Micro-grid scheduling method and device based on large scheduling model, computer device and storage medium

Abstract

The application discloses a micro-grid scheduling method and device based on a scheduling large model, computer equipment and a storage medium. Incremental knowledge newly released is added to a knowledge graph as a new node; a pre-trained scheduling model in which an industry standard is internalized is obtained based on a pre-training of the knowledge graph updated based on the new node; fine-tuning reinforcement training is supervised based on artificially annotated scheduling samples; a subgraph matched with a scheduling scene is extracted and parameter scene migration is performed to obtain a final scheduling model, and scheduling inference is performed on a candidate scheduling scheme verified by rule constraints. The model of the application systematically integrates industry standard constraint systems into the training process, truly learns compliance constraints in the training and inference stages, rather than temporarily retrieving only during inference. The model increases systematic understanding of industry specifications, reduces the generation probability of non-compliant scheduling schemes from the source, ensures that the final output scheduling instruction strictly meets the safety standard requirements, and enhances the micro-grid scheduling effect based on the scheduling large model.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of power system automation and artificial intelligence, and in particular to microgrid dispatching methods, devices, computer equipment and storage media based on large-scale dispatching models. Background Technology

[0002] When microgrids perform scheduling based on existing data-driven scheduling models, the knowledge acquisition methods for these models are mainly twofold: first, incorporating operating procedures and safety standards into the model input prompts as text; and second, encoding industry knowledge into a rule base for post-processing verification. However, the first method essentially relies on the model's contextual understanding capabilities, while the second is a post-event remedial mechanism, leading to the need for repeated revisions of numerous candidate solutions and low reasoning efficiency. Neither approach achieves deep integration of knowledge at the model parameter level, resulting in superficial industry knowledge fusion.

[0003] Furthermore, the model's output scheduling instructions may violate safety standards without its knowledge. In engineering applications, this means that each scheduling instruction requires manual review, severely limiting the level of automation.

[0004] Finally, power industry standards are constantly being updated, new equipment is continuously being added to the grid, and operating conditions are constantly evolving with the seasons and business development. Once existing models are trained, their knowledge state at the training moment is fixed, leading to a lag in knowledge updates and a gradual degradation in model performance over time. Existing microgrid dispatching technologies, while relying solely on data-driven methods to learn statistical patterns of dispatching patterns from historical data, lack a systematic understanding of industry standards, making it difficult to make optimal decisions while ensuring safety and compliance. In summary, existing technologies suffer from the technical problem of ineffective microgrid dispatching based on large-scale dispatching models. Summary of the Invention

[0005] This application provides a microgrid scheduling method, apparatus, computer equipment, and storage medium based on a large scheduling model, which can solve the technical problem of poor microgrid scheduling effect based on the large scheduling model in the prior art.

[0006] In a first aspect, embodiments of this application provide a microgrid scheduling method based on a large-scale scheduling model, which includes: When newly released incremental knowledge is acquired, it is added as a new node to the preset knowledge graph. The incremental knowledge includes newly released standards and specifications and operating parameters of newly connected devices. Based on the knowledge graph updated by the new nodes, the pre-set base model is pre-trained to obtain the first scheduling model with industry standards internalized in the parameters. The first scheduling model is subjected to supervised fine-tuning and reinforcement learning iterative training based on manually labeled scheduling samples to obtain the fully trained second scheduling model. Extract a knowledge subgraph that matches the target scheduling scenario from the knowledge graph updated by the new node, and perform parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain the final scheduling model adapted to the target scheduling scenario. Based on the final scheduling model, scheduling reasoning is performed on the input real-time operating data of the microgrid, and candidate scheduling schemes verified by hard constraints and soft constraints are used as the basis for microgrid scheduling.

[0007] In some embodiments, when newly published incremental knowledge is acquired, it is added as a new node to a preset knowledge graph, including: In response to obtaining newly released standard documents or new device parameters, constraint rules are extracted from the newly released standard documents and operating parameters of new access devices are extracted from the new device parameters, and combined to obtain triples; Based on the preset confidence threshold and the confidence score of the obtained triples, the triples are classified and stored in different levels of the preset knowledge graph. Generate hierarchical map snapshots for later use in tracking changes.

[0008] In some embodiments, a pre-trained base model is performed based on the knowledge graph updated by the new nodes to obtain a pre-trained first scheduling model whose parameters internalize industry standards, including: The triples in the knowledge graph updated by the new nodes are transformed into natural language descriptions and mixed with a preset general power corpus to obtain a mixed corpus. The triples include industry constraint rules. Based on the hybrid corpus, a pre-trained base model with knowledge augmentation is performed to obtain a first scheduling model whose model parameters incorporate industry rules.

[0009] In some embodiments, a knowledge subgraph matching the target scheduling scenario is extracted from the knowledge graph updated by the new node, and the parameters of the fully trained second scheduling model are transferred based on the knowledge subgraph to obtain a final scheduling model adapted to the target scheduling scenario, including: Extract a knowledge subgraph that matches the target scheduling scenario from the updated knowledge graph of the new node; Device features, topology features, and load features are extracted from the knowledge subgraph and input into a preset graph attention coding network to generate the corresponding scheduling scenario description vector. Freeze the main parameters of the second scheduling model, and train the scene encoder, scene private parameters and scheduling style parameters in the second scheduling model based on the scheduling scene description vector to obtain the final scheduling model that follows the scheduling scene. The final scheduling model is adapted to the target scheduling scene.

[0010] In some embodiments, the scheduling samples are in the form of sample triples, which include historical scheduling conditions, scheduling instructions, and historical constraints. The first scheduling model is then subjected to supervised fine-tuning and reinforcement learning iterative training based on manually labeled scheduling samples to obtain a fully trained second scheduling model, including: The pre-trained first scheduling model is trained under supervision based on the scheduling samples, and basic scheduling capabilities are learned from the historical scheduling conditions of the sample triplet. The supervised training model is further reinforced based on the composite reward, and the economic scheduling capability under compliance hard constraints is learned from the composite reward to obtain the fully trained second scheduling model. The composite reward includes compliance reward, security margin reward and economic reward.

[0011] In some embodiments, the knowledge graph comprises a device layer, a topology layer, an operational layer, and a rule layer. Before adding newly released incremental knowledge as a new node to the preset knowledge graph upon acquisition, the process includes: The device layer of the knowledge graph is built based on physical entities and their attribute vectors. The topology layer of the knowledge graph is established based on the device relationships and device parameters between devices. The device relationships include electrical connection relationships, feeder affiliation relationships, and location inclusion relationships. The knowledge graph's operational layer is built by mining time-series patterns, extracting typical operating condition patterns and coping strategies from preset historical operating data. The rule layer of the knowledge graph is obtained by encoding the preset constraint rules.

[0012] In some embodiments, scheduling inference is performed on the input real-time operating data of the microgrid based on the final scheduling model, and the microgrid scheduling is based on candidate scheduling schemes verified by hard constraints and soft constraints, including: The real-time operating data of the microgrid is input into the final scheduling model. The real-time operating data of the microgrid includes at least one of the following: voltage data, line current, active or reactive power, irradiance, temperature, and frequency. Based on the final scheduling model, rule matching and retrieval are performed on the real-time operating data of the microgrid to obtain the applicable hard constraint rules and soft constraint rules corresponding to the operating conditions. Candidate scheduling schemes are generated using an autoregressive approach; The candidate scheduling scheme is subjected to compliance verification. If it does not violate the hard constraint rules and the degree of deviation from the soft constraint rules is within the preset deviation threshold range, the verification is passed. The microgrid scheduling will be based on candidate scheduling schemes that have been verified by both hard and soft constraints.

[0013] Secondly, embodiments of this application also provide a microgrid scheduling device based on a large-scale scheduling model, the device comprising: The knowledge graph update unit is used to add newly released incremental knowledge as a new node to the preset knowledge graph when it is acquired. The incremental knowledge includes newly released standard specifications and operating parameters of newly accessed devices. The pre-training unit is used to pre-train the pre-set base model based on the knowledge graph updated by the new node, so as to obtain the first scheduling model with industry standards internalized in the parameters. The reinforcement training unit is used to perform supervised fine-tuning and reinforcement learning iterative training on the first scheduling model based on manually labeled scheduling samples to obtain the fully trained second scheduling model. The scenario migration unit is used to extract a knowledge subgraph that matches the target scheduling scenario from the knowledge graph updated by the new node, and perform parameter migration on the fully trained second scheduling model based on the knowledge subgraph to obtain a final scheduling model that adapts to the target scheduling scenario. The scheduling reasoning unit is used to perform scheduling reasoning on the input real-time operating data of the microgrid based on the final scheduling model, and to perform microgrid scheduling based on the candidate scheduling schemes verified by hard constraints and soft constraints.

[0014] Thirdly, embodiments of this application also provide a microgrid scheduling computer device based on a large scheduling model, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-mentioned method.

[0015] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, can implement the above-described method.

