationships as edges.
[0029] S102, Maintenance Constraint Diagram A maintenance constraint graph is constructed using constraints such as mutual exclusion (equipment i and j cannot be de-energized at the same time), synchronization (equipment i and j must be inspected at the same time), topology (upstream and downstream equipment linkage constraints), power supply constraint (no inspection is allowed during major events / power supply periods), and time window constraint (earliest start time and latest end time).
[0030] S103, Maintenance Resource Carrying Capacity Map A maintenance resource carrying capacity map is constructed based on constraints such as the number of work teams, the number of personnel, the maximum daily workload, tools, vehicles, and work permissions, as well as geographical distribution and work radius constraints.
[0031] After the above graph is constructed, the graph query interface is encapsulated as a "legitimacy determination function isLegal(Plan)" for subsequent optimization algorithms to call.
[0032] S2, Large Language Model Analysis The system utilizes a large language model to read unstructured data such as initial maintenance work orders and maintenance plan texts. Through the natural language understanding capabilities of the large language model, it identifies the core elements of maintenance tasks, including equipment identification, time information, resource requirements, and priority markers. Equipment identification includes equipment name, number, affiliated plant, and voltage level; time information includes planned start time, duration, and end time; resource requirements include required work teams, tools, and special vehicles; and priority markers include emergency maintenance, planned maintenance, and troubleshooting tasks. By constructing a prompt template containing system role definitions, input examples, and output format constraints, the system guides the large language model to extract structured constraints from the unstructured text. This translates the natural language descriptions of inter-equipment constraints into a structured constraint description language, outputting standardized constraint data containing equipment identification, time information, resource requirements, priority markers, and constraint conditions.
[0033] The LLM outputs structured constraints, which include: the plan number field is "PLAN20260601001", the equipment name is "No. 1 main transformer", the substation is "XX substation", the voltage level is "110kV", the expected start date is "2026-06-15", the duration is 3 days, the constraint condition field includes the synchronous maintenance equipment list as "110kV bus", the mutually exclusive maintenance equipment list as "No. 2 main transformer", and the Boolean flag for whether it is a fixed time plan is "false".
[0034] LLM translates natural language constraints into a structured constraint description language (CDL). The translation rules are shown in Table 1. Table 1 Natural Language Constraint Translation Rules LLM translates constraints using the following prompt template. The system prompts you that you are a power grid maintenance plan constraint translation assistant. Please convert the user-input natural language maintenance requirements into a standardized CDL format. The user inputs [requirement text], and the output format is JSON, containing the fields plan_id, device, and constraints.
[0035] S3. Constructing a multi-objective optimization mathematical model Assume the total number of maintenance plans is The arrangement cycle is Every day, every plan The construction period is The initial reporting start time is .
[0036] S301, Decision Variables Define decision variables to represent the plan Is it in the first The day begins: ; in, S302, Objective Function The multi-objective weighted function is constructed as follows: in, This refers to workload volatility, used to measure the balance of resource allocation. In the formula That is, the first Total workload for the day That is, the average daily workload. For the plan Daily resource consumption; The cost of power outage losses is used to measure power supply reliability. In the formula For the plan exist The cost of power outage losses during the start of construction was obtained by fitting historical load data with voltage level coefficients. The plan execution deviation is used to measure the deviation from the originally reported requirements. Weighting coefficient satisfy The LLM dynamically adjusts the parameters based on business priorities.
[0037] S303, Constraints S3031, Uniqueness Constraint: Each plan can only select one start time. ; S3032, Schedule Integrity Constraints: Maintenance must not exceed the scheduled cycle. ;in, , indicating a plan The start time of the maintenance.
[0038] S3033, Equipment mutual exclusion maintenance constraint: Maintenance times for mutually exclusive devices must not overlap. For any pair of mutually exclusive devices... ( (For a set of mutually exclusive pairs), their maintenance times cannot overlap: ; S3034, Constraints on Synchronous Equipment Maintenance: Synchronous devices need to start maintenance simultaneously; for any pair of synchronous devices... ( To synchronize the sets, they must start simultaneously: ; S3035, Daily Work Limit Constraint: ;in, This represents the maximum allowable workload per day.
