An electricity market large model semantic mapping architecture and a construction method thereof

Abstract

The application relates to the technical field of power market information, in particular to a power market large model semantic mapping architecture and a construction method thereof, which comprises a data access module used for accessing structured transaction data and time series measurement data of a power market, and uniformly cataloging various types of data according to a pre-established metadata system; a semantic mapping module used for mapping various types of data to a same semantic space to obtain semantic vectors, and performing constraint projection on the semantic vectors based on a physical manifold defined by a power grid physical equation, so that the semantic vectors after constraint projection are located on the physical manifold; and a semantic knowledge construction module used for constructing a semantic graph with physical nodes and lines as graph nodes, electrical connection relationships and transaction relationships as edges according to the power grid topology data, hanging the semantic vectors after constraint projection as node representations of corresponding graph nodes, and generating time series event indexes from the semantic vectors corresponding to the time series measurement data after constraint projection.

Description

Technical Field

[0001] This invention relates to the field of electricity market information technology, specifically to a semantic mapping architecture for a large-scale electricity market model and its construction method. Background Technology

[0002] With the deepening of power market reform, the types of electricity transactions are becoming increasingly diverse, and the trading rhythm is becoming more refined. During market operation, multi-source heterogeneous data are continuously generated, including structured transaction data, time-series measurement data, grid topology data, and rule texts. To promptly identify issues such as abnormal price fluctuations, transaction manipulation, and settlement anomalies from this data, it is necessary to introduce the semantic understanding capabilities of large-scale models into power market data analysis.

[0003] Currently, existing technologies for analyzing and processing electricity market data mainly achieve this through the following methods: First, establishing a data catalog or data platform to register and collect data generated by various business systems to support data query and retrieval; Second, using data-driven methods such as neural networks to extract features and vectorize electricity data, and organizing the relationships between data based on vector retrieval or knowledge graph technologies to support downstream analytical applications.

[0004] However, the aforementioned existing technologies have at least the following shortcomings: First, transaction data, measurement data, topology data, and rule texts are generated by different business systems with different formats and definitions, and there is a lack of unified correlation criteria between the data. When conducting anomaly analysis, it is still necessary to manually confirm the data source and applicable rules, making it difficult to achieve automatic data retrieval and correlation analysis. Second, semantic vectors generated by data-driven methods may exhibit situations that do not conform to the physical laws of the power grid, such as an imbalance between node injected power and outflow power, or line power flow exceeding transmission capacity. If such semantic vectors are directly entered into downstream retrieval and analysis, the analysis conclusions will lack physical basis.

[0005] Therefore, this invention provides a semantic mapping architecture for a large-scale electricity market model and a method for constructing it. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is: a semantic mapping architecture for a large-scale electricity market model, which includes: The data access module is used to access structured transaction data, time-series measurement data, power grid topology data and rule texts from the electricity market, and to uniformly catalog all types of data according to a pre-established metadata system. The semantic mapping module is used to map various types of data to the same semantic space to obtain semantic vectors, and based on the physical manifold defined by the power grid physical equations, to constrain the semantic vectors carrying electrical quantity semantics so that the constrained and projected semantic vectors lie on the physical manifold. The semantic knowledge construction module is used to construct a semantic graph based on the power grid topology data, with physical nodes and lines as graph nodes and electrical connection relationships and transaction relationships as edges. It attaches the semantic vectors after constraint projection to the node representations of the corresponding graph nodes, and generates a time-series event index from the semantic vectors corresponding to the time-series measurement data after constraint projection. The ontology and rule base module is used to store the ontology and business rules in the power market field, record the effective versions of each business rule, and when the business rules are changed, compare the differences in the terms of different effective versions to locate the semantic vectors affected by them, and trigger the semantic mapping module to perform incremental remapping on the affected semantic vectors. The dynamic orchestration service module is used to parse business analysis tasks, orchestrate and call atomic services of semantic retrieval, feature fusion and rule reasoning to generate analysis pipelines, write the analysis results back to the semantic graph and time-series event index, and re-execute the corresponding analysis pipelines based on the semantic representation after incremental remapping.