[0016] This application provides a microgrid scheduling method, apparatus, computer equipment, and storage medium based on a large scheduling model. The method includes: upon acquiring newly released incremental knowledge, adding it as a new node to a preset knowledge graph, wherein the incremental knowledge includes newly released standards and specifications and operating parameters of newly connected equipment; pre-training a preset base model based on the knowledge graph updated by the new node to obtain a pre-trained first scheduling model whose parameters internalize industry standards; performing supervised fine-tuning and reinforcement learning iterative training on the first scheduling model based on manually labeled scheduling samples to obtain a fully trained second scheduling model; extracting a knowledge subgraph matching the target scheduling scenario from the knowledge graph updated by the new node, and performing parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain a final scheduling model adapted to the target scheduling scenario; and performing scheduling inference on the input real-time operating data of the microgrid based on the final scheduling model, and using candidate scheduling schemes verified by hard constraints and soft constraints as the basis for microgrid scheduling. In this application, not only are statistical patterns of scheduling patterns learned from historical data, but the large-scale scheduling model also systematically integrates these industry standard constraints into the training process. The model truly learns compliance constraints during both training and inference phases, rather than merely retrieving them temporarily during inference. Internalizing rule constraints into the model parameters reduces the likelihood of rule omissions or conflicts in complex, multi-constraint scenarios, enhances the systematic understanding of industry standards, and enables better scheduling decisions under the premise of safety and compliance. This reduces the probability of generating non-compliant scheduling schemes from the source, ensuring that the final output scheduling instructions strictly meet safety standard requirements, and enhancing the microgrid scheduling effect based on the large-scale scheduling model. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the microgrid scheduling method based on a large scheduling model provided in this application embodiment; Figure 2 A schematic block diagram of a microgrid scheduling device based on a large scheduling model provided in this application embodiment; Figure 3 A schematic block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0021] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] When microgrids perform scheduling based on existing data-driven scheduling models, the knowledge acquisition methods for these models are mainly twofold: first, incorporating operating procedures and safety standards into the model input prompts as text; and second, encoding industry knowledge into a rule base for post-processing verification. However, the first method essentially relies on the model's contextual understanding capabilities, while the second is a post-event remedial mechanism, leading to the need for repeated revisions of numerous candidate solutions and low reasoning efficiency. Neither approach achieves deep integration of knowledge at the model parameter level, resulting in superficial industry knowledge fusion.

[0024] Furthermore, the model's output scheduling instructions may violate safety standards without its knowledge. In engineering applications, this means that each scheduling instruction requires manual review, severely limiting the level of automation.

[0025] Finally, power industry standards are constantly being updated, new equipment is continuously being added to the grid, and operating conditions are constantly evolving with the seasons and business development. Once existing models are trained, their knowledge state at the training moment is fixed, leading to a lag in knowledge updates and a gradual degradation in model performance over time. Existing microgrid dispatching technologies, while relying solely on data-driven methods to learn statistical patterns of dispatching patterns from historical data, lack a systematic understanding of industry standards, making it difficult to make optimal decisions while ensuring safety and compliance. In summary, existing technologies suffer from the technical problem of ineffective microgrid dispatching based on large-scale dispatching models.

[0026] To address the aforementioned technical issues, this application proposes a microgrid scheduling method, device, computer equipment, and storage medium based on a large-scale scheduling model. This application internalizes rule constraints into the model parameters, avoiding rule omissions or conflicts when facing complex, multi-constraint scenarios. The model truly learns to comply with constraints during the training and inference phases, whereas existing large-scale model scheduling schemes systematically encode these hard constraints and integrate them into the training process.

[0027] Figure 1 This is a flowchart illustrating the microgrid scheduling method based on a large scheduling model provided in this application. Figure 1 As shown, the method includes the following steps S110-S150: S110. When newly released incremental knowledge is obtained, it is added as a new node to the preset knowledge graph. The incremental knowledge includes newly released standard specifications and operating parameters of newly accessed devices. The knowledge graph consists of a device layer, a topology layer, an operation layer, and a rule layer. Before S110, steps A1-A4 are included: A1. The device layer of the knowledge graph is established based on physical entities and their attribute vectors; Taking the four-layer knowledge graph of a microgrid in an industrial park as an example, a knowledge graph is constructed.

[0028] The industrial park's microgrid includes 4 photovoltaic arrays, 2 sets of lithium iron phosphate energy storage systems, 6 CNC machining lines, 8 injection molding machines, 20 DC fast charging piles, 4 photovoltaic inverters, 2 bidirectional energy storage converters, 2 10kV / 0.4kV transformers, 1 grid-connected circuit breaker, and 8 feeder circuit breakers. The total installed capacity is 2MWp of photovoltaic power, 1MWh / 500kW of energy storage, and a peak load of approximately 2.8MW.

[0029] Based on the equipment ledger of the aforementioned industrial park microgrid, a device layer of a knowledge graph is established. Specifically, a corresponding entity node is created for each piece of equipment.

[0030] Taking an energy storage system as an example, a device layer node is created for energy storage system 1, with the following attribute vectors: rated capacity = 500kWh, maximum charge / discharge power = 250kW, charging efficiency = 0.95, discharging efficiency = 0.95, SOC (State of Charge) safety lower limit SOC_min = 0.1, SOC safety upper limit SOC_max = 0.9, cycle life = 6000 cycles, battery type = lithium iron phosphate, production date = June 2025.

[0031] Taking a photovoltaic array as another example, a device layer node is created for photovoltaic array 1, with the following attribute vectors: rated power = 500kWp, installation tilt angle = 25°, azimuth angle = 15° south of west, temperature coefficient = -0.35% / ℃, first year degradation rate = 2.5%, annual degradation rate = 0.5%.

[0032] Taking the CNC machining production line as another example, create an equipment layer node for CNC production line 1, with the following attribute vector: rated power = 300kW, load type = flexible and controllable, allowable adjustment range = 30%~100%, adjustment rate limit = 50kW / min, daily allowable running time = 8:00-22:00, production priority = high.

[0033] Create device layer nodes for each physical device and attach complete attribute vectors to complete the construction of the device layer.

[0034] A2. Establish the topology layer of the knowledge graph based on the device relationships and device parameters between devices, wherein the device relationships include electrical connection relationships, feeder affiliation relationships, and location inclusion relationships;

[0035] Based on the primary wiring diagram of the industrial park's microgrid, a topology layer of the knowledge graph is established, and electrical connection relationships are created. An electrical connection edge is established between each pair of directly electrically connected devices. The output terminal of photovoltaic inverter 1 is connected to section A of the 0.4kV bus via a cable, establishing a connection edge. This edge includes line parameters: cable type, line impedance R = 0.075Ω / km, X = 0.080Ω / km, maximum current carrying capacity = 380A, and line length = 0.15km.

[0036] Establish feeder affiliation relationships. This industrial park adopts a 10kV double-feeder power supply structure, with two feeders (Feeder A and Feeder B) each supplying a section of 0.4kV busbar, and the two feeders are mutually redundant via a tie switch. Mark the feeder affiliation of all equipment according to its assigned feeder. For example, section A of 0.4kV busbar and all equipment connected to it belong to feeder A.

[0037] Establish location inclusion relationships, such as photovoltaic arrays 1-4 located in areas A and B on the factory roof, energy storage system located in power distribution room 1, and CNC production line located on the first floor of the processing workshop, and establish inclusion relationships between equipment and their physical locations.

[0038] Once the topology layer is built, inference is supported. When the current monitoring value of feeder A reaches 350A, all devices connected to feeder A are retrieved along the feeder affiliation. By combining the real-time power data of each device, it can be quickly determined which devices have adjustable power, providing topology-aware decision support for overload control.

[0039] A3. Extract time-series patterns, typical operating condition patterns, and coping strategies from preset historical operating data, and establish the operating layer of the knowledge graph;

[0040] Historical operating data of the industrial park's microgrid over the past year were obtained, and K-means clustering algorithm was used to perform cluster analysis on the daily load curves, extracting 12 typical operating modes: As an example, a typical weekday daytime solar PV peak pattern is characterized by solar PV output reaching over 80% of installed capacity between 10:00 and 15:00, with industrial production lines running at full speed and charging pile loads gradually increasing. This pattern occurs on sunny weekdays, with an average frequency of approximately 120 days per year. The corresponding strategies include using energy storage to charge at an appropriate power level to absorb the solar PV surplus, maintaining the grid connection power factor above 0.95, and preparing backup responses for midday solar PV output fluctuations.

[0041] As an example, the peak load pattern on weekday evenings is characterized by a rapid decline in photovoltaic output to zero between 5:00 PM and 9:00 PM, while industrial production lines continue operating, coupled with concentrated charging of electric vehicles, forming the peak load period of the day. This pattern occurs every weekday. The corresponding strategy is to switch energy storage to discharge mode to support the evening peak, prioritizing power supply to production lines, and appropriately reducing the power output of charging piles to participate in demand response.

