A power dispatch scheme generation method and device, electronic equipment and storage medium

Abstract

The application discloses a power dispatching scheme generation method and device, electronic equipment and a storage medium, and belongs to the technical field of intelligent dispatching of a power system. The method comprises the following steps: acquiring power grid operation state measurement data and power grid dispatching event log data of a target power system, performing feature coding, fusion and topological bias superposition processing on the two data, and constructing a space-time joint embedding sequence; inputting the space-time joint embedding sequence into a pre-trained large language model to generate a structured simulation control sequence; performing tide flow numerical calculation based on the structured simulation control sequence, and combining the calculation result to carry out constraint checking and closed-loop adjustment, and finally outputting a power dispatching scheme meeting operation constraints. Therefore, by implementing the application, the problem that the physical compliance rate of the dispatching instructions generated by the large model is low and the power grid safety constraints are easily violated due to insufficient space-time feature fusion of multi-source data of the power grid and lack of physical mechanism closed-loop checking in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent dispatching technology for power systems, specifically to a method, apparatus, electronic device, and storage medium for generating power dispatching schemes. Background Technology

[0002] Power dispatching scheme generation is a core business for ensuring the safe, stable, and economical operation of the power system. With the continuous expansion of the modern power grid and the high proportion of new energy integration, power grid operation exhibits characteristics of high coordination between power sources, grids, loads, and storage, as well as nonlinear and dynamic changes, leading to a surge in the coupling and complexity of dispatching decisions.

[0003] However, existing power dispatching scheme generation methods still face technical challenges in practical applications, such as low physical compliance rates due to insufficient fusion of multi-source data, lack of topology awareness, and lack of mechanism-based closed-loop verification. The main reasons are twofold: First, existing methods fail to deeply encode and fuse power grid operating status measurement data with power grid dispatching event log data when extracting data features. Furthermore, they lack methods to organically integrate the topology bias coding of the power grid structure into the temporal feature level, resulting in generated schemes that cannot accurately perceive complex spatial electrical distances and spatiotemporal multidimensional relationships. Second, there is a severe disconnect between the generation logic of existing schemes and the underlying physical mechanisms of the power system. During the scheme output process, there is a lack of intervention mechanisms for real-time constraint verification and closed-loop adjustment using mechanism models such as power flow numerical calculations. This prevents the system from performing targeted corrections and iterative evolution of the control sequence based on numerical calculation results. Consequently, the generated dispatching schemes still pose a safety hazard of outputting commands that violate power physical constraints when facing overload or over-limit conditions. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for generating power dispatching schemes, which can solve the problem in the prior art that the dispatching instructions generated by large models have low physical compliance rate and are prone to violating power grid security constraints due to insufficient fusion of spatiotemporal features of multi-source data from the power grid and lack of physical mechanism closed-loop verification.

[0005] An embodiment of the present invention provides a method for generating a power dispatching scheme, comprising: Acquire grid operation status measurement data and grid dispatch event log data of the target power system; Feature encoding and fusion are performed on power grid operation status measurement data and power grid dispatch event log data to generate initial sequence labels; a preset topology bias code is superimposed on the initial sequence labels to generate a spatiotemporal joint embedding sequence; The spatiotemporal joint embedding sequence is input into a pre-trained large language model to output a structured simulation control sequence; Power flow numerical solutions are obtained by performing power flow numerical solutions based on structured simulation control sequences; Based on the power flow numerical solution results, the structured simulation control sequence is constrained and closed-loop adjusted to output a power dispatching scheme.

[0006] Furthermore, acquire grid operation status measurement data and grid dispatch event log data of the target power system, including: The continuous physical quantity values ​​of the target power system within a preset time window are obtained as power grid operation status measurement data; wherein, the continuous physical quantity values ​​include node voltage, branch current and injected power; Discrete text records of the target power system within the time window are acquired as power grid dispatch event log data; wherein, the discrete text records include equipment fault alarm records, relay protection action reports, dispatch operation logs, and accident contingency plan texts.

[0007] Furthermore, feature encoding and fusion are performed on power grid operation status measurement data and power grid dispatch event log data to generate initial sequence markers, including: The power grid operation status measurement data is normalized to obtain normalized measurement data, and the normalized measurement data is then mapped to intervals to obtain interval mapping features. Perform feature transformation on the interval mapping features to obtain continuous feature vectors; Based on a pre-defined event dictionary and pre-defined domain labels, the power grid dispatching event log data is mapped to obtain discrete event feature vectors; Extract the first equipment number code, first time code, and first data type code corresponding to the power grid operation status measurement data; Extract the second device number code, second time code, and second data type code corresponding to the power grid dispatch event log data; The first device number code, the first time code, the first data type code, and the continuous feature vector are concatenated to obtain the first concatenated feature vector; The second device number code, the second time code, the second data type code, and the discrete event feature vector are concatenated to obtain the second concatenated feature vector; The first concatenated feature vector is dimensionally aligned to obtain continuous quantity labels; the second concatenated feature vector is dimensionally aligned to obtain discrete event labels. The continuous quantity tags and discrete event tags are combined sequentially according to their time order to generate the initial sequence tags.

[0008] Furthermore, the preset topological bias code is superimposed onto the initial sequence marker to generate a spatiotemporal joint embedding sequence, including: Parse the device number codes contained in each tag in the initial sequence tag; Based on the device number code, the corresponding target topological bias feature is matched from the preset topological bias code; The target topological features are superimposed one by one onto the corresponding labels to generate a spatiotemporal joint embedding sequence.