[0008] Preferably, the establishment of the metadata system includes: sorting out the data requirements of typical anomaly analysis scenarios in the power market, including abnormal price fluctuations, transaction manipulation and settlement anomalies; extracting data types, data sources, related business rules and analysis scenario identifiers from the data requirements as metadata fields, establishing the relationship between the metadata fields, and forming the metadata system; The data access module catalogs various types of data in a unified manner according to a pre-established metadata system, including: parsing the data type and data source of the access data, generating metadata records containing the metadata fields for the access data, and associating the access data with the corresponding business analysis tasks according to the analysis scenario identifier in the metadata records.

[0009] Preferably, the semantic mapping module performs constrained projection on the semantic vector carrying the semantics of electrical quantities, including: The physical manifold is determined by the equality and inequality relationships between the node power balance equation and the line transmission constraints; the deviation of the semantic vector relative to the physical manifold is calculated, and the semantic vector with deviation is projected onto the physical manifold along the normal of the physical manifold.

[0010] Preferably, the semantic knowledge construction module generates the time-series event index by including the following steps: S401, the semantic vectors corresponding to the time series measurement data after constraint projection are arranged into a vector sequence according to the occurrence time, and the vector sequence is divided according to a preset time window; S402, compare the semantic vectors in adjacent time windows, identify the time windows in which the changes in semantic vectors meet the preset event conditions as events, and determine the event elements of the event, the event elements including the time of occurrence, event type, associated physical node and associated measurement; S403, Establish a time-series event index for the identified events according to the time of occurrence; S404, the time-series event index is organized hierarchically according to the day-ahead, intraday, real-time and settlement transaction rhythm of the electricity market, and the event views corresponding to different transaction rhythms are associated with the mapping chain of planned value, deviation and settlement amount.

[0011] Preferably, the semantic knowledge construction module constructs a semantic graph, including: Using physical nodes and lines as graph nodes, if physical nodes are directly electrically connected by lines, establish electrical connection edges between the corresponding graph nodes. If there are transaction records between market members to which physical nodes belong, establish transaction relationship edges between the corresponding graph nodes. And attach the semantic vector after constraint projection to the node representation of the corresponding graph node. The semantic knowledge construction module is also used to attach events in the time-series event index to the corresponding graph nodes in the semantic graph according to the associated physical nodes, and supports retrieving associated events and associated graph nodes in a preset neighborhood range along the edges of the semantic graph, starting from any graph node in the semantic graph.

[0012] Preferably, the electricity market ontology is organized according to the conceptual hierarchy of market participants, trading instruments, clearing mechanisms and constraints, and the conceptual nodes are connected by attributes and relationships. The business rules are represented in a structured form of rule conditions and rule actions, and are attached to the corresponding concept nodes in the domain ontology. The rule conditions reference the attributes of the attached concept nodes, and the rule actions act on the data objects associated with the attached concept nodes.

[0013] Preferably, the ontology and rule base module includes: The business rules of different effective versions are broken down into clauses and compared one by one to determine the added, deleted and modified clauses; based on the concept nodes attached to the modified clauses, the semantic graph nodes associated with the concept nodes and the events in the time sequence event index are determined, and the node representation of the corresponding graph node and the semantic vector corresponding to the event are determined as the affected semantic vectors. The execution scope of the incremental remapping is the determined affected semantic vectors, while the remaining semantic vectors remain unchanged; the ontology and rule base module records the correspondence between the terms in the different effective versions, and performs a backtracking query on the semantic representation corresponding to the historical data according to the business rule version that took effect when the historical data was generated.

[0014] Preferably, the dynamic orchestration service module generates the analysis pipeline by including the following steps: S801, parse the business analysis task to obtain the task type and analysis object of the business analysis task; S802, according to the pre-configured correspondence between task types and atomic service combinations, determine the semantic retrieval, feature fusion and rule reasoning atomic services to be invoked and their execution order, and generate the analysis pipeline; S803, execute the analysis pipeline, wherein when the feature fusion atomic service is invoked, perform intermodal attention interaction on the different modal semantic vectors after constraint projection, calculate the contribution weight of each modal semantic vector relative to the current business analysis task, perform weighted summation on the different modal semantic vectors according to the contribution weight, and project the weighted summation result back onto the physical manifold to obtain the fused semantic vector; S804, during the execution of the analysis pipeline, the source data, mapping path and the version of the business rule on which the called data are recorded as traceability information, so that the output analysis results carry the traceability information and support tracing back to the original data and business rule terms along the traceability information.