[0042] As an example, the low-load pattern on rest days is characterized by a daily load of 30% to 40% of that on weekdays, with photovoltaic power generation primarily used for grid feeding. This pattern occurs on Saturdays, Sundays, and public holidays. The corresponding strategies are: energy storage utilizes off-peak electricity for low-cost charging; photovoltaic power output is prioritized for grid feeding; and energy storage adjusts its charging and discharging strategies to align with the expected load for the following day.

[0043] The above 12 typical operating conditions are stored in the knowledge graph as nodes of the operation layer. Each node contains a pattern feature description vector, time distribution pattern and response strategy summary.

[0044] A4. Encode the preset constraint rules to obtain the rule layer of the knowledge graph.

[0045] As an example, constraint rules were extracted from standard documents such as the "Technical Specification for Electrochemical Energy Storage System Connection to the Power Grid" and the "Power Grid Operation Guidelines," as well as the power company's corporate operation procedures, and a rule layer was established through coding. Hard constraints include voltage deviation hard constraints, frequency deviation hard constraints, feeder current carrying capacity hard constraints, and energy storage SOC safety hard constraints. Soft constraints include SOC comfort range soft constraints and weekday energy storage preparation soft constraints.

[0046] The above steps completed the construction of a four-layer knowledge graph for the industrial park microgrid. The relationships between the four layers are as follows: the device layer provides basic entity nodes and attribute information for the entire graph; the topology layer establishes connections and attribution relationships between devices based on the device layer nodes, providing spatial structural constraints; the operation layer, independent of specific device instances, stores abstract knowledge of typical time-series patterns; and the rule layer uses device attributes from the device layer and connection relationships from the topology layer to make conditional judgments. These four layers work together to provide structured knowledge support for the subsequent training and inference of the larger model.

[0047] S110 includes S1101-S1103: This step enables the automatic extraction and hierarchical storage of newly released standards and specifications and parameters of newly connected devices, allowing the knowledge graph to continuously absorb the latest changes in industry knowledge and providing timeliness assurance for model training.

[0048] S1101. In response to obtaining a newly released standard document or new device parameters, extract constraint rules from the newly released standard document and extract new access devices and operating parameters of the new access devices from the new device parameters, and combine them to obtain a triple. In response to obtaining newly released standard documents or new device parameters, constraint rules are extracted from the newly released standard documents and operating parameters of new access devices are extracted from the new device parameters, and then combined to obtain a triplet.

[0049] Specifically, the newly released standard documents include, but are not limited to: industry standard documents, enterprise standard documents, and revised versions of various power safety regulations. New equipment parameters include nameplate parameters, factory test data, and grid connection test reports for distributed resources such as photovoltaic arrays, wind turbine generators, energy storage systems, converters, and switchgear newly connected to microgrids.

[0050] A BERT-based named entity recognition model is used to parse standard documents, including: identifying device entity names, parameter entity names, and constraint entity names in the document; using a relation classification model to identify the semantic relationship types between entities; and combining the identified entities into constraint rule triples according to the head entity-relationship-tail entity structure. These constraint rules include hard constraints and soft constraints. Hard constraints are physical and safety constraints that must be strictly satisfied, while soft constraints are optimization objectives and operational preferences.

[0051] The Named Entity Recognition (NAME) model identifies the device type entity, device name entity, rated parameter entity (including rated power, rated capacity, rated voltage, rated current, etc.) and operating characteristic entity (including efficiency curve, decay function, charge and discharge characteristics, etc.) in the new device parameter document; constructs the ontology description node of the new device, and attaches the extracted attribute information to the corresponding attribute fields.

[0052] Each extracted head entity-relationship-tail entity pair is combined into a structured triple. Each triple is assigned a confidence score by the named entity recognition model and the relation classification model during generation. The confidence score reflects the reliability of the triple extraction result.

[0053] Furthermore, for rule-based triples, the constraint rules are encoded into the Datalog rule language format to support chained reasoning on the knowledge graph.

[0054] S1102. Based on the preset confidence threshold and the confidence score of the obtained triples, the triples are classified and stored in different levels of the preset knowledge graph. Based on the preset confidence threshold and the confidence score of the obtained triples, the triples are classified and stored in different levels of the preset knowledge graph.

[0055] Specifically, the preset confidence threshold is denoted as τ_c, with a recommended value of 0.85, which can be adjusted within the range of 0.75 to 0.95 depending on the actual application scenario.

[0056] The tiered warehousing strategy is as follows: When the confidence score c of a triple is greater than or equal to τ_c, the triple is considered to have high reliability, and is directly added to the corresponding level of the knowledge graph. Specifically, device-related triples are added to the device layer, topology-related triples to the topology layer, operation mode-related triples to the operation layer, and constraint rule-related triples to the rule layer.

[0057] When the confidence score of a triple satisfies 0.5 ≤ c < τ_c, the triple is considered to have moderate reliability and is sent to the manual review queue for verification by reviewers. Triples that pass manual review are added to the corresponding level of the knowledge graph, while triples that fail manual review are discarded.

[0058] When the confidence score c of a triple is less than 0.5, the triple is deemed unreliable and is discarded.

[0059] The device layer carries the ontology description and attribute vectors of each physical entity in the microgrid. The physical entities include distributed power generation nodes, energy storage system nodes, load equipment nodes, converter equipment nodes, and switch equipment nodes. The topology layer describes the electrical connection relationships and spatial affiliation relationships between devices. The operation layer carries the time-series pattern knowledge of historical operation data. The rule layer encodes power safety standards into computable hard constraint functions and soft constraint functions, supporting chained reasoning.

[0060] S1103. Generate a snapshot of the hierarchical map for later use in tracking changes.

[0061] Generate a snapshot of the knowledge graph after hierarchical storage, so that it can be called when making changes and making retrospectives later.

[0062] Specifically, after each knowledge injection operation, a complete snapshot version of the current knowledge graph is automatically generated. The graph snapshot includes the following: a complete backup of all nodes and relationships in the current graph; a log record of this injection operation, including the injection time, source document name, list of triples added to the database and their confidence scores; and a record of changes since the previous version, recording the types of operations for addition, deletion and modification at the triple granularity.

[0063] When a subsequent detection indicates that a structural change in the knowledge graph corresponding to the snapshot affects the rule query results of the downstream model, an incremental update notification is triggered to initiate the knowledge augmentation pre-training or incremental fine-tuning process in the subsequent stages.

[0064] Furthermore, knowledge graph snapshots support change tracking and rollback operations. When an error is found in the content of a knowledge injection or it needs to be undone, a historical snapshot version can be called to restore the knowledge graph to its state before the injection.

[0065] Furthermore, the knowledge graph is continuously cyclically executed throughout its entire lifecycle, allowing it to dynamically evolve as industry standards are updated and microgrid devices are replaced.

[0066] S120. Based on the knowledge graph updated by the new node, the preset base model is pre-trained to obtain the first scheduling model with industry standards internalized in the parameters. By deeply integrating industry knowledge from knowledge graphs at the model parameter level, large models can internalize industry knowledge during the pre-training stage, rather than only retrieving it temporarily during inference. This overcomes the technical shortcomings of existing technologies, such as shallow knowledge integration and rule omissions or conflicts when facing complex multi-constraint scenarios.

[0067] S120 includes S1201-S1202: S1201. Convert the triples in the knowledge graph updated by the new node into natural language descriptions, and mix them with the preset general power corpus to obtain a mixed corpus. The triples include industry constraint rules. The triples in the knowledge graph updated by the new nodes are transformed into natural language descriptions and mixed with the preset general power corpus to obtain a hybrid corpus. The triples include industry constraint rules.

[0068] Specifically, the knowledge graph updated with new nodes refers to a four-layer knowledge graph after incremental knowledge injection, which includes a device layer, a topology layer, an operational layer, and a rule layer. When the knowledge graph is incrementally updated, the system automatically triggers a corpus update process, converting the newly added or changed knowledge nodes into natural language descriptions.

[0069] A multi-granularity knowledge description generation strategy is adopted, specifically including the following three granularities: fact-level description, rule-level description, and scenario-level description.

[0070] Transform single triples into concise factual statements. Fact-level descriptions are used to express basic factual knowledge in the knowledge graph, including device attributes, parameter values, static relationships, etc. Fact-level descriptions are generated using a template-filling method, with multiple pre-defined sentence templates. The appropriate template is selected for filling based on the relationship type of the triple.

[0071] Transform the Datalog rules or constraint functions in the rule layer into fluent natural language rule expressions, emphasizing the conditions, decision logic, and compliance requirements of the rules, so that the model can understand the meaning of the rules and the applicable scenarios.

[0072] Combining multiple related triples into a coherent scene description paragraph provides richer contextual information, enabling the model to understand the relationships between knowledge and the overall operational logic.

[0073] The pre-defined general power corpus refers to a large-scale general power text corpus collected from publicly available documents, technical reports, operating procedures, dispatch logs, and other sources in the power industry. This corpus is used to maintain the model's understanding of common power industry terminology and basic knowledge. The general power corpus and the knowledge augmentation corpus share the same semantic space, but the general corpus does not contain structured knowledge descriptions generated from knowledge graphs.

