Power grid energy scheduling optimization method based on knowledge graph

Abstract

The invention discloses a power grid energy scheduling optimization method based on a knowledge graph, and the method comprises the steps: collecting multi-source heterogeneous data, and carrying out the preprocessing of the data, and obtaining a standardized power grid data set; constructing a power grid knowledge graph to obtain a scheduling knowledge set; extracting structured features to obtain input sequence data; constructing an improved liquid state time constant network to obtain a rolling prediction parameter set; generating solution input items, and summarizing to obtain scheduling optimization solution input; carrying out iterative search by adopting a dandelion optimization algorithm to obtain a scheduling optimization result; and outputting the power grid energy scheduling scheme, and performing scheduling execution. According to the method, the feasibility, the stability and the economical efficiency of power grid energy scheduling are realized by improving rolling prediction of the liquid time constant network and combining dandelion optimization iterative solution.

Description

Technical Field

[0001] This invention relates to the field of power dispatching technology, and in particular to a power grid energy dispatching optimization method based on knowledge graphs. Background Technology

[0002] Existing power grid energy dispatch typically involves multi-object collaborative decision-making, requiring the integration of heterogeneous information from multiple sources, including measurement data, topology connection data, maintenance plan data, and dispatching procedure data, within the same dispatching cycle, and dynamic updates and verifications as operating conditions change. Current technologies often employ relational databases, scattered model files, text-based procedures, or manually generated rule tables for management. Inconsistent data field definitions and inconsistent identification systems make it difficult to fully express the relationships between equipment, sections, regions, procedure clauses, and operational events. Topology information and procedural experience are often stored offline, failing to form a unified mapping with real-time measurements. This makes it difficult to distill dispatching experience and procedural rules into reusable, queryable, and reasonable structured knowledge, hindering cross-system sharing and automated invocation. Consequently, dispatching knowledge struggles to automatically participate in decision-making and respond promptly to changes in operating conditions.

[0003] Existing scheduling optimization methods typically rely on fixed-form prediction parameters and static boundary settings. When constructing scheduling models, they often simplify equipment availability, cross-sectional capacity, maintenance impacts, and procedural triggering conditions, making it difficult to explicitly incorporate implicit coupling relationships into the solution process. When topology switching, maintenance execution, load fluctuations, or alarm events occur, parameter boundaries are easily distorted, leading to infeasibility, slow convergence, or unstable solution quality in the optimization model. Traditional discrete-time prediction models are insufficient in characterizing continuous-time dynamic processes and abrupt event changes. In scenarios with irregular sampling, missing measurement points, or sudden changes in operating conditions, prediction biases tend to accumulate, increasing the deviation between rolling prediction parameters and actual operation, further affecting the executability, stability, and economy of scheduling optimization results.

[0004] Therefore, how to provide a knowledge graph-based method for optimizing power grid energy dispatch is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based method for optimizing power grid energy dispatch. This invention comprehensively utilizes multi-source heterogeneous data fusion from the power grid, knowledge graph construction and graph query reasoning, improved continuous-time prediction using liquid time constant networks, and iterative solution techniques using the dandelion optimization algorithm. This forms a complete process from power grid data acquisition and preprocessing, knowledge graph generation and dispatch knowledge extraction, construction of time-series prediction input sequences, rolling prediction parameter output, dispatch optimization solution, to energy dispatch plan generation and execution. The improved liquid time constant network enhances its ability to characterize abrupt changes in operating conditions and topology variations through state-inhibited liquid evolution functions, coupled evolution under dynamically connected topologies, and continuous-time jump units. The dandelion optimization algorithm improves the stability and convergence efficiency of the solution through discontinuous drift direction control and fault-based drift step updates. Compared with existing technologies, this invention achieves a unified structured expression and reuse of dispatch knowledge, reduces the infeasibility and instability of dispatch optimization, and has the advantages of high feasibility, strong adaptability to changes in operating conditions, and ease of engineering deployment.

[0006] A knowledge graph-based power grid energy dispatch optimization method according to an embodiment of the present invention includes: Collect multi-source heterogeneous data on power grid operation and preprocess it to form a unified field and identifier system, thus obtaining a standardized power grid dataset. A power grid knowledge graph is constructed based on a standardized power grid dataset. Graph query and reasoning are performed on the power grid knowledge graph to obtain a scheduling knowledge set. By extracting structured features from the scheduling knowledge set, input with scene labels is generated to obtain the input sequence data for time series prediction; An improved liquid time constant network is constructed, and a state-suppressed liquid evolution function is introduced to suppress the network state update process corresponding to the input sequence data. Time-varying update processing is carried out through coupled evolution under dynamic connection topology, and jump update processing is carried out based on continuous time jump unit to obtain the rolling prediction parameter set. The input terms for solving the scheduling optimization problem are generated based on the scheduling knowledge set and the rolling prediction parameter set, and then summarized to obtain the scheduling optimization solution input. The dandelion optimization algorithm is used to iteratively search the input of the scheduling optimization solution. Discontinuous drift direction control is performed to reverse the drift update direction of the candidate scheduling solution. The drift step size of the candidate scheduling solution is switched to the step size of the corresponding interval through the drift step size discontinuous update, and the scheduling optimization result is obtained. The scheduling optimization results are output as a power grid energy scheduling scheme and executed to obtain the power grid energy scheduling plan.

[0007] Optionally, the power grid operation data includes multi-source heterogeneous data, such as power grid measurement data, topology connection data, maintenance plan data, and dispatching procedure data.

[0008] Optionally, obtaining the standardized power grid dataset includes: Acquire power grid measurement data, topology connection data, maintenance plan data, and dispatching procedure data to obtain the raw data set; The power grid measurement data, topology connection data, and maintenance plan data are identified, unified, and time-series processed to form measurement time series, topology status series, and maintenance status series, respectively, resulting in a time-series data set. The scheduling procedure data is parsed to generate a set of rule entries. The time series data set and the rule entry set are standardized to form a unified field and identification system, resulting in a standardized power grid dataset.

[0009] Optionally, obtaining the scheduling knowledge set includes: Based on a standardized power grid dataset, define the entity type set, relation type set, and attribute field set of the power grid knowledge graph, and establish the schema layer of the power grid knowledge graph; Based on the schema layer, entities, relationships, and attribute values ​​are extracted from the standardized power grid dataset and written into the power grid knowledge graph to form the power grid knowledge graph. The power grid knowledge graph is used to perform graph query and reasoning to generate a scheduling knowledge set, which includes equipment operating status information, topology connectivity information, maintenance impact information, and scheduling procedure rule triggering information.