[0015] A method for constructing a semantic mapping architecture for a large-scale electricity market model includes the following steps: S1 accesses structured transaction data, time-series measurement data, grid topology data, and rule texts from the electricity market, and catalogs the accessed data in a unified manner according to a pre-established metadata system; S2, map the various types of data into the same semantic space to obtain semantic vectors, and perform constrained projection on the semantic vectors carrying electrical quantity semantics based on the physical manifold defined by the power grid physical equations, so that the constrained and projected semantic vectors are located on the physical manifold; S3. Construct a semantic graph based on the power grid topology data, with physical nodes and lines as graph nodes and electrical connection relationships and transaction relationships as edges. Attach the constrained projected semantic vectors as node representations of the corresponding graph nodes, and generate a time-series event index from the semantic vectors corresponding to the constrained projected time-series measurement data. S4, construct the ontology and business rule library for the power market field, and record the effective versions of each business rule; S5, parse the business analysis task, dynamically orchestrate and call the atomic services of semantic retrieval, feature fusion and rule reasoning to generate the analysis pipeline, and write the analysis results back to the semantic graph and time-series event index; S6. When business rules change, compare the differences between the terms of different effective versions, locate the semantic vectors affected by the differences in the terms, perform incremental remapping on the affected semantic vectors, and after the incremental remapping is completed, re-execute the corresponding analysis pipeline based on the semantic representation after incremental remapping.

[0016] Preferably, in step S2, constraining the projection of the semantic vector carrying electrical quantity semantics onto the physical manifold defined by the power grid physical equations includes: determining the physical manifold based on the equality and inequality relationships between the node power balance equations and the line transmission constraints; calculating the deviation of the semantic vector relative to the physical manifold; and projecting the semantic vector with deviation onto the physical manifold along the normal of the physical manifold.

[0017] This invention provides a semantic mapping architecture for a large-scale electricity market model and its construction method. It offers the following advantages: 1. By constraining the projection of semantic vectors onto the physical manifold defined by the node power balance equation and line transmission constraints, the semantic vectors entering the downstream retrieval and analysis correspond to the physically reachable power grid state. This avoids the semantic representations generated by data-driven methods from violating the physical laws of the power grid and ensures that the analysis conclusions have a physical basis. Furthermore, the projection is iteratively solved along the normal direction with the goal of minimizing the distance, which eliminates physical inconsistencies while preserving the original effective semantic features in the semantic vectors to the maximum extent.

[0018] 2. When business rules are changed, the affected scope is located along the association path of "modified terms → attached concept nodes → associated graph nodes and events → corresponding semantic vectors". Incremental remapping is performed only on the affected semantic vectors to avoid the computational overhead of full remapping, so that the analysis conclusions are consistent with the current rule standards. At the same time, backtracking queries are performed according to the rule version that took effect when the data was generated, so that historical analysis conclusions can be reproduced according to the rule standards that were effective at that time.

[0019] 3. By encapsulating semantic retrieval, feature fusion, and rule reasoning into atomic services and dynamically combining them according to task type to generate analysis pipelines, the processing flow can be reused as business and rules evolve; during feature fusion, the weighted summation result is projected back onto the physical manifold to ensure that the fused semantic vector still satisfies the physical consistency of the power grid; the analysis results carry traceability information of source data, mapping path, and rule version, which can be traced back to the original data and business rule clauses, making it easy to verify the basis of the analysis conclusions. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the architecture of the present invention; Figure 2 This is a flowchart of the physical manifold constraint projection of the present invention; Figure 3 This is a flowchart of the rule change incremental remapping and version rollback process of the present invention. Detailed Implementation

[0021] 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.