[0074] The generated knowledge-enhanced corpus is combined with the preset general power corpus according to a preset mixing ratio.

[0075] S1202. Based on the hybrid corpus, perform knowledge-enhanced pre-training on the preset base model to obtain a pre-trained first scheduling model whose model parameters incorporate industry rules.

[0076] A key challenge in power dispatching lies in the extreme scarcity of samples of illegal dispatching practices. In practice, safety mechanisms prevent the vast majority of violations from occurring, resulting in models lacking sufficient learning of compliance boundaries. This invention proposes a method for constructing compliance boundary critical point samples based on knowledge graph rules.

[0077] Based on the mixed corpus, a pre-trained base model with knowledge augmentation is performed to obtain a pre-trained first scheduling model whose model parameters incorporate industry rules.

[0078] Specifically, the preset base model is an open-source pre-trained Transformer architecture large language model. The base model already has general natural language understanding and generation capabilities, but lacks power industry expertise.

[0079] The base model is further pre-trained using an autoregressive language modeling task. Specifically, text sequences from the mixed corpus are used as training samples. The cross-entropy loss function is employed, and the predicted probability distribution of the next word is calculated through forward propagation. The model parameters are updated through backpropagation. The training objective is to minimize the negative log-likelihood loss between the predicted sequence and the true sequence.

[0080] Comparative experiments verified that deep internalization of industry knowledge is achieved through parameter-level knowledge fusion. Preferably, after the knowledge graph undergoes incremental updates, it can be incrementally executed, that is, only the newly added knowledge is converted into natural language descriptions, mixed with a subset of the training data from the previous stage, and then subjected to lightweight incremental pre-training, so that the model parameters can quickly absorb the latest knowledge without the need for full retraining.

[0081] As an example, specifically, for an existing compliant scheduling sample y_i, the hard constraint function g_h(x,u) is first continuously relaxed to construct a differentiable version:

[0082] Here, u represents the scheduling action, which is the variable that needs to be decided, optimized, or adjusted during the scheduling process. x represents the state variable, environment state, or scheduling context other than the scheduling action. c(x,u) = c_limit - c_actual(x,u) is the constraint margin, with positive values ​​indicating satisfaction and negative values ​​indicating violation. c_limit is the constraint limit, c_actual is the actual value of the constraint index, σ(·) is the sigmoid function, and κ is the temperature coefficient (κ = 10 is recommended). As it approaches ∞, the continuous relaxation function approaches the original Boolean function g_h. The continuously relaxed function is differentiable with respect to the scheduling action u, and the gradient direction can be calculated.

[0083] By utilizing the gradient of the continuous relaxation function, a minimum perturbation is applied in its opposite direction so that the sample just breaks through the compliance boundary: the negative sample constructed in this way is located at the critical point of the compliance boundary, which is the most informative for the model to learn the compliance boundary.

[0084] After constructing positive and negative sample pairs, the InfoNCE contrastive loss function is used for training. Positive sample pairs are the encoded representations of compliant scheduling and its augmented view under the same working condition, while negative sample pairs include non-compliant scheduling schemes and irrelevant scheduling schemes across working conditions. Through contrastive learning, the encoded representations of the scheduling schemes are made closer to compliant scheduling and further away from non-compliant scheduling.

[0085] S130. Based on manually labeled scheduling samples, the first scheduling model is subjected to supervised fine-tuning and reinforcement learning iterative training to obtain the fully trained second scheduling model. Encoding security and compliance constraints in knowledge graphs into automatically computed training signals forces large models to learn compliance boundaries during fine-tuning, reducing the probability of generating non-compliant scheduling schemes from the source. This overcomes the technical shortcomings of existing technologies, such as reliance on post-processing manual verification for compliance assurance and the lack of compliance awareness in the models themselves.

[0086] The scheduling sample is in the form of a sample triple, which includes historical scheduling conditions, scheduling instructions, and historical constraints. S130 includes S1301-S1302: S1301. Supervised training of the pre-trained first scheduling model is performed based on the scheduling samples, and basic scheduling capabilities are learned from the historical scheduling conditions of the sample triplet. Specifically, the scheduling sample is a pre-constructed training sample set D = {(x_i, y_i, c_i)} of triplets consisting of scheduling conditions, scheduling instructions, and applicable constraints, i = 1, 2, ..., N. Here, x_i is the natural language description of the current operating condition of the i-th sample, including real-time load data, power generation forecast, equipment status, electricity price signals, and other information; y_i is the standard scheduling instruction sequence corresponding to the i-th sample, including the power setpoints of each controllable device; and c_i is the set of constraints related to the current operating condition extracted from the rule layer of the knowledge graph.

[0087] The scheduling samples are derived from compliant scheduling records verified by operations personnel in historical operational data. The pre-trained first scheduling model, whose parameters already incorporate industry rules, undergoes supervised fine-tuning training on these scheduling samples. Specifically, the operational condition description x_i is input into the model, and the model generates a sequence of scheduling instructions using an autoregressive approach. The difference between the generated sequence and the labeled sequence y_i is calculated using the cross-entropy loss function. During training, the model parameters are updated through backpropagation and gradient descent, and the training objective is to minimize the cross-entropy loss function.

[0088] As an example, the training parameters for supervised fine-tuning are configured as follows, with a learning rate of... to With batch sizes ranging from 16 to 64 and training rounds of 2 to 5, the model learns the mapping relationship between input operating conditions and output scheduling instructions through supervised fine-tuning, and masters the basic scheduling decision-making pattern.

[0089] S1302. Based on the composite reward, the supervised training model is further reinforced to learn the economic scheduling capability under compliance hard constraints from the composite reward, and the second scheduling model after complete training is obtained. The composite reward includes compliance reward, security margin reward and economic reward.

[0090] The supervised training model is further reinforced based on the composite reward, and the economic scheduling capability under compliance hard constraints is learned from the composite reward to obtain the fully trained second scheduling model. The composite reward includes compliance reward, security margin reward and economic reward.

[0091] Specifically, composite reward refers to the unified encoding of compliance verification results, security margin indicators, and economic indicators of the knowledge graph rule layer into a multi-objective reward function, which is used to guide the policy optimization of reinforcement learning algorithms.

[0092] As an example, the composite reward function Specifically, it is expressed as follows: ,in, This represents the current operating condition. The scheduling action output by the model. The weighting parameter represents the compliance reward. In compound rewards, the larger the relative weight, the more it is biased towards satisfying the constraints of the rule layer. The weighting coefficient represents the safety margin reward. In compound rewards, a larger relative weight indicates a tendency to move further away from the safety boundary. The KL divergence penalty coefficient controls the intensity of the penalty for the current policy deviating from the reference policy. This is used to ensure the stability of reinforcement learning training and prevent behavioral degradation caused by drastic deviation from the reference policy in pursuit of rewards. For regularization terms, To determine the conditional probability distribution of scheduling actions output by the scheduling model to be optimized under parameter θ, To supervise the policy distribution of the first scheduling model after fine-tuning and before reinforcement learning begins, Indicates relative to.

[0093] The economic incentive, representing the economic efficiency of the dispatch scheme, drives the model to pursue economic optimization while meeting safety constraints. The economic incentive is a weighted sum of the electricity purchase cost saving rate, photovoltaic (PV) grid integration rate, and energy storage utilization efficiency. The weights can be adjusted according to the optimization objectives of the specific application scenario. Specifically, the electricity purchase cost saving rate refers to the proportion of cost savings from purchasing electricity from the main grid under the current dispatch scheme; the lower the purchase cost, the higher the saving rate. The PV grid integration rate refers to the proportion of actual PV power generation to the theoretical maximum power generation under the current dispatch scheme; the higher the PV grid integration rate, the higher the overall PV grid integration rate. Energy storage utilization efficiency refers to the energy conversion efficiency of the energy storage system during charge-discharge cycles, reflecting the economic efficiency of energy storage dispatch.

[0094] The compliance reward is one of the innovations that distinguishes it from existing technologies. It is automatically calculated by querying the knowledge graph rule engine, measuring the degree to which the current scheduling action satisfies all activation constraints in the rule layer. Specifically, for scheduling actions... The system queries the knowledge graph rule engine to obtain the constraint set related to the current scenario and calculates the weighted average of the satisfaction of all constraints.

[0095] The value is 1 when the hard constraint is met and 0 when it is violated. For soft constraints, the satisfaction is a continuous value sat_c(x, u) = 1 - g_s(x, u), where g_s(x, u) is a continuous penalty function with a value range of [0,1].

[0096] When any hard constraint is violated Setting it directly to -1 forces the model to prioritize learning to avoid hard constraint violations during reinforcement learning training.