[0010] Optionally, the input sequence data for obtaining time series prediction includes: The structured information set is obtained by selecting equipment operation status information, topology connectivity status information, maintenance impact information, and scheduling procedure rule triggering information corresponding to the prediction object from the scheduling knowledge set; The structured information set is mapped into numerical feature vectors according to preset fields, and a corresponding scene label is generated for each timestamp to obtain a structured feature sequence. The structured feature sequence and the measurement time series are merged by timestamp, and the input window is truncated from the continuous timestamps according to the input window length to obtain the input sequence data for time series prediction.

[0011] Optionally, obtaining the rolling prediction parameter set includes: An improved liquid time constant network is constructed, including a state-suppressed liquid evolution function, a liquid coupling evolution structure under a dynamically connected topology, and a continuous-time jump unit; The input sequence data is input into the improved liquid time constant network in chronological order. The liquid state nodes are continuously updated based on the state-suppression liquid evolution function. The state-suppression liquid evolution function applies a first-order suppression term and a third-order suppression term to the liquid state and updates the liquid state together with the input driving term, resulting in the state-suppression updated liquid state sequence. During the continuous-time state update process, the connection relationship between nodes in the liquid state is updated based on the liquid coupling evolution under the dynamic connection topology. The node connectivity relationship is generated and the coupling relationship between nodes is updated according to the topological state in the input sequence data. The coupling relationship of the liquid state nodes changes with the topological state, resulting in a topology-adaptive liquid state sequence. The liquid state is updated by continuous time jump unit. The jump time is determined according to the event trigger information in the input sequence data. At the jump time, the liquid state is superimposed with the jump variable obtained by input and the continuous time state update continues. The final liquid state sequence is input to the output layer to obtain the rolling prediction parameter set. The rolling prediction parameter set includes load prediction sequence, renewable power output prediction sequence, available capacity prediction sequence and risk indicator prediction sequence.

[0012] Optionally, the summarization yields the scheduling optimization solution input, including: The scheduling object set is determined by the scheduling knowledge set and the corresponding scheduling decision variable set is generated. The scheduling object set includes generator sets, renewable energy output units, energy storage units and tie-line switching units. Based on the scheduling knowledge set, a set of scheduling conditions and a set of scheduling decision variable boundaries are generated. The set of scheduling conditions includes equipment operating status conditions, topology connectivity conditions, maintenance impact conditions, and scheduling procedure rule triggering conditions. The set of scheduling decision variable boundaries includes upper limit boundaries, lower limit boundaries, and availability boundaries, thus obtaining condition and boundary input items. Based on the rolling forecast parameter set, time period forecast parameter input items are generated, and scheduling target input items including operating cost items, wind and solar curtailment penalty items, and risk indicator penalty items are generated. The scheduling decision variable set, condition and boundary input items, time period forecast parameter input items, and scheduling target input items are summarized to obtain the scheduling optimization solution input.

[0013] Optionally, obtaining the scheduling optimization result includes: A population of the dandelion optimization algorithm is constructed, which consists of multiple candidate scheduling solutions. Each candidate scheduling solution is composed of the values ​​of the scheduling decision variable set in each time period. The fitness value is calculated for each candidate scheduling solution. The fitness value is jointly determined by the operating cost item, the wind curtailment penalty item, and the risk indicator penalty item to obtain the initial optimal candidate scheduling solution. In each iteration, a drift update is performed on each candidate scheduling solution to generate an updated candidate scheduling solution. After generating the updated candidate scheduling solution, the fitness value is calculated and the optimal candidate scheduling solution is updated to obtain the iterative optimal candidate scheduling solution. Perform discontinuous drift direction control processing. If the fitness value of the optimal candidate scheduling solution does not improve within a preset number of consecutive iterations, switch the drift update direction of the candidate scheduling solution from the direction toward the optimal candidate scheduling solution to the direction away from the optimal candidate scheduling solution, and generate a new candidate scheduling solution based on the switched drift update direction. Perform a drift step size fault-based update process, dividing the iteration process into a first iteration interval, a second iteration interval, and a third iteration interval. The first drift step size is used in the first iteration interval, the second drift step size is used in the second iteration interval, and the third drift step size is used in the third iteration interval. The first drift step size is greater than the second drift step size, and the second drift step size is greater than the third drift step size. In each iteration interval, the drift update of the candidate scheduling solution is completed according to the corresponding drift step size. When the termination condition is met, the iterative optimal candidate scheduling solution is output as the scheduling optimization result. The termination condition includes reaching a preset number of iterations or the fitness value of the iterative optimal candidate scheduling solution does not exceed a preset threshold within a consecutive preset number of iterations.

[0014] Optionally, obtaining the power grid energy dispatch plan includes: The scheduling optimization results are converted into a grid energy dispatch plan, which includes generator output plans, renewable energy output arrangements, energy storage charging and discharging plans, reserve capacity plans, and tie-line switching plans for each time period. Based on the power grid energy dispatch plan, a dispatch execution instruction set is generated. The dispatch execution instruction set includes output setpoint instructions, charging and discharging setpoint instructions, standby setpoint instructions, and tie-line exchange setpoint instructions organized by object identifier and time period identifier. The scheduling execution instruction set is sent to the scheduling execution platform to complete the scheduling execution, and the scheduling execution result data is collected. The scheduling execution result data includes the actual output time series, the actual power flow time series, the energy storage state of charge time series, and the wind and solar curtailment time series, thus obtaining the grid energy scheduling execution result.

[0015] The beneficial effects of this invention are: This invention unifies the preprocessing of power grid measurement data, topology connectivity data, maintenance plan data, and dispatching procedure data to form a unified field and identifier system. It then constructs a power grid knowledge graph and performs graph querying and reasoning, outputting a dispatching knowledge set containing information on equipment operating status, topology connectivity status, maintenance impact, and procedure rule triggering. This allows information previously scattered across ledgers, topology models, procedure clauses, and event records to be expressed and consistently referenced within a single semantic framework. Compared to existing processing methods that rely primarily on scattered tables, textual procedures, and manual experience, this invention improves the structuring and reusability of dispatching knowledge. It enables dispatching rules and operational information to be invoked by the dispatching process in a computable form, thereby enhancing the consistency and automation of dispatching decisions and reducing the manual verification costs caused by inconsistent data definitions and difficulties in linking topology information.