[0039] S3036, Fixed Time Constraint: For plans marked as fixed time Its start time cannot be changed: ; S3037, Restrictions on Prohibited Periods for Major Projects: If on a certain day During periods of power supply guarantee or maintenance restrictions for major projects, relevant equipment is assembled. Unrepairable; ; in, This refers to the collection of equipment that is prohibited from maintenance during designated maintenance periods. Gathering for power supply guarantees or during periods of restricted maintenance for major projects.
[0040] S4. Improved Adaptive Ant Colony Algorithm (IACO) Global Optimization like Figure 3 As shown, the algorithm parameters are initialized first, including ant colony parameters and bat parameters. An improved adaptive ant colony algorithm is used for global search. Each ant generates a maintenance plan and calculates the objective function value. Simultaneously, a validity determination function is called to determine the validity of the plan. If valid, the global optimal solution is updated and the pheromone is updated adaptively in a loop; if invalid, the path is blocked and the pheromone is penalized. This process is repeated until the maximum number of iterations is reached or the switching condition is met, and a candidate optimal solution is output. Then, the improved bat algorithm enters a local fine-tuning stage. The better solutions output by the ant colony are used as the initial population. The bat speed and position are updated, and the relationship between random numbers and impulse rate determines whether to perform a local search. After each update, a validity determination function is called to perform constraint verification and update the loudness and impulse rate. Finally, the candidate optimal solution after global search and local fine-tuning is output, and the process ends.
[0041] S401, Problem Modeling This invention abstracts the intelligent scheduling problem of power grid maintenance plans into a constrained graph coloring problem. In this model, each device to be maintained is treated as an independent node; mutually exclusive maintenance relationships between devices (i.e., pairs of devices that cannot be maintained simultaneously) are modeled as conflicting edges connecting two nodes; and the available maintenance time slices (e.g., each day within the scheduling cycle) correspond to the available colors in the graph. Simultaneously, the maximum allowed maintenance workload per day is converted into a coloring quantity constraint, meaning the total number of nodes that can be colored in the same time slice (color) must not exceed the resource carrying capacity limit. Through this graph coloring modeling approach, the maintenance scheduling problem is transformed into assigning a time slice color to each node while satisfying all conflicting edges and coloring quantity constraints, thus providing a clear mathematical foundation for the subsequent improvement of the Adaptive Ant Colony (IACO) algorithm's path search and constraint pruning.
[0042] S402, State Transition Probability Ant For equipment Select time slice The probability of: Heuristic function as follows: in, For pheromone concentration, As a pheromone-inspired factor, As the expected heuristic factor, For ants The set of currently allowed legal time slices (determined by the knowledge graph isLegal interface). For hypothetical plans In the The increments of each sub-target at the beginning of the day Avoid dividing by zero for extremely small positive numbers.
[0043] S403, Adaptive Cyclic Pheromone Update in, () represents the pheromone evaporation coefficient. For pheromone intensity, This represents the objective function value of the current iterative optimal solution. Volatility coefficient. It changes cyclically with the iteration cycle: ,in, The iteration cycle is as follows: the volatility coefficient is small in the early stage to enhance global search; the volatility coefficient is large in the middle stage to accelerate convergence; and the volatility coefficient is small in the later stage to avoid local optima.
[0044] S5, Improved Bat Algorithm (IBA) Local Fine-tuning S501, Population Initialization bat The position is represented as the start time vector of the maintenance plan: Initialization refers to using the Top-K best solutions (i.e., the bat population size) output by I-ACO as the initial population for the I-BA algorithm: ; in, This represents the population size of bats.
[0045] S502, Bat Speed Update ; in, This is the current globally optimal solution. The bat search frequency is coupled with the I-ACO pheromone concentration.