[0097] The safety margin reward represents the margin between the scheduling result and the safety boundary. It is used to drive the model to not only meet the basic requirements of hard constraints, but also to leave sufficient safety margin to improve the safety resilience of system operation. The safety margin reward is obtained by calculating the normalized distance between the current value and the constraint boundary of each key safety indicator such as voltage, frequency, line current, and energy storage SOC, and taking the average value.

[0098] Hard and soft constraints in the rule layer play different roles in reinforcement learning. Hard constraints generate a strong penalty signal to the model through the mechanism of directly decrementing compliance rewards, forcing the policy to avoid illegal actions. Soft constraints generate smooth gradients through continuous satisfaction, guiding the policy towards the ideal operating range. The coercive nature of hard constraints and the guiding nature of soft constraints work together to enable the model to continuously optimize scheduling performance within the safe boundary.

[0099] In some embodiments, a KL divergence regularization term D_KL is designed (or The KL divergence regularization term is used to prevent the policy from deviating too far from the reference policy during reinforcement learning, thereby ensuring training stability and preventing behavioral degradation caused by the model excessively pursuing rewards. By subtracting the KL divergence regularization term from the composite reward, the model is prevented from deviating drastically from the reference policy in pursuit of rewards, thus ensuring training stability.

[0100] After the reward function was optimized, the model was forced to learn to pursue economic optimality while satisfying hard constraints, thereby reducing the probability of violations from the source.

[0101] S140. Extract a knowledge subgraph that matches the target scheduling scenario from the knowledge graph updated by the new node, and perform parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain a final scheduling model that adapts to the target scheduling scenario. This technology enables efficient parameter transfer of large models across different microgrid scenarios, allowing the same model to exhibit differentiated scheduling strategies in different scenarios. It also avoids the need for full retraining in each scenario, thus overcoming the technical shortcomings of existing technologies, such as large data transfer requirements, long transfer cycles, and insufficient representation of scenario features.

[0102] S140 includes S1401-S1403: S1401. Extract a knowledge subgraph that matches the target scheduling scenario from the knowledge graph updated by the new node; Extract a knowledge subgraph that matches the target scheduling scenario from the updated knowledge graph of the new node.

[0103] Specifically, the knowledge graph updated by the new node refers to a four-layer knowledge graph after incremental knowledge injection, including a device layer, a topology layer, an operation layer, and a rule layer. When it is necessary to deploy the trained scheduling model to a new target microgrid scenario, the knowledge subgraph related to the target scenario is first extracted from this global knowledge graph.

[0104] The target scheduling scenario refers to the microgrid system for which the large-scale scheduling model is to be deployed. This scenario may differ significantly from the source scenario used in the model's pre-training in terms of topology, equipment configuration, and load characteristics. Target scheduling scenarios include, but are not limited to: industrial park microgrid scenarios, commercial complex microgrid scenarios, remote independent microgrid scenarios, campus microgrid scenarios, and community microgrid scenarios.

[0105] Based on the equipment list and primary wiring diagram of the target microgrid, all nodes and relationships belonging to this scenario are selected from the global knowledge graph. Specifically, starting from the root node of the target scenario, a breadth-first traversal is performed along the edges of the knowledge graph to retrieve all equipment layer nodes, topology layer relationship edges, and typical operating mode nodes marked as belonging to this scenario in the operation layer.

[0106] The extracted knowledge subgraph is represented by graph structure data, where the node set V contains all device entity nodes and pattern nodes in the scenario, and the edge set E contains all topological relationship edges and pattern association edges in the scenario.

[0107] S1402. Extract device features, topology features, and load features from the knowledge subgraph and input them into a preset graph attention coding network to generate the corresponding scheduling scenario description vector. Device features, topology features, and load features are extracted from the knowledge subgraph and input into a pre-defined graph attention encoding network to generate the corresponding scheduling scenario description vector.

[0108] Specifically, equipment features refer to the type encoding and capacity configuration information of each physical entity extracted from the equipment layer nodes of the knowledge subgraph. The equipment types of various devices in the scene, such as photovoltaics, energy storage, wind power, diesel generators, various loads, converters, and switches, are encoded, and numerical parameters such as rated power and rated capacity are normalized and then concatenated into equipment feature vectors.

[0109] Topological features refer to the structural information about the electrical connections between devices extracted from the topological layer of a knowledge subgraph. The topological structure of the knowledge subgraph is represented as an adjacency matrix, where the elements represent the electrical connections between devices, and the attributes of the edges, such as line impedance and current carrying capacity, are used as edge features.

[0110] Load characteristics refer to the time-series pattern features of load extracted from the historical operating data of the target scenario. For example, typical daily load curves of the target scenario over the past 30 days are obtained, and a Fast Fourier Transform is performed on the sample set of daily load curves to extract the top K frequency domain components as the frequency domain features of the load, reflecting the daily and weekly periodicity of the load.

[0111] Graph attention encoding networks are used to encode the topological structure information of the knowledge subgraph into fixed-dimensional scene representation vectors. The graph attention encoding network outputs graph-level scene representation vectors.

[0112] After L-layer message passing, the final hidden states of all nodes are obtained. The hidden states of all nodes are aggregated using a global read function to obtain a graph-level scene structure representation. The average of the final hidden states of all nodes is then taken to obtain the graph-level scene representation.

[0113] Furthermore, the graph-level scene structure representation is concatenated with device features and load frequency domain features, and then input into a feature fusion multilayer perceptron for feature fusion and dimensional transformation, ultimately obtaining a fixed-dimensional scene description vector.

[0114] S1403. Freeze the main parameters of the second scheduling model, and train the scene encoder, scene private parameters and scheduling style parameters in the second scheduling model based on the scheduling scene description vector to obtain the final scheduling model after migration following the scheduling scene. The final scheduling model is adapted to the target scheduling scene.

[0115] The second scheduling model refers to the large scheduling model obtained after reinforcement learning training. Its Transformer parameters have incorporated industry knowledge and scheduling strategies from the source scenario during the previous training.

[0116] The main Transformer parameters of the second scheduling model remain unchanged and do not participate in gradient updates, so as to retain the general power scheduling knowledge and language understanding ability that the model has learned and prevent it from being forgotten when adapting to new scenarios.

[0117] When migrating to a new scene, the knowledge subgraph of the new scene generates a new scene description vector through the scene encoder. Therefore, the scene encoder parameters need to be updated according to the data of the new scene.

[0118] Scene-specific parameters and scheduling style parameters play distinct yet highly complementary roles in the large model, together forming the core of efficient parameter transfer.

[0119] Scene-specific parameters refer to the parameters of the low-rank adaptation matrix, responsible for storing specific private knowledge of the target scene. For example, commercial complexes need to master the skill of flexible daytime air conditioning regulation, while industrial parks need to master the skill of off-peak electricity storage at night. Scene-specific parameters solidify these differentiated skills with a very small number of parameters. Because the main model is frozen, scene-specific parameters ensure that the model does not overwrite the original general power knowledge when absorbing new scene features. This solves the problem of forgetting the old scene during traditional full-scale fine-tuning.

[0120] The scheduling style parameters refer to the multilayer perceptron parameters normalized at the conditional layer. Conditional layer normalization takes the generated scene description vector as conditional input and generates scaling parameter γ(s) and offset parameter β(s) through the multilayer perceptron.

[0121] A single model body can achieve multiple scheduling styles through scheduling style parameters. For example, when the input is a descriptor for an industrial park, a set of γ / β parameters is generated, causing the model to exhibit peak shaving behavior with heavy energy storage; when the input is a descriptor for a commercial complex, another set of scaling parameters γ(s) and offset parameters β(s) is generated, causing the model to favor heavy air conditioning load following. This modulation is global, acting on every Transformer layer, deeply altering the model's decision logic, unlike feature concatenation which only affects the input layer.

[0122] As an example, the scheduling scenario description vector is injected into the Transformer layers of the large model through conditional layer normalization. For each normalized sublayer of the Transformer layer:

[0123] Where x is the input of the current layer, μ and σ are the mean and standard deviation of the input in the feature dimension, respectively, and ⊙ represents "element-wise multiplication" (Hadamard product), which means that multiplying corresponding elements of two vectors of the same dimension results in a vector of the same dimension. γ(s) and β(s) are scaling and offset parameters generated by the scene descriptor through a small multilayer perceptron (MLP). The MLP contains at least one hidden layer with ReLU or GELU activation function, and the hidden layer dimension is recommended to be 1 / 4 to 1 / 2 of the Transformer hidden layer dimension.

[0124] The design intent of this mechanism is that, for the same model subject in different scenarios, only the scenario descriptor s needs to be changed, and the behavior of the Transformer layer will be adjusted accordingly, so that the model can exhibit a differentiated scheduling strategy for different scenario characteristics.

[0125] By simply changing the scene description vector, different behavioral tendencies can be generated without retraining.

[0126] S150. Based on the final scheduling model, perform scheduling reasoning on the input microgrid real-time operating data, and use the candidate scheduling schemes verified by hard constraints and soft constraints as the basis for microgrid scheduling.