[0016] This invention introduces an improved liquid time constant network and a dandelion optimization algorithm in the rolling prediction and optimization solution stages. The improved liquid time constant network suppresses state update fluctuations through a state-suppressing liquid evolution function, follows topological state changes through coupled evolution under dynamic connection topology, and handles state jumps triggered by events through continuous time jump units. It outputs a rolling prediction parameter set of load prediction sequence, renewable power output prediction sequence, available capacity prediction sequence, and risk index prediction sequence, thereby reducing the accumulation of prediction bias caused by the easy distortion of parameter boundaries. The dandelion optimization algorithm controls the drift direction of the candidate scheduling solution in the iterative search by using discontinuous drift direction control to reduce getting trapped in local optima. It switches the drift step size in different iteration intervals through discontinuous update to balance search efficiency and convergence accuracy, resulting in more stable scheduling optimization results. This improves the feasibility, convergence efficiency, and execution stability of the power grid energy dispatch plan. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0018] Figure 1 This is a flowchart of a knowledge graph-based power grid energy dispatch optimization method proposed in this invention; Figure 2 This is a block diagram of the improved liquid time constant network for a knowledge graph-based power grid energy dispatch optimization method proposed in this invention. Figure 3 This is a functional diagram of the Dandelion optimization algorithm, which is a knowledge graph-based power grid energy dispatch optimization method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A knowledge graph-based method for optimizing power grid energy dispatch includes: Collect multi-source heterogeneous data on power grid operation and preprocess it to form a unified field and identifier system, thus obtaining a standardized power grid dataset. A power grid knowledge graph is constructed based on a standardized power grid dataset. Graph query and reasoning are performed on the power grid knowledge graph to obtain a scheduling knowledge set. By extracting structured features from the scheduling knowledge set, input with scene labels is generated to obtain the input sequence data for time series prediction; An improved liquid time constant network is constructed, and a state-suppressed liquid evolution function is introduced to suppress the network state update process corresponding to the input sequence data. Time-varying update processing is carried out through coupled evolution under dynamic connection topology, and jump update processing is carried out based on continuous time jump unit to obtain the rolling prediction parameter set. The input terms for solving the scheduling optimization problem are generated based on the scheduling knowledge set and the rolling prediction parameter set, and then summarized to obtain the scheduling optimization solution input. The dandelion optimization algorithm is used to iteratively search the input of the scheduling optimization solution. Discontinuous drift direction control is performed to reverse the drift update direction of the candidate scheduling solution. The drift step size of the candidate scheduling solution is switched to the step size of the corresponding interval through the drift step size discontinuous update, and the scheduling optimization result is obtained. The scheduling optimization results are output as a power grid energy scheduling scheme and executed to obtain the power grid energy scheduling plan.

[0021] In this embodiment, the multi-source heterogeneous data of power grid operation includes power grid measurement data, topology connection data, maintenance plan data, and dispatching procedure data.

[0022] In this embodiment, obtaining the standardized power grid dataset includes: Acquire power grid measurement data, topology connection data, maintenance plan data, and dispatching procedure data to obtain the raw data set; The power grid measurement data, topology connection data, and maintenance plan data are identified, unified, and time-series processed to form measurement time series, topology status series, and maintenance status series, respectively, resulting in a time-series data set. The dispatching procedure data is parsed to generate a set of rule entries. The time-series data set and the rule entry set are then standardized to form a unified field and identifier system, resulting in a standardized power grid dataset. The specific steps of parsing the clauses are as follows: The scheduling procedure data is structured and broken down into chapters, sections, and clauses. A unique clause identifier is generated for each clause. The clause text is segmented and elements are extracted to identify the applicable object identifier, applicable time period identifier, triggering condition element, limiting condition element, and execution action element corresponding to the clause. The triggering condition element is converted into a condition item consisting of field name, comparison relationship, and threshold. The limiting condition element is converted into a limiting item consisting of variable name, upper and lower limits, and unit. The execution action element is converted into an action item consisting of action type, target object, and target value. The condition item, limiting item, and action item of each clause are encapsulated into rule entries and written into the rule entry set. At the same time, the object identifier, time identifier, variable name, and unit are uniformly mapped to complete the consistent field and identifier correspondence between the rule entry set and the time series data set.

[0023] In this embodiment, obtaining the scheduling knowledge set includes: Based on a standardized power grid dataset, the entity type set, relation type set, and attribute field set of the power grid knowledge graph are defined, and a schema layer of the power grid knowledge graph is established, wherein: The definition of the entity type set, relation type set, and attribute field set of the power grid knowledge graph is as follows: Based on the unified object identifier and field caliber of the standardized power grid dataset, business objects are classified and abstracted into entity types. A primary key identifier field is determined for each entity type. Object associations that can be directly mapped in the data are abstracted into relation types. A starting entity type, an ending entity type, and a relation direction are determined for each relation type. Fields describing object characteristics and status are merged into an attribute field set. The data type, unit, and update granularity are determined for each attribute field, forming the entity type set, relation type set, and attribute field set. The establishment of the schema layer for the power grid knowledge graph specifically involves: defining node types for the schema layer based on a set of entity types; configuring a primary key field and a set of attribute fields for each node type; defining edge types for the schema layer based on a set of relation types; configuring a start point type, end point type, direction identifier, and a set of relation attribute fields for each edge type; setting uniqueness rules for the primary key field; setting field validation rules for the attribute field; and registering the node type definitions and edge type definitions as the schema layer of the knowledge graph to complete the establishment of the schema layer. Based on the schema layer, entities, relationships, and attribute values ​​are extracted from the standardized power grid dataset and written into the power grid knowledge graph to form the power grid knowledge graph. The execution graph query and reasoning of the power grid knowledge graph generates a scheduling knowledge set, which includes equipment operating status information, topology connectivity information, maintenance impact information, and scheduling procedure rule triggering information. Specifically, the execution graph query and reasoning involves: Using scheduling time period, area identifier, and equipment identifier as query conditions, the system retrieves equipment entities and attributes, connection relationship edges and attributes, maintenance event entities and time interval attributes, and procedure clause entities and condition and restriction items attributes from the power grid knowledge graph to obtain a subgraph corresponding to the scheduling time period. Based on the effective state of the connection relationship edges, the system performs connectivity traversal on the equipment entities to generate topology connectivity state information. Through the association between maintenance event entities and equipment entities and the maintenance time interval, the system determines the set of unavailable equipment and the set of affected relationship edges within the scheduling time period to generate maintenance impact information. The triggering conditions in the procedure clause entity are matched one by one with the measurement time series, equipment status attributes, topology status attributes and maintenance status attributes. When the triggering conditions are met, a rule triggering record is generated and associated with the corresponding equipment entity and time period identifier. This yields equipment operating status information, topology connectivity status information, maintenance impact information and scheduling procedure rule triggering information, which are then summarized to form a scheduling knowledge set.