[0046] S503, Location Update: After each location update, the new location is rounded and clipped to ensure it falls within a valid date range. .
[0047] S504, Pheromone-Frequency Coupling Mechanism: in, The maximum pulse frequency, To minimize the pulse frequency, The activation function maps the difference to the 0-1 interval. This represents the average concentration of I-ACO pheromones in the current iteration. This is the pheromone threshold. The higher the value, the higher the search frequency, indicating a more refined search within that area. When the frequency increases, a more refined search is performed near a high-quality solution; conversely, the search range is expanded.
[0048] S505, Local Search (Random Walk) When random number At that time, the bat entered a local search mode: ; in, For random directions, This represents the current average loudness.
[0049] S506, Parameter Update After each iteration, the loudness and impulse emission rate are updated: ; in, The initial loudness, This is the loudness attenuation coefficient.
[0050] ; in, The initial pulse emission rate, Pulse enhancement coefficient As the iterations proceed, the loudness gradually decreases and the impulse rate gradually increases, and the algorithm gradually transitions from global search to local fine-tuning.
[0051] S6, IACO and IBA switching conditions In this invention, the triggering conditions for switching from the Improved Adaptive Ant Colony Algorithm (IACO) to the Improved Bat Algorithm (IBA) are set as follows: When any one of the following conditions is met, the hybrid optimization algorithm switches from IACO in the global search phase to IBA in the local fine-tuning phase: First, the current iteration number of IACO reaches the preset maximum iteration number. Second, the objective function value of the global optimum did not improve in any of the 20 consecutive iterations of IACO; third, the variance of fitness among the top K individuals in the IACO population was less than a very small positive threshold. This condition indicates a significant loss of population diversity, making it difficult for the algorithm to obtain a better solution through continued global search. The aforementioned switching mechanism ensures that the hybrid algorithm can promptly initiate local fine-tuning when its global search capability becomes saturated or convergence stalls, thereby improving overall optimization efficiency and solution quality.
[0052] S7, Constraint Validation and Automatic Repair After optimization, the knowledge graph's `isLegal` interface is called to perform a full constraint check, and solutions that violate constraints are automatically repaired. Mutual exclusion conflict repair extends the time from the last device to the last device to finish: Synchronization conflict resolution includes unifying to a later start time: Resource overload repair refers to shifting low-priority devices from the overload day to the next day and calling the isLegal function again to check if other constraints are violated after the shift. Repair during restricted maintenance periods refers to shifting the entire maintenance window out of the restricted maintenance zone.
[0053] S8. Multi-objective optimization and output like Figure 4As shown, after inputting the candidate optimal solution, the legality judgment function of the knowledge graph is called to identify the conflict type. If a constraint conflict exists, corresponding repair strategies are executed according to different conflict types, including: for mutual exclusion conflicts, the start time of the later device is postponed until the end time of the previous device; for synchronization conflicts, the start time is unified to the later time; for resource overload, the low-priority device is shifted back by one day; for power supply conflicts, the entire maintenance window is moved out of the prohibited maintenance zone; and for project timeline violations, the start time is adjusted forward to within the cycle. After repair, the maintenance plan is updated and the legality judgment function is called again until there are no constraint conflicts, and finally the final optimal solution is output. The specific steps are as follows:
[0054] S801, Reward Function and Scheme Evaluation The final maintenance plan must meet the following requirements: ; And it must pass all constraints, which are expressed as follows: Set validation. This indicates the search for the objective function. Minimize the independent variable, This represents the set of maintenance plans that the Plan must satisfy all constraints.
[0055] Interpretation of results from S802 and LLM drivers After I-BA optimization, the optimal solution is input into LLM to generate a natural language interpretation: Input the optimal solution vector +Objective function value The output example is: "This optimization has generated a maintenance plan. The overall workload balance is 2.3 (35% lower than the initial plan), and the power outage loss cost is reduced by 12%. The original plan had mutual conflicts between the XX line and the YY line, which have been automatically adjusted to staggered maintenance. It is recommended to pay special attention to the maintenance schedule of the No. 1 main transformer, as this plan avoids the power supply guarantee period for major events from June 6th to June 8th."