[0009] Furthermore, the large language model is trained in the following ways: Obtain the scheduling scenario training dataset; the scheduling scenario training dataset includes several scheduling scenario samples; The training is iteratively performed on the training dataset based on the scheduling scenario until the preset number of iterations is reached, thus completing the training of the large language model. Each iteration of training includes: Extract scheduling scenario samples from the scheduling scenario training dataset and use them as the current training samples; The current training samples are input into the large language model to be trained, so that the large language model to be trained generates multiple independent inference branches in parallel based on the current training samples, and outputs the attention distribution map corresponding to each inference branch; wherein, each inference branch includes a natural language interpretation chain and a candidate structured simulation control sequence. The candidate structured simulation control sequences corresponding to each inference branch are imported into the preset power flow mechanism model for verification, and the physical hard constraint reward values ​​corresponding to each inference branch are obtained. The semantic similarity of the natural language interpretation chain corresponding to each inference branch and the candidate structured simulation control sequence is compared to obtain the logical consistency reward value corresponding to each inference branch. Entropy values ​​are calculated for the attention distribution maps corresponding to each inference branch to obtain the sparse attention reward values ​​for each inference branch. The physical hard constraint reward value, logical consistency reward value, and sparse attention reward value corresponding to each reasoning branch are weighted and summed to obtain the comprehensive reward value of each reasoning branch; Calculate the local mean and local standard deviation of the combined reward value for all inference branches; The relative dominance of each inference branch is calculated based on the comprehensive reward value corresponding to each inference branch, as well as the local mean and local standard deviation. Using a pre-defined optimizer, the network parameters of the large language model to be trained are updated using gradients based on the relative dominance of all inference branches.

[0010] Furthermore, power flow numerical solutions are performed based on the structured simulation control sequence to obtain power flow numerical solution results, including: Repeat the sequence iteration verification operation until the current set of out-of-limit nodes is empty, or the current number of executions equals the preset maximum number of executions; If the current set of nodes that exceed the limit is empty when the iteration stops, the current node voltage of each power grid node is determined as the power flow numerical solution result; If the current set of nodes that exceed the limit is not empty when the iteration stops, the preset safety margin scheme will be determined as the power flow numerical solution result; The sequence iterative verification operation includes: The current control sequence to be verified is analyzed, and the injected active power modification and injected reactive power modification of the node to be controlled are extracted; wherein, the initial control sequence to be verified is a structured simulation control sequence. Based on the modification of injected active power and the modification of injected reactive power, the parameters of the preset power grid mechanism model are updated to obtain the updated power grid mechanism model. Based on the updated power grid mechanism model, a set of nonlinear power flow balance equations is constructed. The nonlinear power flow balance equations are solved to obtain the current node voltage of each power grid node; Perform over-limit verification on the current node voltage, filter out over-limit nodes and calculate the voltage deviation value of each over-limit node to construct the current over-limit node set; If the current set of out-of-limit nodes is not empty, the voltage deviation value of each out-of-limit node is fed back to the pre-trained large language model to generate a reconstructed control sequence, and the reconstructed control sequence is updated to the current control sequence to be verified.

[0011] Furthermore, based on the power flow numerical solution results, constraint verification and closed-loop adjustment are performed on the structured simulation control sequence to output a power dispatching scheme, including: Analyze the power flow numerical solution results to obtain the data type corresponding to the power flow numerical solution results; When the data type indicates that the power flow numerical solution result is the current node voltage of each power grid node, the current branch power of each transmission line is calculated based on the current node voltage of each power grid node. Thermal stability verification is performed on each transmission line based on its current branch power, and the thermal stability verification results are obtained. If the thermal stability verification result is within the limit, the structured simulation control sequence is determined to meet the constraint conditions, and the structured simulation control sequence is determined as the power dispatching scheme. If the data type indicates that the power flow numerical solution result is the preset safety margin scheme, or the thermal stability verification result is out of limit, the structured simulation control sequence is determined to not meet the constraints, and the safety margin scheme is determined as the power dispatch scheme.

[0012] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0013] One embodiment of the present invention provides a power dispatching scheme generation device, including: a data acquisition module, a spatiotemporal sequence generation module, a model inference module, a power flow calculation module, and a verification output module; The data acquisition module is used to acquire power grid operation status measurement data and power grid dispatch event log data of the target power system; The spatiotemporal sequence generation module is used to perform feature encoding and fusion on power grid operation status measurement data and power grid dispatch event log data to generate an initial sequence label; and to superimpose a preset topology bias code onto the initial sequence label to generate a spatiotemporal joint embedding sequence. The model inference module is used to input the spatiotemporal joint embedding sequence into a pre-trained large language model and output a structured simulation control sequence. The power flow calculation module is used to perform power flow numerical calculation based on the structured simulation control sequence to obtain the power flow numerical calculation result; The verification output module is used to perform constraint verification and closed-loop adjustment on the structured simulation control sequence based on the power flow numerical solution results, and output the power dispatching scheme.

[0014] Based on the above method embodiments, the present invention provides corresponding electronic device embodiments.

[0015] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements any of the power dispatching scheme generation methods described in the above-described method embodiments.

[0016] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0017] One embodiment of the present invention provides a storage medium storing a computer program thereon, wherein, when the computer program is running, it controls the device where the storage medium is located to execute any of the power dispatching scheme generation methods described in the above-described method embodiments.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a method, apparatus, electronic device, and storage medium for generating power dispatch schemes. The method acquires grid operation status measurement data and grid dispatch event log data of a target power system, performs feature encoding, fusion, and topology bias overlay processing on the two data to construct a spatiotemporal joint embedding sequence; inputs the spatiotemporal joint embedding sequence into a pre-trained large language model to generate a structured simulation control sequence; performs power flow numerical calculation based on the structured simulation control sequence, and combines the calculation results to perform constraint verification and closed-loop adjustment, ultimately outputting a power dispatch scheme that meets operational constraints.