[0024] In this embodiment, obtaining the input sequence data for time series prediction includes: From the scheduling knowledge set, equipment operating status information, topology connectivity information, maintenance impact information, and scheduling procedure rule triggering information corresponding to the prediction object are selected to obtain a structured information set. Specifically, the obtained structured information set includes: The system determines the object identifier and the region identifier of the predicted object, retrieves the corresponding equipment operation status information entries in the scheduling knowledge set based on the object identifier, retrieves the topology connectivity status information entries associated with the predicted object in the scheduling knowledge set based on the region identifier, and filters out entries that have a connection relationship or a same section relationship with the predicted object. The system also retrieves maintenance impact information entries in the scheduling knowledge set based on the object identifier and extracts the availability marker and impact range marker within the prediction period. Based on the object identifier and the predicted time period, retrieve the scheduling procedure rule trigger information entries from the scheduling knowledge set and extract the trigger clause identifier and trigger type marker. Align all entries according to a unified set of fields and organize them into a record set in timestamp order to obtain a structured information set. The structured information set is mapped into numerical feature vectors according to preset fields, and a corresponding scene label is generated for each timestamp to obtain a structured feature sequence. The mapping into numerical feature vectors according to preset fields specifically involves: Define a list of feature fields and their order. The list of feature fields includes continuous fields and discrete fields. For continuous fields, write the values ​​directly into the corresponding dimensions and perform unit consistency. For discrete fields, encode the values ​​and the encoding table and write them into the corresponding dimensions. The encoding table assigns a unique integer code to each discrete value. For Boolean fields, write the value 1 if the value is true and write the value 0 if the value is false. For missing fields, write the default value. Concatenate the values ​​of the corresponding dimensions of each field under the same timestamp in the field order to form the numerical feature vector of that timestamp, and obtain the structured feature sequence. The structured feature sequence and the measurement time series are merged by timestamp, and an input window is truncated from the continuous timestamps according to the input window length to obtain the input sequence data for time series prediction. The input window length is specifically: A fixed window based on the prediction step size contains a preset number of consecutive timestamp records. When the prediction step size is five minutes, the input window length is set to contain twenty-four consecutive timestamp records. When the prediction step size is fifteen minutes, the input window length is set to contain thirty-two consecutive timestamp records. When the prediction step size is sixty minutes, the input window length is set to contain twenty-four consecutive timestamp records.

[0025] In this embodiment, obtaining the rolling prediction parameter set includes: An improved liquid time constant network is constructed, comprising a state-inhibited liquid evolution function, a liquid coupling evolution structure under a dynamically connected topology, and continuous-time jump units, wherein: Multiple liquid state nodes are set up between the input and output of the improved liquid time constant network and the connection between the nodes is established. A time constant and a continuous time state update rule are configured for each liquid state node. The state-suppressed liquid evolution function is embedded into the state decay part of the continuous time state update rule so that each state update includes a suppression term and a third suppression term to suppress the excessive fluctuation of the state under abrupt input. The connection relationship between nodes is set to be able to be updated with the change of topology state, so that the liquid coupling evolution structure under dynamic connection topology can be used to generate node connectivity and update the coupling relationship between nodes according to the topology state at different time points. A continuous time jump unit is set in the continuous time state update path, so that the continuous time jump unit is used to superimpose the jump variable obtained by input to the liquid state at the event trigger time, and continue to perform continuous time state update after the jump, thus completing the construction of the improved liquid time constant network. The input sequence data is fed into the improved liquid time constant network in chronological order. The liquid state nodes are continuously updated based on a state-suppressed liquid evolution function. This function simultaneously applies a first-order and a third-order suppression term to the liquid state, which, together with the input driving term, updates the liquid state, resulting in a state-suppressed updated liquid state sequence. The state-suppressed liquid evolution function is specifically defined as follows: the state-suppressed liquid evolution function is a function term used for state decay in the continuous-time state update of liquid state nodes. The function term consists of a first-order suppression term and a third-order suppression term. The first-order suppression term is the product of the suppression coefficient and the liquid state, and the third-order suppression term is the product of the suppression coefficient and the cube of the liquid state. The continuous-time change of the liquid state is jointly determined by the suppression term and the input driving term. The input driving term is formed by combining the driving value obtained by the input sequence data through input transformation and the coupling value obtained by the connection relationship between liquid state nodes. The process of obtaining the state-suppressed updated liquid state sequence is as follows: Adjacent timestamp records of the input sequence data are read sequentially by time. The interval between adjacent timestamps is used as the time step for continuous time updates. The liquid state within each time step is iteratively updated. First, a first-order suppression term and a third-order suppression term are calculated based on the current liquid state and synthesized into a state suppression value. Then, an input driving value is calculated based on the current input record, and a coupling value is calculated based on the connection relationship between nodes. The input driving value, coupling value, and state suppression value are used together to update the liquid state corresponding to the next timestamp. This process is repeated for all timestamps, and the liquid state corresponding to each timestamp is summarized in timetamp order to obtain the state-suppressed updated liquid state sequence. During continuous-time state updates, the inter-node connectivity relationships in the liquid state are updated based on liquid coupling evolution under a dynamic connection topology. Node connectivity relationships are generated and inter-node coupling relationships are updated according to the topological state in the input sequence data. The coupling relationships of the liquid state nodes change with the topological state, resulting in a topology-adaptive liquid state sequence. Specifically, the update process for inter-node connectivity relationships in the liquid state based on liquid coupling evolution under a dynamic connection topology is as follows: During the initialization of the improved liquid time constant network, a basic connection table between nodes is established, and a basic connection strength is configured for each pair of connectable nodes. At each time point of continuous time state update, the topology state of the corresponding time point is read from the input sequence data, and a node connectivity table for the current time point is generated based on the topology connection data. The node connectivity table is used to characterize whether any two nodes are connected at the current time point. The basic connection table and the node connectivity table are matched and updated according to the corresponding node pairs. When the node connectivity table indicates connectivity, the basic connection strength of the current node pair is retained. When the node connectivity table indicates non-connectivity, the connection strength of the node pair is set to zero. The updated connection strength set is used to calculate the coupling value between nodes and participate in the liquid state update at time points. The liquid state at each time point is repeatedly updated and summarized in chronological order to obtain the topology adaptive liquid state sequence. The liquid state is updated using a continuous-time jump unit. The jump time is determined based on event triggering information in the input sequence data. At the jump time, the liquid state is superimposed with the jump variable obtained from the input drive, and continuous-time state updates continue. The final liquid state sequence is input to the output layer to obtain a rolling prediction parameter set. The rolling prediction parameter set includes a load prediction sequence, a renewable energy output prediction sequence, an available capacity prediction sequence, and a risk indicator prediction sequence, among which: The method of determining the jump time according to the event trigger information in the input sequence data is as follows: when generating the input sequence data, an event trigger field is configured for each timestamp record. The event trigger field has a value of 1, which means that the event corresponding to the timestamp has occurred, and a value of 0 means that no event has occurred at the timestamp. During the continuous time state update process, the event trigger field of each timestamp record is read in chronological order. When the event trigger field has a value of 1, the timestamp is determined as the jump time and marked as the jump update point. When the event trigger field has a value of 1 in multiple adjacent timestamps, the timestamp with the first value of 1 is determined as the jump time and a jump update is performed at the jump time. The step of inputting the final liquid state sequence into the output layer to obtain the rolling prediction parameter set is as follows: extract the liquid state vector corresponding to each timestamp in the final liquid state sequence in chronological order, and convert the liquid state vector into a prediction parameter vector through output mapping. The output mapping includes the product of the state vector and the output parameter matrix and the superposition of the output bias. The prediction parameter vector of each timestamp is split into load prediction value, renewable power output prediction value, available capacity prediction value and risk index prediction value according to the field position, and summarized in chronological order to form load prediction sequence, renewable power output prediction sequence, available capacity prediction sequence and risk index prediction sequence, thereby obtaining the rolling prediction parameter set.