[0056] S803, Iteration Schedulers can make adjustment suggestions through natural language, such as "postpone the XX plan by 3 days". After LLM parsing, the requirements are transformed into constraint updates, and the hybrid optimization algorithm solution steps are triggered again (based on the existing solution, the I-BA algorithm is used for local re-optimization to avoid completely recalculating from scratch).
[0057] Example Taking the monthly maintenance plan of a prefecture-level city's power grid in June 2026 as an example, the specific implementation process of this invention is explained in detail. First, basic parameters are set: the total number of equipment to be maintained is N=50, including 3 220kV transformers, 12 110kV lines, 25 35kV circuit breakers, and 10 10kV switchgear; the scheduling period is T=30 days, i.e., from June 1st to June 30th, 2026; there are 5 maintenance teams, with a maximum daily allowable workload of 20 person-days; the initial multi-objective weighting coefficients are set as α=0.4 (workload balance), β=0.3 (power outage loss), and γ=0.3 (plan deviation); the power supply guarantee period is from June 15th to June 17th (power supply guarantee for the middle school entrance examination), during which maintenance of the main grid equipment in the urban area is prohibited.
[0058] Next, a power grid maintenance knowledge graph was constructed. Data was extracted from the power grid's EMS and PMS systems to construct three types of graphs: an equipment topology graph, containing the electrical connection relationships of 50 devices, such as "220kV XX substation No. 1 main transformer, 110kV I section busbar, 110kV XX line"; a maintenance constraint graph, including mutually exclusive groups (such as 220kV XX substation No. 1 and No. 2 main transformers, 110kV XX line and 110kV YY line), synchronization groups (such as 220kV XX substation No. 1 main transformer and 110kV I section busbar, 35kV XX circuit breaker and 35kV XX bay) and power supply constraints (maintenance of 220kV and 110kV equipment in urban areas is prohibited from June 15th to 17th); and a resource carrying capacity graph, recording the personnel configuration, tool list, and working radius of 5 work teams.
[0059] Subsequently, a large language model was used to parse the natural language requirements. Fifty natural language maintenance requests submitted by various units were input, such as "Request for preventative testing of the No. 1 main transformer at 220kV XX substation on June 10th, with a duration of 2 days. It needs to be simultaneously de-energized with the 110kV I-section busbar, and cannot be maintained simultaneously with the No. 2 main transformer; high priority." The large language model output standardized constraints, including plan_id, equipment name, substation, voltage level, expected date, duration, priority, and specific constraints (synchronous equipment, mutually exclusive equipment, prohibited maintenance periods, whether fixed, etc.).
[0060] Based on this, a multi-objective optimization mathematical model is constructed, which includes three objective functions and seven types of constraints according to the aforementioned formula. Based on the business requirement that "power supply security during the high school entrance examination has the highest priority," the weighting coefficients of the large language model are dynamically adjusted to α=0.3, β=0.4, and γ=0.3.
[0061] Then, an improved adaptive ant colony algorithm is executed for global optimization. The number of ants is set to 100, the initial pheromone value τ0 = 1.0, and the pheromone intensity Q = 100, with 200 iterations. In the 120th iteration, the optimal solution shows no improvement for 20 consecutive iterations, triggering a switching condition and entering the local fine-tuning of the improved bat algorithm. At this point, the objective function values of the top-30 better solutions output by the ant colony are in the range [125.6, 138.2].
[0062] Next, the improved bat algorithm was fine-tuned locally. The bat population size was set to 30, and the Top-30 solutions output by the ant colony were directly used as the initial population; the initial loudness A0=0.5, the initial impulse rate r0=0.5, and the iteration was performed for 50 rounds. After 50 rounds of fine-tuning, the objective function value of the candidate optimal solution was 112.3, which is 10.6% higher than the optimal solution of the ant colony.