[0127] Establishing a compliance closed-loop verification mechanism for scheduling schemes during the inference phase ensures that the scheduling schemes generated by the model undergo independent compliance verification by the knowledge graph rule engine before output. When hard constraints are not met, multiple rounds of correction are automatically triggered to ensure that the final output scheduling instructions strictly meet safety standards. This overcomes the technical defects of existing technologies, such as open-loop inference without verification and reliance on manual review for non-compliant schemes.

[0128] S150 includes S1501-S1505: S1501. Input the real-time operating data of the microgrid into the final scheduling model. The real-time operating data of the microgrid includes at least one of voltage data, line current, active or reactive power, irradiance, temperature and frequency. The real-time operating data of the microgrid is input into the final scheduling model. The real-time operating data of the microgrid includes at least one of the following: voltage data, line current, active or reactive power, irradiance, temperature, and frequency.

[0129] Specifically, the final scheduling model has incorporated industry knowledge and is adapted to the current microgrid scenario.

[0130] The real-time operating data of the microgrid comes from the real-time measurement data of the microgrid data acquisition and monitoring control system, including but not limited to the following physical quantities: voltage amplitude and phase angle of each node, line current of each branch, active power injected or outflowed from each node, state of charge of the energy storage system, ambient irradiance, ambient temperature, system frequency, etc.

[0131] Real-time operating data is a snapshot of the data at the current scheduling time t, which is collected and input into the model according to the preset scheduling cycle.

[0132] Furthermore, real-time operating data needs to undergo data preprocessing before being input into the model, including: outlier detection and removal, missing value imputation, and normalization.

[0133] S1502. Based on the final scheduling model, perform rule matching and retrieval on the real-time operating data of the microgrid to obtain the applicable hard constraint rules and soft constraint rules corresponding to the operating conditions. The retrieval employs a dual-path approach. First, a semantic similarity retrieval path encodes the current operating condition as a query vector q, retrieving the Top-K rule description fragments most semantically similar to the current operating condition from the text description space of the knowledge graph rule layer. This path can retrieve rules semantically related to the current operating condition but whose condition expressions do not precisely match, demonstrating good generalization ability. Second, a rule condition matching path matches the antecedent conditions of all rules one by one in the rule layer based on the specific numerical value of the current operating condition. For each rule, the threshold in its antecedent conditions is compared with the measured value of the current operating condition to determine whether the rule is activated. For example, when the measured SOC of the energy storage is 0.15, the rule "SOC < 0.2" prohibiting discharge is activated; when the measured voltage at the grid connection point is 1.05 pu, the rule |ΔU| / UN > 0.07 activates voltage exceeding the limit, where ΔU represents the voltage change and UN represents the rated voltage. After merging and deduplicating the results of the two search paths, the constraint set C_t = C_hard ∪ C_soft applicable to the current working condition is obtained, where C_hard is the set of activated hard constraint rules and C_soft is the set of activated soft constraint rules.

[0134] Furthermore, the retrieved hard and soft constraint rules are organized into a knowledge context K_t in the form of natural language descriptions, which serves as the reference input for subsequent model inference.

[0135] S1503. Generate candidate scheduling schemes using an autoregressive approach; The acquired real-time operating condition data is converted into a structured text description, including the current values ​​of each key physical quantity and their relative relationship with the rated values / safety boundaries, to obtain a natural language description of the current operating condition.

[0136] The retrieved activated hard and soft constraint rules are organized in natural language as context.

[0137] The scheduling instructions and operating conditions at previous scheduling moments are used to obtain the recent historical scheduling sequence.

[0138] The above-mentioned scheduling instruction sequence, the natural language description of the current working condition, and the recent historical scheduling sequence are concatenated and fed into the final scheduling model. The model predicts word by word from left to right to generate a complete scheduling instruction sequence.

[0139] S1504. Perform compliance verification on the candidate scheduling scheme. If it does not violate the hard constraint rules and the degree of deviation from the soft constraint rules is within the preset deviation threshold range, the verification is passed. The candidate scheduling scheme is verified for compliance. If it does not violate the hard constraint rules and the degree of deviation from the soft constraint rules is within the preset deviation threshold range, the verification is successful.

[0140] The candidate scheduling scheme u_t is sent to the knowledge graph rule engine, and the constraint set C_t retrieved in the steps is verified one by one.

[0141] For each rule c_h in the hard constraint rule C_hard, the rule engine calculates the hard constraint satisfaction Boolean value g_h(u_t) ∈ {0,1}, where 1 indicates satisfaction and 0 indicates violation.

[0142] For each rule c_s in the soft constraint rule C_soft, the rule engine calculates the soft constraint continuous penalty value g_s(u_t) ∈ [0,1]. The larger the value, the greater the degree of deviation from the ideal operating range.

[0143] V_h is the set of hard constraints, S_s is the total degree of soft constraint violation or the comprehensive deviation of soft constraints, and θ_s is the soft constraint tolerance threshold, which is a preset boundary used to determine whether the comprehensive deviation of soft constraints S_s is acceptable.

[0144] If |V_h| = 0 and S_s < θ_s, where θ_s is the soft constraint tolerance threshold, then the scheme is deemed to have passed the compliance verification and proceeds to the next step of execution and distribution.

[0145] If |V_h| > 0, meaning any hard constraint is violated, the solution is deemed unsuccessful and the process proceeds to the correction phase. The set of violated hard constraints is described in natural language and fed back to the model, triggering regeneration.

[0146] If |V_h| = 0 and S_s ≥ θ_s, meaning there is no violation of hard constraints but the deviation of soft constraints exceeds the tolerance threshold, then the scheme is recommended for adjustment and marked as a soft constraint deviation warning. At the same time, details of each violation of soft constraints are given, and the operators decide whether to accept the current scheme or require the model to be regenerated.

[0147] S1505. Microgrid scheduling will be based on candidate scheduling schemes that have been verified by hard and soft constraints.

[0148] Specifically, the verified candidate scheduling scheme u_t is used as the final scheduling instruction sequence at the current moment and is output by the final scheduling model to the microgrid energy management system.

[0149] The microgrid energy management system converts the dispatch command sequence according to the communication protocol format of each device and encapsulates it into messages, then sends them to the local controllers of each controllable device through the industrial communication network. Upon receiving the dispatch commands, each local controller converts them into actual control signals for the physical devices: the photovoltaic inverter adjusts its power output according to the active / reactive power setpoints; the energy storage bidirectional converter adjusts its charging / discharging current and voltage according to the charging / discharging power setpoints; the controllable load adjusts its power consumption according to the power reduction command; and the switching equipment performs opening or closing operations according to the opening / closing signals.

[0150] After the above physical control actions are implemented, the grid connection point voltage is maintained within the qualified range, the control system frequency deviation is within ±0.2Hz, line overload is prevented, the energy storage SOC is kept within the safety boundary, the photovoltaic absorption rate is improved, the electricity purchase cost is reduced, and the power supply reliability is guaranteed. Ultimately, the physical scheduling closed-loop control of the microgrid distributed power output, energy storage charging and discharging, and controllable loads is realized.

[0151] This application provides a microgrid scheduling method, apparatus, computer equipment, and storage medium based on a large scheduling model. The method includes: upon acquiring newly released incremental knowledge, adding it as a new node to a preset knowledge graph, wherein the incremental knowledge includes newly released standards and specifications and operating parameters of newly connected equipment; pre-training a preset base model based on the knowledge graph updated by the new node to obtain a pre-trained first scheduling model whose parameters internalize industry standards; performing supervised fine-tuning and reinforcement learning iterative training on the first scheduling model based on manually labeled scheduling samples to obtain a fully trained second scheduling model; extracting a knowledge subgraph matching the target scheduling scenario from the knowledge graph updated by the new node, and performing parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain a final scheduling model adapted to the target scheduling scenario; and performing scheduling inference on the input real-time operating data of the microgrid based on the final scheduling model, and using candidate scheduling schemes verified by hard constraints and soft constraints as the basis for microgrid scheduling. In this application, not only are statistical patterns of scheduling patterns learned from historical data, but the large-scale scheduling model also systematically integrates these industry standard constraints into the training process. The model truly learns compliance constraints during both training and inference phases, rather than merely retrieving them temporarily during inference. Internalizing rule constraints into the model parameters reduces the likelihood of rule omissions or conflicts in complex, multi-constraint scenarios, enhances the systematic understanding of industry standards, and enables better scheduling decisions under the premise of safety and compliance. This reduces the probability of generating non-compliant scheduling schemes from the source, ensuring that the final output scheduling instructions strictly meet safety standard requirements, and enhancing the microgrid scheduling effect based on the large-scale scheduling model.