[0019] This invention overcomes the shortcomings of existing technologies, such as insufficient data fusion and inability to accurately perceive complex spatial electrical distances, by feature encoding and fusion of power grid operation status measurement data and power grid dispatch event log data, and by superimposing a preset topology bias code onto the initial sequence label. Simultaneously, this invention effectively bridges the gap between semantic generation and physical mechanisms by performing power flow numerical calculations on the structured simulation control sequences output by the large language model, and by executing strict constraint verification and closed-loop adjustments based on the calculation results. This solves the safety hazard of existing dispatch schemes easily outputting commands that violate power physical constraints. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a method for generating a power dispatching scheme according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of a power dispatching scheme generation device provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, to address the problem in existing technologies where insufficient fusion of spatiotemporal features of multi-source power grid data and lack of physical mechanism closed-loop verification lead to low physical compliance rates of dispatch instructions generated by large models and a high likelihood of violating power grid security constraints, an embodiment of the present invention provides a power dispatch scheme generation method, comprising at least the following steps: Step S1: Obtain power grid operation status measurement data and power grid dispatch event log data of the target power system; In a preferred embodiment, acquiring grid operation status measurement data and grid dispatch event log data of the target power system includes: The continuous physical quantity values ​​of the target power system within a preset time window are obtained as power grid operation status measurement data; wherein, the continuous physical quantity values ​​include node voltage, branch current and injected power; Discrete text records of the target power system within the time window are acquired as power grid dispatch event log data; wherein, the discrete text records include equipment fault alarm records, relay protection action reports, dispatch operation logs, and accident contingency plan texts.

[0024] Specifically, in the process of generating a power dispatching scheme, it is first necessary to acquire the grid operation status measurement data and grid dispatching event log data of the target power system. For the grid operation status measurement data, continuous physical quantity values ​​of the target power system within a preset time window are obtained as the grid operation status measurement data. Continuous physical quantity values ​​include node voltage, branch current, and injected power. The acquired continuous physical quantity values ​​specifically originate from the monitoring and data acquisition equipment and synchronous phasor measurement devices configured in the target power system. The preset time window is set to a complete continuous time period encompassing long-cycle historical operating conditions. To ensure the integrity of the time-series physical signals, full data acquisition must be performed within the preset time window to construct an original continuous data set containing node voltage, branch current, and injected power at all data sampling times.

[0025] For power grid dispatch event log data, discrete text records of the target power system within a preset time window are synchronously acquired as power grid dispatch event log data. Discrete text records include equipment fault alarm records, relay protection action reports, dispatch operation logs, and emergency response plan texts. Equipment fault alarm records reflect sudden abnormal states of power grid equipment and the time of fault events. Relay protection action reports present the action type of physical protection devices and fault equipment information when a fault occurs. Dispatch operation logs cover structured operation steps and operational basis issued by dispatchers for specific operating conditions. Emergency response plan texts provide standardized post-event risk analysis and key control objectives. Discrete text records and continuous physical quantity values ​​are collected in full within a preset time window, avoiding time window slicing or text truncation operations, to ensure the full retention of long-term causal chain information spanning days and hours of continuous operating conditions.

[0026] By fully acquiring continuous physical quantity values ​​and discrete text records from multiple heterogeneous sources within a preset time window, the blind spots of observation in a single monitoring dimension are eliminated. This enables the complete capture of long-cycle evolutionary characteristics such as chronic equipment degradation and continuous disturbance accumulation, thus providing solid and reliable data support for subsequent discovery of key causal trigger points and generation of scheduling schemes with causal chains.

[0027] Step S2: Perform feature encoding and fusion on the power grid operation status measurement data and power grid dispatch event log data to generate an initial sequence label; superimpose the preset topology bias code onto the initial sequence label to generate a spatiotemporal joint embedding sequence; In a preferred embodiment, feature encoding and fusion are performed on power grid operation status measurement data and power grid dispatch event log data to generate initial sequence markers, including: The power grid operation status measurement data is normalized to obtain normalized measurement data, and the normalized measurement data is then mapped to intervals to obtain interval mapping features. Perform feature transformation on the interval mapping features to obtain continuous feature vectors; Based on a pre-defined event dictionary and pre-defined domain labels, the power grid dispatching event log data is mapped to obtain discrete event feature vectors; Extract the first equipment number code, first time code, and first data type code corresponding to the power grid operation status measurement data; Extract the second device number code, second time code, and second data type code corresponding to the power grid dispatch event log data; The first device number code, the first time code, the first data type code, and the continuous feature vector are concatenated to obtain the first concatenated feature vector; The second device number code, the second time code, the second data type code, and the discrete event feature vector are concatenated to obtain the second concatenated feature vector; The first concatenated feature vector is dimensionally aligned to obtain continuous quantity labels; the second concatenated feature vector is dimensionally aligned to obtain discrete event labels. The continuous quantity tags and discrete event tags are combined sequentially according to their time order to generate the initial sequence tags.

[0028] In a preferred embodiment, a preset topological bias code is superimposed onto the initial sequence marker to generate a spatiotemporal joint embedding sequence, including: Parse the device number codes contained in each tag in the initial sequence tag; Based on the device number code, the corresponding target topological bias feature is matched from the preset topological bias code; The target topological features are superimposed one by one onto the corresponding labels to generate a spatiotemporal joint embedding sequence.

[0029] Specifically, after obtaining power grid operation status measurement data and power grid dispatch event log data, a feature encoding and fusion process is executed to generate initial sequence markers.

[0030] For feature extraction of continuous numerical values, the power grid operation status measurement data is normalized to obtain normalized measurement data. Normalization aims to eliminate scaling differences in physical quantities across different equipment scales. An interval mapping operation is performed on the normalized measurement data to generate interval-mapped features. Feature transformation operations are then applied to the interval-mapped features to calculate the continuous quantity feature vector.

[0031] For the vectorization of discrete text records, a pre-defined event dictionary and pre-defined domain labels are used to map power grid dispatching event log data, outputting discrete event feature vectors. The pre-defined event dictionary is configured with a standardized dispatching operation specification action table, and the pre-defined domain labels are specifically used to identify the unique attributes of different types of power anomalies and faults.

[0032] Extract the first device number code, first time code, and first data type code corresponding to the power grid operation status measurement data. Simultaneously extract the second device number code, second time code, and second data type code corresponding to the power grid dispatch event log data. Concatenate the first device number code, first time code, and first data type code with the continuous quantity feature vector to construct the first concatenated feature vector. Concatenate the second device number code, second time code, and second data type code with the discrete event feature vector to obtain the second concatenated feature vector.

[0033] Dimension alignment is applied to the first concatenated feature vector to generate corresponding continuous quantity labels. The same dimension alignment is applied to the second concatenated feature vector to generate discrete event labels with mapped dimensions. The continuous quantity labels and discrete event labels are then sequentially combined according to strict chronological order to construct a complete initial sequence label.