[0063] Constraint verification and automatic repair were then performed. A full constraint verification was conducted using the knowledge graph's `isLegal` interface, revealing a resource overload conflict: the total workload on June 8th was 22 person-days, exceeding the daily limit of 20 person-days. The system automatically repaired this by postponing the lowest priority task, "10kV XX switchgear maintenance," scheduled for June 8th to June 9th. After the repair, all constraints were satisfied.
[0064] The final optimal maintenance plan is output. Partial arrangements are as follows: The 220kV XX substation's No. 1 main transformer will be maintained from 08:30 on June 10th to 18:30 on June 11th, 2026, a period of 2 days, handled by shift X1, with a workload of 8 person-days; the 110kV I section busbar will be maintained by shift X1 during the same period, with a workload of 4 person-days; the 110kV XX line will be maintained from 08:30 on June 12th to 18:30 on June 13th, 2026, a period of 2 days, handled by shift X2, with a workload of 6 person-days; the 10kV XX switchgear will be maintained from 08:00 to 18:00 on June 9th, 2026, a period of 1 day, handled by shift X3, with a workload of 2 person-days.
[0065] The structured constraints output by this invention include fields for plan number, equipment name, substation, voltage level, expected start date, duration, priority, and constraint conditions. The constraint conditions field further includes a list of equipment to be maintained concurrently, a list of equipment to be maintained mutually exclusively, a list of prohibited maintenance periods, and a Boolean flag indicating whether it is a fixed-time plan. Table 2 shows the final optimal maintenance plan arrangement in the output section.
[0066] Table 2 Example of Optimal Maintenance Plan Output The overall architecture of this invention consists of a four-layer collaborative orchestration architecture, comprising a large language model layer, a maintenance knowledge graph layer, a hybrid optimization algorithm layer, and a constraint verification and iteration layer, used to achieve intelligent orchestration of power grid maintenance plans. Specifically, the large language model is responsible for natural language understanding and constraint translation, the knowledge graph provides dynamic constraint verification and reasoning, the hybrid algorithm completes global optimization and local fine-tuning, and the constraint verification layer implements conflict detection and automatic repair.
[0067] This invention encapsulates the constructed equipment topology map, maintenance constraint map, and resource carrying capacity map into a unified legality judgment function. During the process of improving the ant colony algorithm to select time slices for equipment (construction path) and improving the bat algorithm to fine-tune the maintenance plan position (position update), this function is called in real time to perform legality judgment, realize the immediate pruning of illegal solutions, and ensure that the optimization process is always carried out within the feasible domain.
[0068] This invention employs a pheromone evaporation coefficient that cyclically changes over time. A smaller evaporation coefficient is maintained in the early stages of the iteration cycle to enhance global search capabilities; the evaporation coefficient is increased in the middle stages to accelerate convergence; and the evaporation coefficient is decreased again in the later stages to avoid getting trapped in local optima, thereby achieving a dynamic balance between global exploration and local exploitation.
[0069] This invention establishes a mathematical correlation between the search frequency of bats in the improved bat algorithm and the average pheromone concentration in the current iteration of the improved ant colony algorithm. When the pheromone concentration is high, the search frequency is increased to perform a more refined local search near the high-quality solution; when the pheromone concentration is low, the search frequency is decreased to expand the search range, thus achieving organic linkage and complementary advantages of the hybrid algorithms.
[0070] This invention utilizes a large language model to parse and translate the natural language feedback from dispatchers into structured constraint updates. Without completely recalculating, it triggers an improved bat algorithm to perform local re-optimization based on the existing solution, achieving iterative semantic understanding, constraint updates, local re-optimization, and result output, supporting dynamic adjustment of maintenance plans.
[0071] After optimization, this invention calls a legality determination function to perform full constraint verification and executes automatic repair strategies for different types of non-compliant constraints: mutual exclusion conflict repair (postponing the start time of the later device to the end time of the previous device), synchronization conflict repair (unifying the start time to a later one), resource overload repair (shifting the low-priority devices on overload days to the later ones), and prohibited maintenance period repair (shifting the entire maintenance window out of the prohibited maintenance interval), ensuring that the final maintenance plan meets all constraints.