[0022] Overall architecture of a dedicated large-scale model for the electricity market: like Figures 1 to 3 As shown, this invention proposes a semantic mapping architecture for a large-scale electricity market model and its construction method, including: The data access module is used to access structured transaction data, time-series measurement data, power grid topology data and rule texts from the electricity market, and to uniformly catalog all types of data according to a pre-established metadata system. The semantic mapping module is used to map various types of data to the same semantic space to obtain semantic vectors. Based on the physical manifold defined by the power grid physical equations, it performs constrained projection on the semantic vectors carrying electrical quantity semantics (including semantic vectors corresponding to time series measurement data and semantic vectors generated by fusion), so that the constrained projected semantic vectors lie on the physical manifold. In addition to the node power balance equation and line transmission constraints, the power grid physical equations may also include one or more of the following: reactive power balance equation and node voltage amplitude upper and lower limit constraints. Those skilled in the art can select according to the required analytical precision. The semantic knowledge construction module is used to construct a semantic graph based on power grid topology data, with physical nodes and lines as graph nodes and electrical connection relationships and transaction relationships as edges. It attaches the semantic vectors after constraint projection to the node representations of the corresponding graph nodes, and generates a time-series event index from the semantic vectors corresponding to the time-series measurement data after constraint projection. The ontology and rule base module is used to store the ontology and business rules in the power market field, record the effective versions of each business rule, and when the business rules are changed, compare the differences in the terms of different effective versions to locate the semantic vectors affected by them, and trigger the semantic mapping module to perform incremental remapping on the affected semantic vectors. The dynamic orchestration service module is used to parse business analysis tasks, orchestrate and call atomic services of semantic retrieval, feature fusion and rule reasoning to generate analysis pipelines, write the analysis results back to the semantic graph and time-series event index, and re-execute the corresponding analysis pipelines based on the semantic representation after incremental remapping.

[0023] The overall architecture is designed as follows: using a metadata system to unify data definitions; using a unified semantic space and physical manifold constraints to achieve semanticization and physical trustworthiness of multimodal data; using semantic graphs and time-series event indexes to achieve queryable organization of semantic knowledge; using domain ontology and rule bases to achieve formal management of business knowledge; and using dynamic orchestration of atomic services to achieve on-demand combination of analytical capabilities. Each module is connected sequentially to form a complete technical link of "data access - semantic mapping - knowledge construction - rule management - analytical services".

[0024] Electricity trading data metadata system and unified cataloging: As an optional implementation, the transaction data, measurement data, power grid topology data and rule text of the electricity market are generated by different business systems, with different formats and standards. There is no unified basis for the correlation between the data, and manual confirmation of the data source and applicable rules is required when conducting anomaly analysis.

[0025] Therefore, a metadata system is established during the data access phase. The metadata system refers to a data attribute registration framework consisting of metadata fields and the relationships between fields, which is used to provide a unified registration standard for various heterogeneous data and support the automatic retrieval of data for subsequent business analysis tasks. Specifically, the metadata system is established in the following way: First, the data requirements for typical anomaly analysis scenarios in the power market are identified. These typical anomaly analysis scenarios are determined by summarizing the anomaly analysis business and report styles of domestic and foreign power trading institutions, including abnormal price fluctuations, transaction manipulation, and settlement anomalies. The data requirements for abnormal price fluctuation scenarios are nodal prices, market clearing prices, and load and renewable energy output for the corresponding time periods. The data requirements for transaction manipulation scenarios are market member bid sequences, transaction records, and relationships between related members. The data requirements for settlement anomaly scenarios are settlement electricity volume, settlement price, and details of assessment fees.

[0026] Based on this, metadata fields are extracted from the above data requirements, and attribute registration entries are generated for the access data according to the metadata fields. These attribute registration entries are the metadata records; let the metadata record be denoted as... Where: t is the data type field, taking values ​​from a set of data types including transaction data, measurement data, topology data, and rule text; s is the data source field, recording the identifier of the entity or business system that generated the data; B is the associated business rule field, i.e., the set of identifiers of business rules that constrain the business meaning of this data. , n is the number of associated business rules; a is the analysis scenario identifier field, which is a marker representing the category of abnormal analysis scenarios served by the data, and its value is taken from the set of abnormal analysis scenarios including abnormal price fluctuations, transaction manipulation and settlement anomalies.