[0034] After obtaining the initial sequence labels, a preset topological bias code needs to be superimposed onto the initial sequence labels to generate a spatiotemporal joint embedding sequence. The device ID codes contained in each label of the initial sequence labels are parsed. Based on the device ID codes, the precisely corresponding target topological bias features are matched from the preset topological bias codes.

[0035] To transform the impedance properties of physical nodes into rigorous electrical correlation features in mathematical space, the node admittance matrix at the topology level of the target power system is extracted. The row vectors representing the self-admittance and mutual admittance of the target physical nodes in the node admittance matrix are selected, and these row vectors are identified as the electrical correlation features of the target physical nodes. The Laplace matrix corresponding to the node admittance matrix is ​​calculated, and eigenvalue decomposition is performed on the Laplace matrix to extract low-frequency eigenvector components. Based on the extracted low-frequency eigenvector components and the electrical correlation features, the partial features of the target topology are calculated. The specific calculation formula for the partial features of the target topology is as follows: In the formula, Indicates the target topological partial features; This represents the topological dimension alignment projection matrix; Represents low-frequency eigenvector components; Indicates electrical association characteristics; Represents the topological projection bias vector; symbol This indicates that the low-frequency feature vector components and the electrical correlation features are concatenated along a specified dimension. The physical node index represents the spatial topology of the target power system.

[0036] The target topological features determined by the matching are successively superimposed onto the corresponding initial sequence label elements to finally generate a spatiotemporal joint embedding sequence. The mathematical expression for the superposition update is as follows: In the formula, This represents the joint embedding element in the spatiotemporal joint embedding sequence corresponding to the time dimension index; Represents an initial sequence marker element within a specific time dimension of the initial sequence marker; Represents the time dimension index of the sequence, and its value is a positive integer greater than or equal to one.

[0037] By dimensionally aligning and integrating multi-scale continuous measurement data with discrete scheduling text, and deeply overlaying target topological features containing spatial physical connections, the semantic barriers at the level of multi-source heterogeneous data are effectively broken down. This enables the algorithm to accurately perceive the physical ability of electrical distance between adjacent nodes, greatly improving the accuracy of multi-track inference mechanism in cross-temporal causal tracing when facing complex abnormal power grid conditions.

[0038] Step S3: Input the spatiotemporal joint embedding sequence into a pre-trained large language model and output a structured simulation control sequence; In a preferred embodiment, the large language model is trained in the following manner: Obtain the scheduling scenario training dataset; the scheduling scenario training dataset includes several scheduling scenario samples; The training is iteratively performed on the training dataset based on the scheduling scenario until the preset number of iterations is reached, thus completing the training of the large language model. Each iteration of training includes: Extract scheduling scenario samples from the scheduling scenario training dataset and use them as the current training samples; The current training samples are input into the large language model to be trained, so that the large language model to be trained generates multiple independent inference branches in parallel based on the current training samples, and outputs the attention distribution map corresponding to each inference branch; wherein, each inference branch includes a natural language interpretation chain and a candidate structured simulation control sequence. The candidate structured simulation control sequences corresponding to each inference branch are imported into the preset power flow mechanism model for verification, and the physical hard constraint reward values ​​corresponding to each inference branch are obtained. The semantic similarity of the natural language interpretation chain corresponding to each inference branch and the candidate structured simulation control sequence is compared to obtain the logical consistency reward value corresponding to each inference branch. Entropy values ​​are calculated for the attention distribution maps corresponding to each inference branch to obtain the sparse attention reward values ​​for each inference branch. The physical hard constraint reward value, logical consistency reward value, and sparse attention reward value corresponding to each reasoning branch are weighted and summed to obtain the comprehensive reward value of each reasoning branch; Calculate the local mean and local standard deviation of the combined reward value for all inference branches; The relative dominance of each inference branch is calculated based on the comprehensive reward value corresponding to each inference branch, as well as the local mean and local standard deviation. Using a pre-defined optimizer, the network parameters of the large language model to be trained are updated using gradients based on the relative dominance of all inference branches.

[0039] Specifically, after obtaining the spatiotemporal joint embedding sequence, it needs to be input into a pre-trained large language model to output a structured simulation control sequence. In order for the large language model to have the ability to perceive physical laws and maintain logical consistency, it needs to be specifically trained in advance.

[0040] Obtain the scheduling scenario training dataset. The scheduling scenario training dataset includes several scheduling scenario samples. Perform iterative training based on the scheduling scenario training dataset until the preset number of iterations is reached, completing the training of the large language model.

[0041] In each training iteration, a specific scheduling scenario sample is extracted from the scheduling scenario training dataset as the current training sample. This current training sample is input into the large language model to be trained, prompting the model to generate multiple independent inference branches in parallel based on the current training sample, and outputting the attention distribution map corresponding to each inference branch. The total number of inference branches generated in parallel is set to... The independent indexes for each inference branch are: , Belongs to a group that is greater than or equal to one and less than or equal to one. Positive integers. Each inference branch contains a natural language interpretation chain and a candidate structured simulation control sequence.

[0042] The Natural Language Interpretation Chain (NLC) is the explicit inference process text output by a pre-trained large language model. The NLC explicitly covers state monitoring text, fault diagnosis text, and expected action text. The candidate structured simulation control sequence is the physical computation code built based on standard simulation engine rules and matched with the NLC. By synchronously generating the NLC and candidate structured simulation control sequences within the same inference branch, a dual-track inference architecture is established to provide transparent evidence for the model's inference process.

[0043] Candidate structured simulation control sequences corresponding to each inference branch are imported into a pre-defined power flow mechanism model for verification. The physical hard constraint reward value corresponding to each inference branch is obtained based on the limit violation verification state. If the candidate structured simulation control sequence satisfies all physical constraints, a positive physical hard constraint reward value is assigned; if a limit violation occurs, a negative physical hard constraint penalty value is assigned.