[0026] In this embodiment, the summarization of the scheduling optimization solution input includes: The process involves determining a set of scheduling objects using a scheduling knowledge set and generating a corresponding set of scheduling decision variables. The set of scheduling objects includes generator sets, renewable energy output units, energy storage units, and tie-line switching units. Specifically, determining the set of scheduling objects and generating the corresponding set of scheduling decision variables using the scheduling knowledge set involves: The system reads equipment operation status information and topology connectivity status information from the scheduling knowledge set, filters out object identifiers that are available and connected to the scheduling area during the scheduling period, and classifies the object identifiers into generator sets, renewable energy output units, energy storage units and tie-line switching units according to the object type field to form a scheduling object set. For generator sets, the system generates unit output variables and unit standby variables for each object identifier in each scheduling period. For renewable energy output units, the system generates renewable energy output variables and abandoned electricity variables for each object identifier in each scheduling period. For each object identifier of the energy storage unit, charging power variables, discharging power variables, and state of charge variables are generated in each scheduling period. For each object identifier of the tie-line switching unit, switching power variables are generated in each scheduling period. All types of variables are arranged and summarized according to object identifier and time period identifier to obtain a set of scheduling decision variables. Based on the scheduling knowledge set, a set of scheduling conditions and a set of scheduling decision variable boundaries are generated. The set of scheduling conditions includes equipment operating status conditions, topology connectivity conditions, maintenance impact conditions, and scheduling procedure rule triggering conditions. The set of scheduling decision variable boundaries includes upper limit boundaries, lower limit boundaries, and availability boundaries, thus obtaining the condition and boundary input items. Specifically, the generation of the set of scheduling conditions and the set of scheduling decision variable boundaries based on the scheduling knowledge set is as follows: The system reads equipment operation status information from the scheduling knowledge set, extracts start / stop markers, availability markers, and fault markers for each scheduling object in each scheduling period, and forms equipment operation status conditions. It also reads topology connectivity status information from the scheduling knowledge set, extracts connectivity and on / off markers for each scheduling object with the scheduling area, key sections, and connecting lines, and forms topology connectivity status conditions. Finally, it reads maintenance impact information from the scheduling knowledge set, extracts the object identifiers covered by maintenance, the start and end times of the impact, and the derating markers, and forms maintenance impact conditions and corresponding availability boundaries. The system reads the scheduling procedure rule triggering information from the scheduling knowledge set, extracts the trigger clause identifier, trigger time period, trigger object identifier and trigger type mark to form the scheduling procedure rule triggering conditions. At the same time, it reads the capacity attribute, power upper and lower limit attribute, ramp attribute and exchange limit attribute of each object from the scheduling knowledge set, converts the capacity attribute and power upper and lower limit attribute into upper limit boundary and lower limit boundary, converts start-stop mark, availability mark and maintenance derating mark into availability boundary, and summarizes them into condition and boundary input items according to object identifier and time period identifier. Based on the rolling forecast parameter set, time-period forecast parameter input items are generated, and scheduling target input items including operating cost items, wind and solar curtailment penalty items, and risk indicator penalty items are generated. The scheduling decision variable set, condition and boundary input items, time-period forecast parameter input items, and scheduling target input items are summarized to obtain the scheduling optimization solution input. Specifically, the process of generating time-period forecast parameter input items based on the rolling forecast parameter set and generating scheduling target input items including operating cost items, wind and solar curtailment penalty items, and risk indicator penalty items involves: Read the load forecast sequence, renewable output forecast sequence, available capacity forecast sequence and risk indicator forecast sequence from the rolling forecast parameter set according to the scheduling period identifier, write the forecast value of each forecast sequence in each scheduling period into the parameter record that matches the scheduling period and form the period forecast parameter input item, and generate the scheduling target input item at the same time. The operating cost item of the dispatch target input consists of the output variable of each generator unit in each dispatch period and the corresponding cost coefficient. The wind and solar curtailment penalty item consists of the curtailment of each renewable power unit in each dispatch period and the corresponding penalty coefficient. The risk indicator penalty item consists of the predicted value of the risk indicator and the risk penalty coefficient for each dispatch period.