[0027] Furthermore, relationships are established between metadata fields to form a metadata system. These relationships fall into two categories: first, the type constraint relationship between data type field t and associated business rule field B, where each data type corresponds to a range of business rules that can be attached, for example, measurement data can only be attached to measurement management rules; second, the scenario requirement relationship between analysis scenario identifier field a and data type field t, where each anomaly analysis scenario corresponds to a necessary combination of data types, for example, a settlement anomaly scenario requires a combination of transaction data and measurement data. Based on these relationships, if any data's metadata record lacks associated data of the data type required for its analysis scenario, the cataloging is deemed incomplete, and the data is returned to the data access stage for supplementary access. In some implementations, the data access module catalogs various types of data. Cataloging refers to the process of generating metadata records for the accessed data and registering them in the database. The unified cataloging process includes: parsing the data type t and data source s of the accessed data; generating a metadata record R containing metadata fields for the accessed data; and associating the accessed data with the corresponding business analysis task according to the analysis scenario identifier field a in the metadata record, so that when the dynamic orchestration service module parses the business analysis task, it can directly retrieve the cataloged data based on the analysis scenario identifier.

[0028] Unified representation of heterogeneous multimodal data in the electricity market: Specifically, the semantic mapping module maps various types of data to the same semantic space to obtain semantic vectors. Specifically, the semantic mapping module configures a corresponding feature extraction method for each modality. For structured transaction data, the numerical fields (such as transaction volume and transaction price) and categorical fields (such as transaction type and market participants) in each transaction record are encoded separately and then concatenated to obtain transaction characteristics; For time-series measurement data, the measurement sequence is segmented according to a preset time window, and the numerical change characteristics (including window mean, window extreme value and rate of change) within each time window are extracted to obtain the measurement characteristics. For power grid topology data, physical nodes and lines are used as graph elements. The attributes of each graph element (such as node type and line voltage level) and the connection relationships between graph elements are encoded together to obtain topology features. For the rule text, the text is segmented into clauses, and each clause is semantically encoded to obtain text features.

[0029] Based on this, the four modal features mentioned above are mapped to a semantic space of the same dimension through their respective transformation layers. The output dimension of each transformation layer is consistent, and different modal features belonging to the same analysis object are aligned during the transformation process. The alignment process refers to reducing the distance between different modal features corresponding to the same physical node, the same transaction period, or the same business rule in the semantic space based on the correlation between the cataloged data, so that semantically related data are adjacent to each other in the semantic space. After the transformation and alignment are completed, the representation of each type of data in the semantic space is a semantic vector. The semantic encoding of the rule text and the conversion layer for each modality are implemented by a pre-trained large language model: the organized description of the clause text or numerical field is input into the pre-trained large language model, and its hidden layer output is taken as the semantic feature of the corresponding modality; the conversion layer is the linear mapping layer at the end of the large language model, which maps the hidden layer output to a semantic space of the same dimension.

[0030] The pre-trained large language model is obtained as follows: The large language model refers to a deep neural network that is pre-trained in a large-scale text corpus under self-supervised conditions. Its pre-training task is to predict the occluded or subsequent words based on the context. By minimizing the prediction error, the network parameters are continuously adjusted so that the model can learn general language and semantic representation capabilities. This invention can directly use the pre-training parameters of publicly available pre-trained large language models (such as BERT and GPT series models) as initial parameters.

[0031] Furthermore, to adapt the model to the electricity market domain, domain-specific training is performed based on the initial parameters mentioned above: electricity market domain corpus is collected, including electricity market trading rules, dispatching and operation procedures, market announcements and settlement details texts. The domain corpus is then used to train the large language model according to a self-supervised pre-training task until the prediction error converges. After training is completed, the network parameters are fixed to obtain the pre-trained large language model used in this invention.

[0032] When in use, the organized description of the clause text or numerical field is input into the model, and the output of the pooled last hidden state is taken as the semantic feature of the corresponding modality. The transformation layer is a linear mapping layer connected to the output of the large language model, which linearly transforms the pooled hidden layer output to a unified semantic space dimension. The parameters of the linear mapping layer are optimized together with the alignment target during modality alignment training and then fixed.