[0072] Regarding algorithm substitution, this invention provides the following explanations: In the hybrid optimization architecture, in addition to using the improved adaptive ant colony algorithm as the primary optimizer, the particle swarm optimization algorithm can also be used as a substitute; in addition to using the improved bat algorithm as the local fine-tuner, the simulated annealing algorithm or the tabu search algorithm can also be used as substitutes, as long as the hybrid optimization architecture of "global search + local fine-tuning" is followed. Furthermore, a non-dominated sorting genetic algorithm can be used for multi-objective Pareto solutions, or an optimization solver can be used to construct a hybrid integer programming model to achieve accurate solutions. The dynamic constraint verification module can be implemented using a knowledge graph or a vector database combined with a rule engine, but the latter is weaker than a knowledge graph in terms of reasoning ability. For smaller-scale power grid scenarios, a hybrid storage mode combining relational databases and graph databases can be used as an alternative. The application of the large language model can be replaced as follows: the large language model can only be used to generate prompts for optimized code, without directly participating in process control; or a domain-specific model (such as the BERT model for the power sector) can be used instead of the general large language model to reduce computational overhead, but the level of intelligence will be correspondingly reduced.
[0073] Compared with the prior art, the advantages of the present invention include: This invention achieves a dynamic balance between global exploration and local development by improving the adaptive ant colony algorithm for global search, improving the bat algorithm for local fine-tuning, and combining the pheromone-frequency coupling mechanism. This effectively avoids the problem of a single algorithm easily getting trapped in local optima and significantly improves the optimization accuracy of maintenance plans under multiple constraints.
[0074] This invention uses knowledge graphs as a dynamic constraint engine to perform real-time legality checks during the process of each ant constructing a solution path and the bat fine-tuning its position. This ensures that the generated maintenance plan strictly follows the power grid topology logic and equipment relationships, significantly reducing the power grid operation risks that may be caused by human arrangement.
[0075] The introduction of the large language model enables dispatchers to directly interact with the system through natural language to adjust maintenance plans. The system can automatically understand the semantics and convert them into executable constraints for the algorithm, while generating interpretable optimization results, thus improving the system's usability and intelligence.
[0076] The hybrid algorithm architecture of this invention supports flexible replacement of the specific implementation strategies of the master optimizer or local fine tuner, such as introducing differential evolution or adaptive weights; it also supports replacing the knowledge graph with other technical solutions such as vector databases, making it easy to adapt to power grid application scenarios of different scales and complexities.
[0077] Based on the same inventive concept, this invention also provides an intelligent scheduling system for power grid maintenance plans based on AI algorithms, comprising: The data acquisition module is used to acquire maintenance tasks reported by maintenance requesting units; acquire power grid equipment models, equipment status data, maintenance work orders, dispatching procedures and historical maintenance plan data of the power grid to be maintained; construct a knowledge graph about equipment and maintenance; and encapsulate the knowledge graph into a legality judgment function.
[0078] The function construction module is used to extract equipment identifiers, time information, resource requirements, and priority markers of the maintenance task as maintenance elements using a pre-trained large language model; based on the maintenance elements, the constraint relationships between the equipment in the maintenance task are translated into structured constraint data for each maintenance task; using the structured constraint data and the preset scheduling cycle days, constraints including equipment mutual exclusion constraints, equipment synchronization constraints, daily work limit constraints, and prohibited time period constraints are constructed, along with a multi-objective optimization function that minimizes the workload balance of the power grid to be maintained, minimizes the power outage loss cost, and minimizes the plan execution deviation.