[0044] Semantic similarity is compared between the natural language interpretation chain corresponding to each inference branch and the candidate structured simulation control sequence to obtain the logical consistency reward value for each inference branch. The formula for calculating the logical consistency reward value is as follows: In the formula, Indicates that the independent index is The logical consistency reward value corresponding to the reasoning branch; This represents the cosine similarity calculation operator; This represents a predefined domain semantic embedding extraction function; Indicates that the independent index is The reasoning branches contain chains of natural language interpretation; Indicates that the independent index is The inference branches contain candidate structured simulation control sequences.

[0045] Entropy values ​​are calculated for the attention distribution maps corresponding to each inference branch to obtain the sparse attention reward value for each inference branch. The formula for calculating the sparse attention reward value is as follows: In the formula, Indicates that the independent index is The sparse attention reward value corresponding to the reasoning branch; This represents the information entropy calculation operator; Indicates that the independent index is Attention distribution map corresponding to the reasoning branches.

[0046] To standardize the dimensions of multidimensional reward signals and establish robust local comparative significance, batch normalization logic is introduced to perform real-time batch normalization operations on physical hard constraint reward values, logical consistency reward values, and sparse attention reward values. The normalization calculation formula is as follows: In the formula, This represents the standardized reward score. The reward type label index represents the physical hard constraint direction, the logical consistency direction, and the sparse attention direction, respectively. This represents the original category reward value before normalization. This represents the generated data corresponding to a single current training sample. Real-time batch local mean of reward scores for a specific category within an independent reasoning branch; This represents the generated data corresponding to a single current training sample. Real-time batch local standard deviation of reward scores for a specific category within an independent inference branch; This represents a preset minimal smoothing constant to prevent the denominator from being zero.

[0047] The weighted sum of the normalized physical hard constraint reward value, logical consistency reward value, and sparse attention reward value for each inference branch is used to obtain the comprehensive reward value for each inference branch. The formula for calculating the comprehensive reward value is as follows: In the formula, Indicates that the independent index is The overall reward value corresponding to the reasoning branch; This represents the preset weight coefficients for physical hard constraints; This represents the physical hard constraint reward value after standardization. This represents the pre-defined weighting coefficient for logical consistency items; This represents the logical consistency reward value after standardization. This represents the preset weight coefficients for the sparse attention term; This represents the sparse attention reward value after standardization.

[0048] Calculate the combined local mean and combined local standard deviation of the overall reward value for all inference branches. Based on the overall reward value, combined local mean, and combined local standard deviation for each inference branch, calculate the relative advantage of each inference branch. The formula for calculating the relative advantage is as follows: In the formula, Indicates that the independent index is The relative dominance of the reasoning branches; This represents the combined local mean constructed from the combined reward values ​​of all independent inference branches; This represents the combined local standard deviation of the combined reward values ​​of all independent inference branches.

[0049] Using a pre-defined optimizer, the network parameters of the large language model to be trained are updated using gradients based on the relative dominance of all inference branches.

[0050] After updating the network parameters of the large language model, the comprehensive reward values ​​of all inference branches corresponding to the current training sample are numerically sorted. The inference branch with the largest comprehensive reward value is extracted as the optimal policy branch, and the natural language interpretation chain contained in the optimal policy branch and the candidate structured simulation control sequence are synchronously stored in a pre-set scheduling knowledge base. Incorporating the optimal policy branch into the knowledge base provides high-quality domain demonstration samples for the large language model, thereby supporting the model's experience reuse and adaptive evolution when facing new access devices or unknown abnormal operating conditions.

[0051] By introducing parallel sampling with multidimensional rewards and strictly defining the local batch statistical domain, the traditional alignment algorithm's dependence on a large external evaluation network is removed, and the model achieves highly consistent adaptive alignment of power physics rules and textual logic under limited computing resources.

[0052] Step S4: Perform power flow numerical solution based on the structured simulation control sequence to obtain the power flow numerical solution results; In a preferred embodiment, power flow numerical calculation is performed based on the structured simulation control sequence to obtain power flow numerical calculation results, including: Repeat the sequence iteration verification operation until the current set of out-of-limit nodes is empty, or the current number of executions equals the preset maximum number of executions; If the current set of nodes that exceed the limit is empty when the iteration stops, the current node voltage of each power grid node is determined as the power flow numerical solution result; If the current set of nodes that exceed the limit is not empty when the iteration stops, the preset safety margin scheme will be determined as the power flow numerical solution result; The sequence iterative verification operation includes: The current control sequence to be verified is analyzed, and the injected active power modification and injected reactive power modification of the node to be controlled are extracted; wherein, the initial control sequence to be verified is a structured simulation control sequence. Based on the modification of injected active power and the modification of injected reactive power, the parameters of the preset power grid mechanism model are updated to obtain the updated power grid mechanism model. Based on the updated power grid mechanism model, a set of nonlinear power flow balance equations is constructed. The nonlinear power flow balance equations are solved to obtain the current node voltage of each power grid node; Perform over-limit verification on the current node voltage, filter out over-limit nodes and calculate the voltage deviation value of each over-limit node to construct the current over-limit node set; If the current set of out-of-limit nodes is not empty, the voltage deviation value of each out-of-limit node is fed back to the pre-trained large language model to generate a reconstructed control sequence, and the reconstructed control sequence is updated to the current control sequence to be verified.

[0053] Specifically, after generating the structured simulation control sequence, power flow numerical calculations are required to verify the physical feasibility of the control commands. To construct a complete iterative verification closed loop, the initial control sequence to be verified is defined as the obtained structured simulation control sequence, and the initial current execution count is set to zero. The sequence iterative verification process is then initiated, continuously repeating the sequence iterative verification operation until the current set of out-of-limit nodes is determined to be empty, or the current execution count equals the preset maximum execution count. If the stop iteration condition is triggered and the current set of out-of-limit nodes is determined to be empty, the current node voltage of each grid node is determined as the final power flow numerical calculation result. If the current set of out-of-limit nodes is not empty when the stop iteration condition is triggered, it means that even reaching the maximum number of solution rounds has failed to eliminate the physical out-of-limit risks. In this case, a forced truncation and degradation mechanism is triggered, and the preset safety margin scheme is determined as the final power flow numerical calculation result. The preset safety margin scheme is a conservative fixed control strategy pre-set to ensure the safety of the baseline operation.