[0152] Figure 2This is a schematic block diagram of a microgrid scheduling device based on a large scheduling model, provided in an embodiment of this application. Figure 2 As shown, corresponding to the above methods, this application also provides a microgrid dispatching device 600 based on a large-scale dispatching model. This microgrid dispatching device 600 based on a large-scale dispatching model can be configured in terminals such as desktop computers, tablet computers, and laptops. For details, please refer to... Figure 2 The microgrid scheduling device 600 based on a large scheduling model includes a graph update unit 601, a pre-training unit 602, a reinforcement training unit 603, a scene transfer unit 604, and a scheduling inference unit 605, wherein: The knowledge graph update unit 601 is used to add newly released incremental knowledge as a new node to a preset knowledge graph when it is acquired. The incremental knowledge includes newly released standard specifications and operating parameters of newly accessed devices. The pre-training unit 602 is used to pre-train the preset base model based on the knowledge graph updated by the new node, so as to obtain the first scheduling model with industry standard pre-trained parameters. The reinforcement training unit 603 is used to perform supervised fine-tuning and reinforcement learning iterative training on the first scheduling model based on manually labeled scheduling samples to obtain the fully trained second scheduling model. The scene transfer unit 604 is used to extract a knowledge subgraph that matches the target scheduling scene from the knowledge graph updated by the new node, and perform parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain a final scheduling model that adapts to the target scheduling scene. The scheduling reasoning unit 605 is used to perform scheduling reasoning on the input real-time operating data of the microgrid based on the final scheduling model, and to perform microgrid scheduling based on the candidate scheduling schemes verified by hard constraints and soft constraints.

[0153] In some embodiments, when the graph update unit 601 obtains newly published incremental knowledge, it adds it as a new node to a preset knowledge graph, specifically for: In response to obtaining newly released standard documents or new device parameters, constraint rules are extracted from the newly released standard documents and operating parameters of new access devices are extracted from the new device parameters, and combined to obtain triples; Based on the preset confidence threshold and the confidence score of the obtained triples, the triples are classified and stored in different levels of the preset knowledge graph. Generate hierarchical map snapshots for later use in tracking changes.

[0154] In some embodiments, the pre-training unit 602 performs pre-training on a preset base model based on the knowledge graph updated by new nodes to obtain a first scheduling model whose parameters internalize industry standards, specifically used for:

[0155] The triples in the knowledge graph updated by the new nodes are transformed into natural language descriptions and mixed with a preset general power corpus to obtain a mixed corpus. The triples include industry constraint rules. Based on the hybrid corpus, a pre-trained base model with knowledge augmentation is performed to obtain a first scheduling model whose model parameters incorporate industry rules.

[0156] In some embodiments, the scene migration unit 604 extracts a knowledge subgraph matching the target scheduling scene from the knowledge graph updated by the new node, and performs parameter migration on the fully trained second scheduling model based on the knowledge subgraph to obtain a final scheduling model adapted to the target scheduling scene, specifically used for: Extract a knowledge subgraph that matches the target scheduling scenario from the updated knowledge graph of the new node; Device features, topology features, and load features are extracted from the knowledge subgraph and input into a preset graph attention coding network to generate the corresponding scheduling scenario description vector. Freeze the main parameters of the second scheduling model, and train the scene encoder, scene private parameters and scheduling style parameters in the second scheduling model based on the scheduling scene description vector to obtain the final scheduling model that follows the scheduling scene. The final scheduling model is adapted to the target scheduling scene.

[0157] In some embodiments, the scheduling sample is in the form of a sample triple, which includes historical scheduling conditions, scheduling instructions, and historical constraints. The reinforcement training unit 603 performs supervised fine-tuning and reinforcement learning iterative training on the first scheduling model based on manually labeled scheduling samples to obtain a fully trained second scheduling model, specifically used for: The pre-trained first scheduling model is trained under supervision based on the scheduling samples, and basic scheduling capabilities are learned from the historical scheduling conditions of the sample triplet. The supervised training model is further reinforced based on the composite reward, and the economic scheduling capability under compliance hard constraints is learned from the composite reward to obtain the fully trained second scheduling model. The composite reward includes compliance reward, security margin reward and economic reward.

[0158] In some embodiments, the knowledge graph comprises a device layer, a topology layer, an operation layer, and a rule layer. The microgrid scheduling device 600 based on a large scheduling model further includes a graph construction unit. Before adding newly released incremental knowledge as a new node to the preset knowledge graph, the graph construction unit is specifically used for: The device layer of the knowledge graph is built based on physical entities and their attribute vectors. The topology layer of the knowledge graph is established based on the device relationships and device parameters between devices. The device relationships include electrical connection relationships, feeder affiliation relationships, and location inclusion relationships. The knowledge graph's operational layer is built by mining time-series patterns, extracting typical operating condition patterns and coping strategies from preset historical operating data. The rule layer of the knowledge graph is obtained by encoding the preset constraint rules.

[0159] In some embodiments, the scheduling inference unit 605 performs scheduling inference on the input real-time operating data of the microgrid based on the final scheduling model, and uses candidate scheduling schemes verified by hard and soft constraints as the basis for microgrid scheduling, specifically for: The real-time operating data of the microgrid is input into the final scheduling model. The real-time operating data of the microgrid includes at least one of the following: voltage data, line current, active or reactive power, irradiance, temperature, and frequency. Based on the final scheduling model, rule matching and retrieval are performed on the real-time operating data of the microgrid to obtain the applicable hard constraint rules and soft constraint rules corresponding to the operating conditions. Candidate scheduling schemes are generated using an autoregressive approach; The candidate scheduling scheme is subjected to compliance verification. If it does not violate the hard constraint rules and the degree of deviation from the soft constraint rules is within the preset deviation threshold range, the verification is passed. The microgrid scheduling will be based on candidate scheduling schemes that have been verified by both hard and soft constraints.

[0160] In summary, this application provides a microgrid scheduling method, apparatus, computer equipment, and storage medium based on a large scheduling model. The method includes: upon acquiring newly released incremental knowledge, adding it as a new node to a preset knowledge graph, wherein the incremental knowledge includes newly released standard specifications and operating parameters of newly connected equipment; pre-training a preset base model based on the knowledge graph updated by the new node to obtain a pre-trained first scheduling model whose parameters internalize industry standards; performing supervised fine-tuning and reinforcement learning iterative training on the first scheduling model based on manually labeled scheduling samples to obtain a fully trained second scheduling model; extracting a knowledge subgraph matching the target scheduling scenario from the knowledge graph updated by the new node, and performing parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain a final scheduling model adapted to the target scheduling scenario; and performing scheduling inference on the input real-time operating data of the microgrid based on the final scheduling model, and using candidate scheduling schemes verified by hard constraints and soft constraints as the basis for microgrid scheduling. In this application, not only are statistical patterns of scheduling patterns learned from historical data, but the large-scale scheduling model also systematically integrates these industry standard constraints into the training process. The model truly learns compliance constraints during both training and inference phases, rather than merely retrieving them temporarily during inference. Internalizing rule constraints into the model parameters reduces the likelihood of rule omissions or conflicts in complex, multi-constraint scenarios, enhances the systematic understanding of industry standards, and enables better scheduling decisions under the premise of safety and compliance. This reduces the probability of generating non-compliant scheduling schemes from the source, ensuring that the final output scheduling instructions strictly meet safety standard requirements, and enhancing the microgrid scheduling effect based on the large-scale scheduling model.

[0161] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the microgrid scheduling device and its various units based on the large scheduling model can be found in the corresponding descriptions in the aforementioned method embodiments. For the sake of convenience and brevity, these details will not be repeated here.

[0162] The aforementioned microgrid dispatching device based on a large-scale dispatching model can be implemented as a computer program, which can be used in various ways, such as... Figure 3 It runs on the computer device shown.

[0163] Please see Figure 3 , Figure 3 This is a schematic block diagram of a computer device 700 provided in an embodiment of this application. The computer device 700 can be a terminal or a server. The terminal can be an electronic device with communication functions, such as a smartphone, tablet, laptop, desktop computer, personal digital assistant, or wearable device. The server can be a standalone server or a server cluster composed of multiple servers.

[0164] See Figure 3 The computer device 700 includes a processor 702, a memory, and a network interface 705 connected via a system bus 701. The memory may include a non-volatile storage medium 703 and internal memory 704.

[0165] The non-volatile storage medium 703 can store an operating system 7031 and a computer program 7032. The computer program 7032 includes program instructions that, when executed, cause the processor 702 to perform a microgrid scheduling method based on a large scheduling model.

[0166] The processor 702 provides computing and control capabilities to support the operation of the entire computer device 700.

[0167] The internal memory 704 provides an environment for the execution of the computer program 7032 in the non-volatile storage medium 703. When the computer program 7032 is executed by the processor 702, the processor 702 can execute a microgrid scheduling method based on a large scheduling model.