[0079] The orchestration module employs an improved adaptive ant colony algorithm, which enhances global search in the early stages, accelerates convergence in the mid-stage, and avoids local optima in the later stages, using the pheromone evaporation coefficient to solve a multi-objective optimization function. Each ant selects a start time for each maintenance task to construct a solution path and calls a validity determination function to prune the solution path in real time. When preset switching conditions are met, the optimal solution output by the ant colony algorithm is used as the initial population, and an improved bat algorithm, which correlates the search frequency of bats with the pheromone concentration of the ant colony, is used to locally optimize the optimal solution. After each position update by the bat algorithm, a validity determination function is called to perform constraint verification on the updated position to obtain the final solution. Finally, a full constraint verification is performed on the final solution, and the optimal maintenance orchestration execution plan that satisfies all constraints is output.
[0080] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the aforementioned intelligent scheduling method for power grid maintenance plans based on AI algorithms.
[0081] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described intelligent scheduling method for power grid maintenance plans based on AI algorithms.
[0082] Specific limitations regarding the computational system for the AI-based intelligent scheduling method for power grid maintenance planning can be found in the limitations outlined above, and will not be repeated here. Each module in the aforementioned AI-based intelligent scheduling system for power grid maintenance planning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, facilitating processor execution of the corresponding operations.
[0083] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A method for intelligent scheduling of power grid maintenance plans based on AI algorithms, characterized in that, Includes the following steps: Obtain maintenance tasks reported by maintenance requesting units; obtain power grid equipment models, equipment status data, maintenance work orders, dispatching procedures and historical maintenance plan data of the power grid to be maintained; construct a knowledge graph about equipment and maintenance; and encapsulate the knowledge graph into a legality judgment function. The equipment identifier, time information, resource requirements and priority markers of the maintenance task are extracted using a pre-trained large language model as maintenance elements; based on the maintenance elements, the constraint relationships between the equipment in the maintenance task are translated into structured constraint data for each maintenance task. Using structured constraint data and a preset scheduling cycle of days, we construct constraints including equipment mutual exclusion constraints, equipment synchronization constraints, daily work limit constraints, and prohibited time period constraints, and a multi-objective optimization function to minimize the workload balance of the power grid to be maintained, minimize the power outage loss cost, and minimize the plan execution deviation. An improved adaptive ant colony algorithm, which enhances global search in the early stage, accelerates convergence in the middle stage, and avoids local optima in the later stage, is used to solve the multi-objective optimization function based on the pheromone evaporation coefficient. Each ant selects a start time for each maintenance task to construct a solution path, and a validity determination function is called to prune the solution path in real time. When a preset switching condition is met, the optimal solution output by the ant colony algorithm is used as the initial population. An improved bat algorithm, which correlates the search frequency of bats with the pheromone concentration of the ant colony, is used to locally optimize the optimal solution. After each position update by the bat algorithm, a validity determination function is called to constrain and verify the updated position, thus obtaining the final solution. The validity determination function is called to perform a full constraint check on the final solution and output the optimal maintenance scheduling execution plan that satisfies all constraints.
2. The intelligent scheduling method for power grid maintenance plans based on AI algorithms according to claim 1, characterized in that, The knowledge graph concerning equipment and maintenance includes an equipment topology subgraph, a maintenance constraint subgraph, and a maintenance resource carrying capacity subgraph. The equipment topology subgraph uses transformers, circuit breakers, lines, switchgear, busbars, and substations in the power grid equipment model as nodes, and electrical connections as edges. The maintenance constraint subgraph is based on scheduling procedures, maintenance work orders, and the constraints between equipment in the maintenance tasks, using the equipment to be maintained as nodes, and mutual exclusion constraints, synchronization constraints, topology constraints, power supply constraints, and time window constraints as edges. The maintenance resource carrying capacity subgraph is based on scheduling procedures and historical maintenance plan data, using work teams as nodes, and constraints on the number of work teams, number of personnel, average daily maximum workload, tools, vehicles, and operating radius as edges.