[0054] When entering the sequence iteration verification operation, the current control sequence to be verified is first parsed to extract the injected active power modification and injected reactive power modification of the node to be regulated. The injected active power modification represents the expected physical change in the actual active power consumed or emitted by the node to be regulated by the control command, while the injected reactive power modification represents the expected physical change in reactive power throughput by the control command. Based on the extracted injected active power modification and injected reactive power modification, the physical parameters of the preset power grid mechanism model are updated to obtain the updated power grid mechanism model. The internal operating state of the updated power grid mechanism model can completely map the underlying control intent issued by the current control sequence to be verified.

[0055] The injected active and reactive power of each grid node in the updated grid mechanism model are obtained, and a set of nonlinear power flow balance equations is constructed. To avoid the risk of object review caused by purely functional feature descriptions, the nonlinear power flow balance equations are specifically expanded into polar coordinate form. The specific mathematical expressions of the nonlinear power flow balance equations are as follows: In the formula, This represents the injected active power of the grid node corresponding to the node index in the updated grid mechanism model; This indicates the node voltage amplitude corresponding to the power grid node at the node index; Indicates the total number of power grid nodes; Indicates the index of adjacent nodes; This indicates the node voltage amplitude corresponding to the power grid node with the adjacent node index; This indicates the conductance parameters between the power grid node corresponding to the node index and the power grid nodes corresponding to the adjacent node index; This represents the phase angle difference of the node voltages; This indicates the electrical allowance parameters between the power grid node corresponding to the node index and the power grid nodes corresponding to the adjacent node index; This represents the injected reactive power of the grid node corresponding to the node index in the updated grid mechanism model; Indicates the node index.

[0056] Numerical algorithms are applied to solve the constructed nonlinear power flow equilibrium equations, yielding the current node voltages of each grid node. After obtaining the current node voltages, each node voltage is checked for exceeding the safety thresholds set by the grid operation procedures. Nodes that do not meet the safety thresholds are selected, and the absolute difference between each node's voltage exceeding the safety threshold is calculated as the voltage deviation value. All selected nodes are aggregated to construct a current set of over-limit nodes. If the constructed current set of over-limit nodes is not empty, the voltage deviation values ​​of each over-limit node are fed back as physical hard constraints to a pre-trained large language model. The pre-trained large language model performs logical backtracking based on the received voltage deviation values, regenerating and reconstructing the control sequence.

[0057] The specific corrective actions involved in regenerating the reconfigured control sequence include reallocating the active power output of generator units, adjusting load transfer between regional nodes, switching backup transmission lines in the topology, and reducing the control amplitude of specific power equipment. By mapping logical backoff to actual grid regulation measures, the model reconfiguration actions are ensured to have clear electrical physical execution significance.

[0058] The reconstructed control sequence is updated to the current control sequence to be verified. At the same time, the current execution count is increased by one according to the preset step size to update the current execution count. Then, the next round of sequence iteration verification operation is started.

[0059] By combining the polar coordinate nonlinear power flow balance equations at the physical level with the inference depth of the large language model, and setting rigorous maximum execution count cutoff conditions and a safety guarantee output mechanism, not only is the abstract scheduling black-box verification completely white-boxed, but the security risk of the algorithm getting stuck in an infinite loop is also eliminated, effectively ensuring the absolute security and underlying compliance of the output scheduling scheme in actual industrial application scenarios.

[0060] Step S5: Based on the power flow numerical solution results, perform constraint verification and closed-loop adjustment on the structured simulation control sequence, and output the power dispatching scheme.

[0061] In a preferred embodiment, the structured simulation control sequence is constrained and closed-loop adjusted based on the power flow numerical solution results, and a power dispatching scheme is output, including: Analyze the power flow numerical solution results to obtain the data type corresponding to the power flow numerical solution results; When the data type indicates that the power flow numerical solution result is the current node voltage of each power grid node, the current branch power of each transmission line is calculated based on the current node voltage of each power grid node. Thermal stability verification is performed on each transmission line based on its current branch power, and the thermal stability verification results are obtained. If the thermal stability verification result is within the limit, the structured simulation control sequence is determined to meet the constraint conditions, and the structured simulation control sequence is determined as the power dispatching scheme. If the data type indicates that the power flow numerical solution result is the preset safety margin scheme, or the thermal stability verification result is out of limit, the structured simulation control sequence is determined to not meet the constraints, and the safety margin scheme is determined as the power dispatch scheme.

[0062] Specifically, after completing the power flow numerical calculation, a data type parsing operation is performed on the output power flow numerical calculation results to obtain the data type corresponding to the results. The data type is clearly divided into two independent categories: the current node voltage of each grid node and the preset safety margin scheme.

[0063] When the data type indicates that the power flow numerical solution result is the current node voltage of each grid node, it means that the structured simulation control sequence has passed the underlying node voltage over-limit check. In the current state, based on the current node voltage of each grid node and combined with preset line impedance parameters, the current branch power of each transmission line is calculated. Branch power represents the actual flowing load value of electrical energy passing through the physical transmission line. After obtaining the current branch power of each transmission line, a thermal stability check is performed on the current branch power of each transmission line. The thermal stability check aims to assess whether the physical line has the risk of overheating and damage under the current transmission power. After comprehensive verification and comparison, the thermal stability check result is output. The thermal stability check result is clearly divided into non-over-limit state and over-limit state.

[0064] If the thermal stability verification result is within acceptable limits, the structured simulation control sequence is determined to fully satisfy all rigid physical constraints of the power grid operation. Once the constraints are satisfied, the structured simulation control sequence is formally adopted as the power dispatching scheme.

[0065] If the data type indicates that the power flow numerical solution result is the preset safety margin scheme, or the thermal stability verification result is in an out-of-limit state, the structured simulation control sequence is determined to not meet the constraints of power grid operation. Since the strategy produced by the solution has the potential to cause a power grid accident, the closed-loop adjustment process is directly cut off, and the preset safety margin scheme is forcibly determined as the final power dispatch scheme to ensure the bottom-line operational safety of the physical power grid.