[0027] In this embodiment, obtaining the scheduling optimization result includes: A population for the Dandelion optimization algorithm is constructed, consisting of multiple candidate scheduling solutions. Each candidate scheduling solution is composed of the values ​​of the scheduling decision variables in each time period. A fitness value is calculated for each candidate scheduling solution, determined jointly by the operating cost item, the wind and solar curtailment penalty item, and the risk indicator penalty item, to obtain the initial optimal candidate scheduling solution. The calculation of the fitness value specifically involves: For each candidate scheduling solution, the values ​​of various scheduling decision variables in each scheduling period are read, and the values ​​are substituted into the scheduling target input items for calculation in each period. The operating cost item is calculated according to the output value of each generator unit in each scheduling period and the corresponding cost coefficient to obtain the operating cost of each period and then summed. The wind and solar curtailment penalty item is calculated according to the curtailment value of each renewable power unit in each scheduling period and the corresponding penalty coefficient to obtain the curtailment penalty of each period and then summed. The risk index penalty item is calculated according to the risk index prediction value of each scheduling period and the corresponding risk penalty coefficient to obtain the risk penalty of each period and then summed. The fitness value of the candidate scheduling solution is obtained by adding the sum of the operating cost, the sum of the wind and solar curtailment penalty, and the sum of the risk indicator penalty. In each iteration, a drift update is performed on each candidate scheduling solution to generate an updated candidate scheduling solution. After generating the updated candidate scheduling solution, the fitness value is calculated and the optimal candidate scheduling solution is updated to obtain the iterative optimal candidate scheduling solution. The specific steps for performing the drift update to generate the updated candidate scheduling solution are as follows: At the beginning of each iteration, the optimal candidate scheduling solution for the current iteration is determined, and a position vector is formed for each candidate scheduling solution in the population according to the order of the corresponding variables. For any candidate scheduling solution, the difference vector between the position vector and the position vector of the optimal candidate scheduling solution for the current iteration is calculated. The difference vector is used as the drift direction vector, and the drift direction vector is scaled according to the drift step size used in the current iteration to obtain the drift increment vector. The drift increment vector is added to the position vector of the candidate scheduling solution to obtain the updated position vector. The updated position vector is restored to the value of each scheduling decision variable in each scheduling period according to the variable index to form the updated candidate scheduling solution. Boundary pruning is performed on the updated candidate scheduling solution to ensure that the value of each scheduling decision variable does not exceed the corresponding upper and lower boundary to obtain the updated candidate scheduling solution. A discontinuous drift direction control process is performed. If the fitness value of the optimal candidate scheduling solution does not improve within a preset number of iterations, the drift update direction of the candidate scheduling solution is switched from the direction towards the optimal candidate scheduling solution to the direction away from the optimal candidate scheduling solution. A new candidate scheduling solution is then generated based on the switched drift update direction. Specifically, the discontinuous drift direction control process involves: During the iteration process, a stall counter is set and the historical best fitness value of the iterative optimal candidate scheduling solution is recorded. When the fitness value of the iterative optimal candidate scheduling solution is not less than the historical best fitness value after the iteration ends, the stall counter is incremented by one. When the fitness value of the iterative optimal candidate scheduling solution is less than the historical best fitness value, the stall counter is cleared and the historical best fitness value is updated. When the stall counter reaches the number of consecutive preset iterations, the drift direction vector of each candidate scheduling solution in the population is switched from the difference vector to the opposite vector of the difference vector, so that the drift update changes from moving towards the optimal candidate scheduling solution to moving away from the optimal candidate scheduling solution. Based on the switched drift direction vector and the current iteration drift step size, a new candidate scheduling solution is generated by drift update, and the stall counter is cleared. The number of consecutive preset iterations is set to five. A fault-based update process with a drift step size is performed, dividing the iteration process into a first iteration interval, a second iteration interval, and a third iteration interval. The first iteration interval uses the first drift step size, the second iteration interval uses the second drift step size, and the third iteration interval uses the third drift step size. The first drift step size is greater than the second drift step size, and the second drift step size is greater than the third drift step size. Within each iteration interval, the drift update of the candidate scheduling solution is completed according to the corresponding drift step size, where: The process of dividing the iteration process into a first iteration interval, a second iteration interval, and a third iteration interval is as follows: a preset total number of iterations of one hundred is set as the maximum number of iterations, and the maximum number of iterations is divided into three consecutive intervals in sequence. The first iteration interval is defined as the iteration round corresponding to one-third of the maximum number of iterations from the first iteration round. The second iteration interval is defined as the iteration round corresponding to two-thirds of the maximum number of iterations from one-third plus one iteration round. The third iteration interval is defined as the iteration round corresponding to the maximum number of iterations from two-thirds plus one iteration round. During the iteration process, the corresponding drift step size is selected according to the interval to which the current iteration round belongs for the drift update of the candidate scheduling solution. The first drift step size, the second drift step size, and the third drift step size are set to 0.1, 0.01, and 0.001, respectively; When the termination condition is met, the iterative optimal candidate scheduling solution is output as the scheduling optimization result. The termination condition includes reaching a preset number of iterations or the fitness value of the iterative optimal candidate scheduling solution does not change more than a preset threshold within a consecutive preset number of iterations. The preset number of iterations and the preset threshold are set to one hundred and 0.001, respectively.

[0028] In this embodiment, obtaining the power grid energy dispatch plan includes: The scheduling optimization results are converted into a grid energy dispatch plan, which includes generator output plans for each time period, renewable energy output arrangements, energy storage charging and discharging plans, reserve capacity plans, and tie-line switching plans. Specifically, the conversion of the scheduling optimization results into a grid energy dispatch plan involves: Read the values ​​of scheduling decision variables for each scheduling object in each scheduling period from the scheduling optimization results, and collect them according to the scheduling period. Arrange the values ​​of generator output variables corresponding to generator sets according to the generator set identifier and the scheduling period identifier to form a generator set output plan. Arrange the values ​​of renewable output variables and abandoned power variables corresponding to renewable energy output units according to the unit identifier and the scheduling period identifier to form a renewable energy output arrangement. Arrange the values ​​of charging power variables, discharging power variables and state of charge variables corresponding to energy storage units according to the energy storage identifier and the scheduling period identifier to form an energy storage charging and discharging plan. The standby variable values ​​corresponding to the generator sets are arranged according to the generator set identifier and the dispatch time period identifier to form a standby capacity plan. The switching power variable values ​​corresponding to the tie line switching unit are arranged according to the tie line identifier and the dispatch time period identifier to form a tie line switching plan. All plans are summarized and output in a unified dispatch plan format to obtain the power grid energy dispatch plan. Based on the power grid energy dispatch plan, a dispatch execution instruction set is generated. The dispatch execution instruction set includes output setpoint instructions, charging and discharging setpoint instructions, standby setpoint instructions, and tie-line exchange setpoint instructions organized by object identifier and time period identifier. The scheduling execution instruction set is sent to the scheduling execution platform to complete the scheduling execution, and the scheduling execution result data is collected. The scheduling execution result data includes the actual output time series, the actual power flow time series, the energy storage state of charge time series, and the wind and solar curtailment time series, thus obtaining the grid energy scheduling execution result.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the intraday rolling energy dispatching of a dispatch center, covering the kV main grid and some important load access areas. The dispatching cycle was 15 minutes, and the rolling prediction time domain was 4 hours. The scenario features concentrated grid connection of wind and solar power, frequent power exchange across regional interconnections, and overlapping equipment maintenance and mode adjustments. The raw data comes from four sources: real-time measurements, topology connections, maintenance plans, and dispatching procedures. There are over 15,000 measurement points, over 2,000 topology objects, and maintenance events covering multiple lines and substation bays. After parsing the procedure clauses, over 400 executable rule entries were generated. In this scenario, it is necessary to unify and reuse the procedures, topology, and maintenance impacts, while avoiding the problems of plan inexecutability and repeated recalculations caused by implicit coupling relationships and parameter boundary distortions in traditional optimization.