3. The intelligent scheduling method for power grid maintenance plans based on AI algorithms according to claim 1, characterized in that, The search frequency of the bats is related to the pheromone concentration of the ant colony. When the average pheromone concentration of the ant colony is higher than the threshold, the search frequency of the bats increases to conduct a fine search near the high-quality solution; when the average pheromone concentration of the ant colony is lower than the threshold, the search frequency of the bats decreases to expand the search range.
4. The intelligent scheduling method for power grid maintenance plans based on AI algorithms according to claim 1, characterized in that, In the multi-objective optimization function, the workload balance is the minimization of the deviation between the daily workload and the average daily workload within the scheduling cycle; the power outage loss cost is the minimization of the power outage loss corresponding to the start date of each maintenance task; and the plan execution deviation is the minimization of the deviation between the actual start time and the initial reporting time of each maintenance task. The weights of each objective are dynamically adjusted by the large language model according to the business priority.
5. The intelligent scheduling method for power grid maintenance plans based on AI algorithms according to claim 1, characterized in that, The full constraint verification also includes an automatic repair strategy, which specifically includes: when repairing mutual exclusion conflicts, the device that starts later will be postponed until the device that starts earlier finishes; when repairing synchronization conflicts, the start time will be unified to a later start time; when repairing resource overload, the low-priority devices on the overload day will be shifted to the back and the legality judgment function will be called again for verification; when repairing prohibited repair periods, the entire maintenance window will be shifted out of the prohibited repair interval.
6. The intelligent scheduling method for power grid maintenance plans based on AI algorithms according to claim 1, characterized in that, It also includes inputting the optimal maintenance scheduling execution plan into a large language model to generate a natural language explanation that includes information on changes in workload balance, changes in power outage loss costs, conflict adjustment status, and power supply protection period avoidance.
7. The intelligent scheduling method for power grid maintenance plans based on AI algorithms according to claim 1, characterized in that, The translation rules for natural language constraints into structured constraint data include: two devices needing to be repaired simultaneously are translated into synchronization constraints; two devices not being repaired simultaneously are translated into mutual exclusion constraints; one device must complete its repair before another device is translated into timing constraints; devices not being allowed to be repaired during a specified period are translated into prohibited time period constraints; and a preset number of devices being repaired simultaneously per day are translated into resource capacity constraints.
8. A smart scheduling system for power grid maintenance plans based on AI algorithms, characterized in that, include: The data acquisition module is used to acquire maintenance tasks reported by the maintenance requesting unit; Acquire the power grid equipment model, equipment status data, maintenance work orders, dispatching procedures and historical maintenance plan data of the power grid to be inspected, construct a knowledge graph about equipment and maintenance, and encapsulate the knowledge graph into a legality judgment function; The function construction module is used to extract the equipment identifier, time information, resource requirements and priority markers of the maintenance task as maintenance elements using a pre-trained large language model; based on the maintenance elements, the constraint relationship between each piece of equipment in the maintenance task is translated into structured constraint data for each maintenance task. Using structured constraint data and a preset scheduling cycle of days, we construct constraints including equipment mutual exclusion constraints, equipment synchronization constraints, daily work limit constraints, and prohibited time period constraints, and a multi-objective optimization function to minimize the workload balance of the power grid to be maintained, minimize the power outage loss cost, and minimize the plan execution deviation. The orchestration module employs an improved adaptive ant colony algorithm, which enhances global search in the early stages, accelerates convergence in the mid-stage, and avoids local optima in the later stages, based on the pheromone evaporation coefficient. Each ant selects a start time for each maintenance task to construct a solution path and calls a validity determination function to prune the solution path in real time. When preset switching conditions are met, the optimal solution output by the ant colony algorithm is used as the initial population, and an improved bat algorithm, which correlates the search frequency of bats with the pheromone concentration of the ant colony, is used to locally optimize the optimal solution. After each position update by the bat algorithm, a validity determination function is called to constrain and verify the updated position, yielding the final solution. The validity determination function is called to perform a full constraint check on the final solution and output the optimal maintenance scheduling execution plan that satisfies all constraints.
9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.