[0066] To further quantify the impact of historical input events on the final decision, a counterfactual interference mechanism is introduced after determining the power dispatch scheme to evaluate the marginal causal contribution of key historical event elements. The original dispatch action vector and original natural language interpretation chain corresponding to the determined power dispatch scheme are extracted. Key historical event elements in the input sequence are masked, and the data is re-inferred through a pre-trained large language model to output counterfactual dispatch action vectors and counterfactual natural language interpretation chains.

[0067] Calculate the policy difference metric between the original scheduling action vector and the counterfactual scheduling action vector. The formula for calculating the policy difference metric is as follows: In the formula, This represents a measure of strategy difference. Represents the original scheduling action vector; Represents the counterfactual scheduling action vector; This represents the maximum theoretical Euclidean distance constant in the normalized action space; Indicates auxiliary weights for semantic distance between texts; This represents the cosine similarity calculation operator; Represents a text feature mapping function; This represents the original natural language interpretation chain; This represents a counterfactual natural language interpretation chain; Indicates the parameter indicating changes in key action types.

[0068] Based on the calculated strategy difference metric, the marginal causal contribution assessment value of the masked key historical event elements is calculated. The formula for calculating the marginal causal contribution assessment value is as follows: In the formula, This represents the marginal causal contribution assessment value; This indicates key historical event elements that are being concealed; The zero norm represents the key historical event element being covered up, and the absolute number of key historical event elements being covered up is represented by . This represents the preset minimum smoothing constant.

[0069] After calculating the marginal causal contribution assessment value, the final determined power dispatch scheme, together with the marginal causal contribution assessment value, will be output as the final result.

[0070] After outputting the final results, an archiving operation is performed on the scheduling inference process of the target power system within the current time window. The natural language interpretation chain, power flow numerical solution limit exceedance feedback records, and marginal causal contribution assessment values ​​are obtained during the generation of the power dispatch scheme. The obtained natural language interpretation chain, power flow numerical solution limit exceedance feedback records, and marginal causal contribution assessment values ​​are encapsulated according to their occurrence sequence to generate a causal attribution report with a traceable logical chain. This causal attribution report is then sent to the dispatch display terminal for operator review.

[0071] By combining numerical solution data types with line thermal stability verification for dual physical fallback, and introducing a counterfactual causal quantification mechanism under a strict measurement space after the scheme is generated, the global risk of issuing illegal instructions is completely blocked, while giving scheduling decisions extremely high traceability.

[0072] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0073] like Figure 2 As shown, an embodiment of the present invention provides a power dispatching scheme generation device, including: a data acquisition module, a spatiotemporal sequence generation module, a model inference module, a power flow calculation module, and a verification output module; The data acquisition module is used to acquire power grid operation status measurement data and power grid dispatch event log data of the target power system; The spatiotemporal sequence generation module is used to perform feature encoding and fusion on power grid operation status measurement data and power grid dispatch event log data to generate an initial sequence label; and to superimpose a preset topology bias code onto the initial sequence label to generate a spatiotemporal joint embedding sequence. The model inference module is used to input the spatiotemporal joint embedding sequence into a pre-trained large language model and output a structured simulation control sequence. The power flow calculation module is used to perform power flow numerical calculation based on the structured simulation control sequence to obtain the power flow numerical calculation result; The verification output module is used to perform constraint verification and closed-loop adjustment on the structured simulation control sequence based on the power flow numerical solution results, and output the power dispatching scheme.

[0074] It should be noted that the embodiments of the apparatus described above correspond to the embodiments of the present invention described above, and can realize the power dispatching scheme generation method described in any one of the present invention. Furthermore, the embodiments of the apparatus described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Additionally, in the accompanying drawings of the apparatus embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0075] Based on the above-described method embodiments of the present invention, a corresponding embodiment of an electronic device is provided.

[0076] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power dispatching scheme generation method according to any one of the present invention, or, when the processor executes the computer program, it implements the functions of each module in the above-described device embodiments.

[0077] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0078] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0079] The processor can be a Central Processing Unit (CPU), or 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. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0080] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0081] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments; Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is running, the device where the storage medium is located executes any of the above-described power dispatching scheme generation methods of the present invention.

[0082] The aforementioned storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0083] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0084] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for generating a power dispatching scheme, characterized in that, include: Acquire grid operation status measurement data and grid dispatch event log data of the target power system; Feature encoding and fusion are performed on power grid operation status measurement data and power grid dispatch event log data to generate initial sequence labels; The preset topological bias code is superimposed on the initial sequence label to generate a spatiotemporal joint embedding sequence; The spatiotemporal joint embedding sequence is input into a pre-trained large language model to output a structured simulation control sequence; Power flow numerical solutions are obtained by performing power flow numerical solutions based on structured simulation control sequences; Based on the power flow numerical solution results, the structured simulation control sequence is constrained and closed-loop adjusted to output a power dispatching scheme.

2. The power dispatching scheme generation method as described in claim 1, characterized in that, Acquire grid operation status measurement data and grid dispatch event log data of the target power system, including: The continuous physical quantity values ​​of the target power system within a preset time window are obtained as power grid operation status measurement data; wherein, the continuous physical quantity values ​​include node voltage, branch current and injected power; Discrete text records of the target power system within the time window are acquired as power grid dispatch event log data; wherein, the discrete text records include equipment fault alarm records, relay protection action reports, dispatch operation logs, and accident contingency plan texts.

3. The power dispatching scheme generation method as described in claim 2, characterized in that, Feature encoding and fusion are performed on power grid operation status measurement data and power grid dispatch event log data to generate initial sequence labels, including: The power grid operation status measurement data is normalized to obtain normalized measurement data, and the normalized measurement data is then mapped to intervals to obtain interval mapping features. Perform feature transformation on the interval mapping features to obtain continuous feature vectors; Based on a pre-defined event dictionary and pre-defined domain labels, the power grid dispatching event log data is mapped to obtain discrete event feature vectors; Extract the first equipment number code, first time code, and first data type code corresponding to the power grid operation status measurement data; Extract the second device number code, second time code, and second data type code corresponding to the power grid dispatch event log data; The first device number code, the first time code, the first data type code, and the continuous feature vector are concatenated to obtain the first concatenated feature vector; The second device number code, the second time code, the second data type code, and the discrete event feature vector are concatenated to obtain the second concatenated feature vector; The first concatenated feature vector is dimensionally aligned to obtain continuous quantity labels; the second concatenated feature vector is dimensionally aligned to obtain discrete event labels. The continuous quantity tags and discrete event tags are combined sequentially according to their time order to generate the initial sequence tags.