[0168] This network interface 705 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 700 to which the present application is applied. The specific computer device 700 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0169] The processor 702 is used to run the computer program 7032 stored in the memory to perform the following steps: When newly released incremental knowledge is acquired, it is added as a new node to the preset knowledge graph. The incremental knowledge includes newly released standards and specifications and operating parameters of newly connected devices. Based on the knowledge graph updated by the new nodes, the pre-set base model is pre-trained to obtain the first scheduling model with industry standards internalized in the parameters. The first scheduling model is subjected to supervised fine-tuning and reinforcement learning iterative training based on manually labeled scheduling samples to obtain the fully trained second scheduling model. Extract a knowledge subgraph that matches the target scheduling scenario from the knowledge graph updated by the new node, and perform parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain the final scheduling model adapted to the target scheduling scenario. Based on the final scheduling model, scheduling reasoning is performed on the input real-time operating data of the microgrid, and candidate scheduling schemes verified by hard constraints and soft constraints are used as the basis for microgrid scheduling.

[0170] It should be understood that in the embodiments of this application, the processor 702 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0171] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0172] It should be noted that the specific embodiments are merely illustrative examples intended to aid in understanding the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. The scope of protection of this application is determined by the contents of the claims. Specific embodiments, parameter ranges, or technical features described in the specification should not be construed as undue limitation or expansion of the scope of protection of the claims. The technical effects described in the specification are only used to illustrate the innovation of the present invention. Any technical solution that does not simultaneously possess all the technical features of the present invention, even if it claims to solve the same technical problem, does not fall within the scope of protection of the present invention.

[0173] Therefore, this application also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the following steps: When newly released incremental knowledge is acquired, it is added as a new node to the preset knowledge graph. The incremental knowledge includes newly released standards and specifications and operating parameters of newly connected devices. Based on the knowledge graph updated by the new nodes, the pre-set base model is pre-trained to obtain the first scheduling model with industry standards internalized in the parameters. The first scheduling model is subjected to supervised fine-tuning and reinforcement learning iterative training based on manually labeled scheduling samples to obtain the fully trained second scheduling model. Extract a knowledge subgraph that matches the target scheduling scenario from the knowledge graph updated by the new node, and perform parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain the final scheduling model adapted to the target scheduling scenario. Based on the final scheduling model, scheduling reasoning is performed on the input real-time operating data of the microgrid, and candidate scheduling schemes verified by hard constraints and soft constraints are used as the basis for microgrid scheduling.

[0174] The storage medium can be any computer-readable storage medium that can store program code, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0175] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0176] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0177] The steps in the methods of this application embodiment can be adjusted, merged, or deleted according to actual needs. The units in the apparatus of this application embodiment can be merged, divided, or deleted according to actual needs. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0178] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.

[0179] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A microgrid scheduling method based on a large-scale scheduling model, characterized in that, The method includes: When newly released incremental knowledge is acquired, it is added as a new node to the preset knowledge graph. The incremental knowledge includes newly released standards and specifications and operating parameters of newly connected devices. Based on the knowledge graph updated by the new nodes, the pre-set base model is pre-trained to obtain the first scheduling model with industry standards internalized in the parameters. The first scheduling model is subjected to supervised fine-tuning and reinforcement learning iterative training based on manually labeled scheduling samples to obtain the fully trained second scheduling model. Extract a knowledge subgraph that matches the target scheduling scenario from the knowledge graph updated by the new node, and perform parameter transfer on the fully trained second scheduling model based on the knowledge subgraph to obtain the final scheduling model adapted to the target scheduling scenario. Based on the final scheduling model, scheduling reasoning is performed on the input real-time operating data of the microgrid, and candidate scheduling schemes verified by hard constraints and soft constraints are used as the basis for microgrid scheduling.

2. The method according to claim 1, characterized in that, When newly released incremental knowledge is acquired, it is added as a new node to the pre-defined knowledge graph, including: In response to obtaining newly released standard documents or new device parameters, constraint rules are extracted from the newly released standard documents and operating parameters of new access devices are extracted from the new device parameters, and combined to obtain triples; Based on the preset confidence threshold and the confidence score of the obtained triples, the triples are classified and stored in different levels of the preset knowledge graph. Generate hierarchical map snapshots for later use in tracking changes.

3. The method according to claim 1, characterized in that, Based on the knowledge graph updated by the new nodes, the pre-trained base model is performed to obtain a first scheduling model whose parameters internalize industry standards, including: The triples in the knowledge graph updated by the new nodes are transformed into natural language descriptions and mixed with a preset general power corpus to obtain a mixed corpus. The triples include industry constraint rules. Based on the hybrid corpus, a pre-trained base model with knowledge augmentation is performed to obtain a first scheduling model whose model parameters incorporate industry rules.

4. The method according to claim 1, characterized in that, From the updated knowledge graph of the new nodes, a knowledge subgraph matching the target scheduling scenario is extracted. Based on the knowledge subgraph, the parameters of the fully trained second scheduling model are transferred to obtain a final scheduling model adapted to the target scheduling scenario, including: Extract a knowledge subgraph that matches the target scheduling scenario from the updated knowledge graph of the new node; Device features, topology features, and load features are extracted from the knowledge subgraph and input into a preset graph attention coding network to generate the corresponding scheduling scenario description vector. Freeze the main parameters of the second scheduling model, and train the scene encoder, scene private parameters and scheduling style parameters in the second scheduling model based on the scheduling scene description vector to obtain the final scheduling model that follows the scheduling scene. The final scheduling model is adapted to the target scheduling scene.

5. The method according to claim 1, characterized in that, The scheduling samples are in the form of sample triplets, which include historical scheduling conditions, scheduling instructions, and historical constraints. Based on manually labeled scheduling samples, the first scheduling model is subjected to supervised fine-tuning and reinforcement learning iterative training to obtain a fully trained second scheduling model, including: The pre-trained first scheduling model is trained under supervision based on the scheduling samples, and basic scheduling capabilities are learned from the historical scheduling conditions of the sample triplet. The supervised training model is further reinforced based on the composite reward, and the economic scheduling capability under compliance hard constraints is learned from the composite reward to obtain the fully trained second scheduling model. The composite reward includes compliance reward, security margin reward and economic reward.

6. The method according to claim 1, characterized in that, The knowledge graph consists of a device layer, a topology layer, an operation layer, and a rule layer. Before adding newly released incremental knowledge as a new node to the preset knowledge graph, it includes: The device layer of the knowledge graph is built based on physical entities and their attribute vectors. The topology layer of the knowledge graph is established based on the device relationships and device parameters between devices. The device relationships include electrical connection relationships, feeder affiliation relationships, and location inclusion relationships. The knowledge graph's operational layer is built by mining time-series patterns, extracting typical operating condition patterns and coping strategies from preset historical operating data. The rule layer of the knowledge graph is obtained by encoding the preset constraint rules.

7. The method according to claim 1, characterized in that, Based on the final scheduling model, scheduling inference is performed on the input real-time operating data of the microgrid. Candidate scheduling schemes verified through hard and soft constraints are used as the basis for microgrid scheduling, including: The real-time operating data of the microgrid is input into the final scheduling model. The real-time operating data of the microgrid includes at least one of the following: voltage data, line current, active or reactive power, irradiance, temperature, and frequency. Based on the final scheduling model, rule matching and retrieval are performed on the real-time operating data of the microgrid to obtain the applicable hard constraint rules and soft constraint rules corresponding to the operating conditions. Candidate scheduling schemes are generated using an autoregressive approach; The candidate scheduling scheme is subjected to compliance verification. If it does not violate the hard constraint rules and the degree of deviation from the soft constraint rules is within the preset deviation threshold range, the verification is passed. The microgrid scheduling will be based on candidate scheduling schemes that have been verified by both hard and soft constraints.

8. A microgrid dispatching device based on a large-scale dispatching model, characterized in that, The device includes: The knowledge graph update unit is used to add newly released incremental knowledge as a new node to the preset knowledge graph when it is acquired. The incremental knowledge includes newly released standard specifications and operating parameters of newly accessed devices. The pre-training unit is used to pre-train the pre-set base model based on the knowledge graph updated by the new node, so as to obtain the first scheduling model with industry standards internalized in the parameters. The reinforcement training unit is used to perform supervised fine-tuning and reinforcement learning iterative training on the first scheduling model based on manually labeled scheduling samples to obtain the fully trained second scheduling model. The scenario migration unit is used to extract a knowledge subgraph that matches the target scheduling scenario from the knowledge graph updated by the new node, and perform parameter migration on the fully trained second scheduling model based on the knowledge subgraph to obtain a final scheduling model that adapts to the target scheduling scenario. The scheduling reasoning unit is used to perform scheduling reasoning on the input real-time operating data of the microgrid based on the final scheduling model, and to perform microgrid scheduling based on the candidate scheduling schemes verified by hard constraints and soft constraints.

9. A microgrid scheduling computer device based on a large-scale scheduling model, characterized in that, The system includes a memory, a processor, and a microgrid scheduling program based on a large scheduling model stored in the memory and executable on the processor. The processor executes the microgrid scheduling program based on the large scheduling model to implement the steps of the microgrid scheduling method based on a large scheduling model as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a program that implements a microgrid scheduling method based on a large scheduling model. The program that implements the microgrid scheduling method based on a large scheduling model is executed by a processor to implement the steps of the microgrid scheduling method based on a large scheduling model as described in any one of claims 1 to 7.

Images