[0030] When applying this invention in the current scenario, firstly, a unified field and identification system is organized to form a standardized power grid dataset that can be directly stored and associated. Then, a power grid knowledge graph is constructed and a scheduling knowledge set is generated through graph query and reasoning. The scheduling knowledge set can provide the equipment availability status, critical connectivity status, maintenance impact range, and procedure triggering results on the same time base. Subsequently, structured features are extracted from the scheduling knowledge set to generate a time-series input sequence with scene labels. This time-series input sequence is fed into an improved liquid time constant network to obtain a rolling prediction parameter set. A state-suppressive liquid evolution function is used to suppress excessive state fluctuations caused by abrupt inputs. Coupled evolution under dynamic connection topology is used to update the coupling relationship between nodes synchronously with changes in topology state. Continuous time jump units are used to correct state jumps at event triggers and continue continuous time evolution. Finally, the scheduling knowledge set and the rolling prediction parameter set are combined into the scheduling optimization solution input. The dandelion optimization algorithm is used for iterative search. Discontinuous drift direction control is used to reverse the drift direction to break out of stagnation when there is no improvement after several consecutive iterations. Drift step size discontinuous update is used to switch the step size in different iteration intervals to balance exploration and convergence, thereby outputting a grid energy scheduling plan that can be directly issued and completing the scheduling execution.

[0031] The dispatch center records show multiple photovoltaic fluctuations caused by rapid changes in cloud cover, one temporary load increase, and two equipment mode switching events. The procedure was triggered over 300 times, affecting more than 50 objects. After adopting this invention, the rule entries generated by the procedure clause parsing can be automatically invoked in rolling dispatch. The number of times manual checks between procedures and topology are performed has decreased from over ten times per day to single digits, and the number of times dispatchers manually recalculate unexecutable plans has significantly decreased. The operation logs on the dispatch execution side show that after the plan is issued, the equipment setpoint instructions and actual execution records can be aligned and closed within time periods. No instructions were lost during mode switching, and the dispatch plan can still maintain continuous output even under scenarios with overlapping maintenance impacts. This demonstrates the effect of unified expression of multi-source heterogeneous information and improved plan executability under dynamic operating conditions.

[0032] Table 1 Summary of Comparison Indicators for Energy Dispatch Optimization Methods

[0033] As shown in Table 1, in terms of economic efficiency and renewable energy consumption, this invention performs best in both total cost and the two key indicators of wind and solar curtailment. The total cost is 31.2 million yuan, lower than the rule-based method's 34.5 million yuan, and also lower than the traditional MILP's 33 million yuan and various schemes incorporating knowledge graphs or predictions. The wind and solar curtailment is 410 MWh, significantly lower than the rule-based method's 680 MWh, and a substantial improvement compared to MILP's 520 MWh. It is also superior to KG+MILP's 470 MWh, KG+LSTM's 455 MWh, and KG+DO's 448 MWh, indicating that within the same dispatch window, the plan output by this invention is more closely aligned with changes in renewable energy output and improves the level of consumption.

[0034] From the perspectives of feasibility and safety, this invention is more stable in terms of the number of limit violations, infeasibility rate, and number of insufficient reserves. The number of limit violations is 1, on par with KG+MILP, KG+LSTM, and KG+DO, and significantly better than the rule-based method (5 violations) and MILP (2 violations). The infeasibility rate is 0.8%, significantly lower than the rule-based method (6.5%), MILP (4.2%), and also lower than KG+MILP (2.1%), KG+LSTM (1.8%), and KG+DO (1.6%), demonstrating higher executability under complex conditions. The number of insufficient reserves is 0, while the rule-based method has 3, MILP has 2, and the other three comparison methods all have 1. This invention is more robust in terms of reserve configuration, which is beneficial for coping with load fluctuations and changes in equipment status.

[0035] In terms of computational efficiency and tracking performance, this invention exhibits a more balanced overall advantage in terms of solution time, prediction error, and plan tracking deviation. The solution time is 7.4 seconds, significantly faster than MILP's 18.6 seconds and KG+MILP and KG+LSTM's approximately 20 seconds, and also better than KG+DO's 9.8 seconds, only 1.2 seconds slower than the rule-based method. The prediction error is 2.9%, a significant decrease compared to the rule-based method's 6.8% and MILP's 5.9%, and also better than KG+LSTM and KG+DO's 3.5%. The plan tracking deviation is 42MW, significantly less than the rule-based method's 95MW and MILP's 70MW, and better than KG+MILP's 58MW, KG+LSTM's 50MW, and KG+DO's 48MW. This indicates that the scheduling plan output by this invention is more closely aligned with actual execution, and the linkage between rolling prediction and optimization enhances the traceability on the execution side.

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

Claims

1. A knowledge graph-based method for optimizing power grid energy dispatch, characterized in that, include: Collect multi-source heterogeneous data on power grid operation and preprocess it to form a unified field and identifier system, thus obtaining a standardized power grid dataset. A power grid knowledge graph is constructed based on a standardized power grid dataset. Graph query and reasoning are performed on the power grid knowledge graph to obtain a scheduling knowledge set. By extracting structured features from the scheduling knowledge set, input with scene labels is generated to obtain the input sequence data for time series prediction; An improved liquid time constant network is constructed, and a state-suppressed liquid evolution function is introduced to suppress the network state update process corresponding to the input sequence data. Time-varying update processing is carried out through coupled evolution under dynamic connection topology, and jump update processing is carried out based on continuous time jump unit to obtain the rolling prediction parameter set. The input terms for solving the scheduling optimization problem are generated based on the scheduling knowledge set and the rolling prediction parameter set, and then summarized to obtain the scheduling optimization solution input. The dandelion optimization algorithm is used to iteratively search the input of the scheduling optimization solution. Discontinuous drift direction control is performed to reverse the drift update direction of the candidate scheduling solution. The drift step size of the candidate scheduling solution is switched to the step size of the corresponding interval through the drift step size discontinuous update, and the scheduling optimization result is obtained. The scheduling optimization results are output as a power grid energy scheduling scheme and executed to obtain the power grid energy scheduling plan.

2. The knowledge graph-based power grid energy dispatch optimization method according to claim 1, characterized in that, The power grid operation data consists of multi-source heterogeneous data, including power grid measurement data, topology connection data, maintenance plan data, and dispatching procedure data.

3. The power grid energy dispatch optimization method based on knowledge graphs according to claim 1, characterized in that, The obtained standardized power grid dataset includes: Acquire power grid measurement data, topology connection data, maintenance plan data, and dispatching procedure data to obtain the raw data set; The power grid measurement data, topology connection data, and maintenance plan data are identified, unified, and time-series processed to form measurement time series, topology status series, and maintenance status series, respectively, resulting in a time-series data set. The scheduling procedure data is parsed to generate a set of rule entries. The time series data set and the rule entry set are standardized to form a unified field and identification system, resulting in a standardized power grid dataset.