4. The power dispatching scheme generation method as described in claim 3, characterized in that, The preset topological bias code is superimposed onto the initial sequence marker to generate a spatiotemporal joint embedding sequence, including: Parse the device number codes contained in each tag in the initial sequence tag; Based on the device number code, the corresponding target topological bias feature is matched from the preset topological bias code; The target topological features are superimposed one by one onto the corresponding labels to generate a spatiotemporal joint embedding sequence.

5. The power dispatching scheme generation method as described in claim 4, characterized in that, Large language models can be trained in the following ways: Obtain the scheduling scenario training dataset; the scheduling scenario training dataset includes several scheduling scenario samples; The training is iteratively performed on the training dataset based on the scheduling scenario until the preset number of iterations is reached, thus completing the training of the large language model. Each iteration of training includes: Extract scheduling scenario samples from the scheduling scenario training dataset and use them as the current training samples; The current training samples are input into the large language model to be trained, so that the large language model to be trained generates multiple independent inference branches in parallel based on the current training samples, and outputs the attention distribution map corresponding to each inference branch; wherein, each inference branch includes a natural language interpretation chain and a candidate structured simulation control sequence. The candidate structured simulation control sequences corresponding to each inference branch are imported into the preset power flow mechanism model for verification, and the physical hard constraint reward values ​​corresponding to each inference branch are obtained. The semantic similarity of the natural language interpretation chain corresponding to each inference branch and the candidate structured simulation control sequence is compared to obtain the logical consistency reward value corresponding to each inference branch. Entropy values ​​are calculated for the attention distribution maps corresponding to each inference branch to obtain the sparse attention reward values ​​for each inference branch. The physical hard constraint reward value, logical consistency reward value, and sparse attention reward value corresponding to each reasoning branch are weighted and summed to obtain the comprehensive reward value of each reasoning branch; Calculate the local mean and local standard deviation of the combined reward value for all inference branches; The relative dominance of each inference branch is calculated based on the comprehensive reward value corresponding to each inference branch, as well as the local mean and local standard deviation. Using a pre-defined optimizer, the network parameters of the large language model to be trained are updated using gradients based on the relative dominance of all inference branches.

6. The power dispatching scheme generation method as described in claim 5, characterized in that, Power flow numerical solutions are performed based on structured simulation control sequences to obtain power flow numerical solution results, including: Repeat the sequence iteration verification operation until the current set of out-of-limit nodes is empty, or the current number of executions equals the preset maximum number of executions; If the current set of nodes that exceed the limit is empty when the iteration stops, the current node voltage of each power grid node is determined as the power flow numerical solution result; If the current set of nodes that exceed the limit is not empty when the iteration stops, the preset safety margin scheme will be determined as the power flow numerical solution result; The sequence iterative verification operation includes: The current control sequence to be verified is analyzed, and the injected active power modification and injected reactive power modification of the node to be controlled are extracted; wherein, the initial control sequence to be verified is a structured simulation control sequence. Based on the modification of injected active power and the modification of injected reactive power, the parameters of the preset power grid mechanism model are updated to obtain the updated power grid mechanism model. Based on the updated power grid mechanism model, a set of nonlinear power flow balance equations is constructed. The nonlinear power flow balance equations are solved to obtain the current node voltage of each power grid node; Perform over-limit verification on the current node voltage, filter out over-limit nodes and calculate the voltage deviation value of each over-limit node to construct the current over-limit node set; If the current set of out-of-limit nodes is not empty, the voltage deviation value of each out-of-limit node is fed back to the pre-trained large language model to generate a reconstructed control sequence, and the reconstructed control sequence is updated to the current control sequence to be verified.

7. The power dispatching scheme generation method as described in claim 6, characterized in that, Based on the power flow numerical solution results, constraint verification and closed-loop adjustment are performed on the structured simulation control sequence to output a power dispatching scheme, including: Analyze the power flow numerical solution results to obtain the data type corresponding to the power flow numerical solution results; When the data type indicates that the power flow numerical solution result is the current node voltage of each power grid node, the current branch power of each transmission line is calculated based on the current node voltage of each power grid node. Thermal stability verification is performed on each transmission line based on its current branch power, and the thermal stability verification results are obtained. If the thermal stability verification result is within the limit, the structured simulation control sequence is determined to meet the constraint conditions, and the structured simulation control sequence is determined as the power dispatching scheme. If the data type indicates that the power flow numerical solution result is the preset safety margin scheme, or the thermal stability verification result is out of limit, the structured simulation control sequence is determined to not meet the constraints, and the safety margin scheme is determined as the power dispatch scheme.

8. A power dispatching scheme generation device, characterized in that, include: The module includes a data acquisition module, a spatiotemporal sequence generation module, a model inference module, a power flow calculation module, and a verification output module. The data acquisition module is used to acquire power grid operation status measurement data and power grid dispatch event log data of the target power system; The spatiotemporal sequence generation module is used to perform feature encoding and fusion on power grid operation status measurement data and power grid dispatch event log data to generate initial sequence markers. The preset topological bias code is superimposed on the initial sequence label to generate a spatiotemporal joint embedding sequence; The model inference module is used to input the spatiotemporal joint embedding sequence into a pre-trained large language model and output a structured simulation control sequence. The power flow calculation module is used to perform power flow numerical calculation based on the structured simulation control sequence to obtain the power flow numerical calculation result; The verification output module is used to perform constraint verification and closed-loop adjustment on the structured simulation control sequence based on the power flow numerical solution results, and output the power dispatching scheme.

9. An electronic device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the power dispatch scheme generation method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the power dispatching scheme generation method as described in any one of claims 1 to 7.

Images