4. The knowledge graph-based power grid energy dispatch optimization method according to claim 1, characterized in that, The obtained scheduling knowledge set includes: Based on a standardized power grid dataset, define the entity type set, relation type set, and attribute field set of the power grid knowledge graph, and establish the schema layer of the power grid knowledge graph; Based on the schema layer, entities, relationships, and attribute values ​​are extracted from the standardized power grid dataset and written into the power grid knowledge graph to form the power grid knowledge graph. The power grid knowledge graph is used to perform graph query and reasoning to generate a scheduling knowledge set, which includes equipment operating status information, topology connectivity information, maintenance impact information, and scheduling procedure rule triggering information.

5. The knowledge graph-based power grid energy dispatch optimization method according to claim 1, characterized in that, The input sequence data for obtaining time series prediction includes: The structured information set is obtained by selecting equipment operation status information, topology connectivity status information, maintenance impact information, and scheduling procedure rule triggering information corresponding to the prediction object from the scheduling knowledge set; The structured information set is mapped into numerical feature vectors according to preset fields, and a corresponding scene label is generated for each timestamp to obtain a structured feature sequence. The structured feature sequence and the measurement time series are merged by timestamp, and the input window is truncated from the continuous timestamps according to the input window length to obtain the input sequence data for time series prediction.

6. The power grid energy dispatch optimization method based on knowledge graphs according to claim 1, characterized in that, The obtained rolling prediction parameter set includes: An improved liquid time constant network is constructed, including a state-suppressed liquid evolution function, a liquid coupling evolution structure under a dynamically connected topology, and a continuous-time jump unit; The input sequence data is input into the improved liquid time constant network in chronological order. The liquid state nodes are continuously updated based on the state-suppression liquid evolution function. The state-suppression liquid evolution function applies a first-order suppression term and a third-order suppression term to the liquid state and updates the liquid state together with the input driving term, resulting in the state-suppression updated liquid state sequence. During the continuous-time state update process, the connection relationship between nodes in the liquid state is updated based on the liquid coupling evolution under the dynamic connection topology. The node connectivity relationship is generated and the coupling relationship between nodes is updated according to the topological state in the input sequence data. The coupling relationship of the liquid state nodes changes with the topological state, resulting in a topology-adaptive liquid state sequence. The liquid state is updated by continuous time jump unit. The jump time is determined according to the event trigger information in the input sequence data. At the jump time, the liquid state is superimposed with the jump variable obtained by input and the continuous time state update continues. The final liquid state sequence is input to the output layer to obtain the rolling prediction parameter set. The rolling prediction parameter set includes load prediction sequence, renewable power output prediction sequence, available capacity prediction sequence and risk indicator prediction sequence.

7. The power grid energy dispatch optimization method based on knowledge graphs according to claim 1, characterized in that, The summarization yields the input for the scheduling optimization solution, including: The scheduling object set is determined by the scheduling knowledge set and the corresponding scheduling decision variable set is generated. The scheduling object set includes generator sets, renewable energy output units, energy storage units and tie-line switching units. Based on the scheduling knowledge set, a set of scheduling conditions and a set of scheduling decision variable boundaries are generated. The set of scheduling conditions includes equipment operating status conditions, topology connectivity conditions, maintenance impact conditions, and scheduling procedure rule triggering conditions. The set of scheduling decision variable boundaries includes upper limit boundaries, lower limit boundaries, and availability boundaries, thus obtaining condition and boundary input items. Based on the rolling forecast parameter set, time period forecast parameter input items are generated, and scheduling target input items including operating cost items, wind and solar curtailment penalty items, and risk indicator penalty items are generated. The scheduling decision variable set, condition and boundary input items, time period forecast parameter input items, and scheduling target input items are summarized to obtain the scheduling optimization solution input.

8. The power grid energy dispatch optimization method based on knowledge graphs according to claim 1, characterized in that, The obtained scheduling optimization results include: A population of the dandelion optimization algorithm is constructed, which consists of multiple candidate scheduling solutions. Each candidate scheduling solution is composed of the values ​​of the scheduling decision variable set in each time period. The fitness value is calculated for each candidate scheduling solution. The fitness value is jointly determined by the operating cost item, the wind curtailment penalty item, and the risk indicator penalty item to obtain the initial optimal candidate scheduling solution. In each iteration, a drift update is performed on each candidate scheduling solution to generate an updated candidate scheduling solution. After generating the updated candidate scheduling solution, the fitness value is calculated and the optimal candidate scheduling solution is updated to obtain the iterative optimal candidate scheduling solution. Perform discontinuous drift direction control processing. If the fitness value of the optimal candidate scheduling solution does not improve within a preset number of consecutive iterations, switch the drift update direction of the candidate scheduling solution from the direction toward the optimal candidate scheduling solution to the direction away from the optimal candidate scheduling solution, and generate a new candidate scheduling solution based on the switched drift update direction. Perform a drift step size fault-based update process, dividing the iteration process into a first iteration interval, a second iteration interval, and a third iteration interval. The first drift step size is used in the first iteration interval, the second drift step size is used in the second iteration interval, and the third drift step size is used in the third iteration interval. The first drift step size is greater than the second drift step size, and the second drift step size is greater than the third drift step size. In each iteration interval, the drift update of the candidate scheduling solution is completed according to the corresponding drift step size. When the termination condition is met, the iterative optimal candidate scheduling solution is output as the scheduling optimization result. The termination condition includes reaching a preset number of iterations or the fitness value of the iterative optimal candidate scheduling solution does not exceed a preset threshold within a consecutive preset number of iterations.

9. The power grid energy dispatch optimization method based on knowledge graphs according to claim 1, characterized in that, The obtained power grid energy dispatch plan includes: The scheduling optimization results are converted into a grid energy dispatch plan, which includes generator output plans, renewable energy output arrangements, energy storage charging and discharging plans, reserve capacity plans, and tie-line switching plans for each time period. Based on the power grid energy dispatch plan, a dispatch execution instruction set is generated. The dispatch execution instruction set includes output setpoint instructions, charging and discharging setpoint instructions, standby setpoint instructions, and tie-line exchange setpoint instructions organized by object identifier and time period identifier. The scheduling execution instruction set is sent to the scheduling execution platform to complete the scheduling execution, and the scheduling execution result data is collected. The scheduling execution result data includes the actual output time series, the actual power flow time series, the energy storage state of charge time series, and the wind and solar curtailment time series, thus obtaining the grid energy scheduling execution result.

Images