[0033] It should be noted that the semantic vectors obtained by mapping multimodal data are generated in a data-driven manner, which may lead to situations that do not conform to the physical laws of the power grid, such as an imbalance between the injected and outflowing power at nodes or power flow exceeding the transmission capacity of lines. If such semantic vectors are directly used for downstream retrieval and analysis, the resulting analytical conclusions will lack physical basis.

[0034] Therefore, as an embodiment of the present invention, after mapping various types of data to the same semantic space, the semantic mapping module performs constrained projection on the semantic vectors carrying electrical quantity semantics (including semantic vectors corresponding to time-series measurement data and fused semantic vectors) based on the physical manifold defined by the power grid physical equations, so that the constrained projected semantic vectors lie on the physical manifold; wherein the physical manifold refers to the set of vectors in the semantic space that satisfy the physical laws of the power grid, defined by the equality and inequality relationships of the power grid physical equations; constrained projection refers to the process of correcting semantic vectors that do not satisfy the above set conditions into the set.

[0035] It should be noted that the semantic vectors of transaction data and rule text do not carry electrical quantity semantics, so there is no need to perform constraint projection, and they can directly enter the semantic knowledge construction stage. Specifically, let the semantic vector be z, that is, z is a real-valued vector in the semantic space, and each component of z carries the multimodal semantic features of the corresponding data. Let the decoding mapping from the semantic space to the electrical quantity space be denoted as z. Decoding mapping The pre-configured mapping relationships in the semantic mapping module are used to extract the corresponding electrical quantities from the semantic vector z; Decoding mapping The decoding network (such as a multi-layer fully connected network) is obtained through pre-training. Historical measurement data is used as training samples. The semantic vector obtained by semantic mapping of each sample is used as input, and the corresponding real electrical quantities (voltage amplitude, voltage phase angle, and injected active power at each node) are used as supervision labels. The output of the decoding network is then trained to approximate the supervision labels. After training, the decoding network is solidified as the decoding mapping. This function is directly called during constraint projection and subsequent processing and is no longer updated. but The electrical quantities corresponding to the semantic vector z include the voltage magnitude, voltage phase angle, and injected active power at each node. Where u is the node number, and u = 1, 2, ..., N, and N is the total number of power grid nodes. The unit is MW; The node power balance equation is then: , where v is the node number; The magnitude of the element in the u-th row and v-th column of the nodal admittance matrix, in units of s, when there is no direct electrical connection between node u and node v. Therefore, the summation only contributes to the connected nodes. Let be the voltage amplitude at node u. The voltage magnitude at node v is expressed in kV. Let be the voltage phase angle at node u. The voltage phase angle at node v is expressed in rad. The phase angle of the admittance matrix element is expressed in rad; the left side of the equation represents the injected active power at node u, and the right side represents the sum of the active power flowing out of node u through each connected line, both in MW. The reason for using the above formula is that power balance is a fundamental physical law that power grid operation must satisfy. For any truly reachable power grid operating state, the active power injected into each node must be equal to the sum of the active power flowing out through the connected lines; conversely, electrical quantities that do not satisfy power balance cannot exist physically. Therefore, by constraining semantic vectors with the power balance equation, semantic vectors located on the physical manifold can all correspond to physically reachable power grid states.

[0036] Based on this, the power balance equations of all nodes are taken as equalities, and the active power transmitted by each line does not exceed its transmission power limit as an inequality. The above equalities and inequalities together define the physical manifold in the semantic space: when the electrical quantity corresponding to the semantic vector z satisfies the power balance equations of all nodes, and the active power transmitted by each line does not exceed the corresponding limit, it is determined that z is located on the physical manifold.

[0037] Furthermore, the deviation of the semantic vector relative to the physical manifold is calculated: Substituting into the power balance equations of each node, the difference between the two sides of the equation is the power balance residual of the corresponding node; the portion of active power transmitted by each line exceeding its limit is the excess of the corresponding line. When all power balance residuals are zero and there is no excess, no projection is needed; otherwise, it is determined that z has a deviation, and normal projection is performed; normal projection is to find the vector on the physical manifold that has the smallest distance from the original semantic vector: , where z is the semantic vector before projection, that is, the real vector in the semantic space to be corrected, and each component is a semantic feature value, which is dimensionless; The variable used for optimization is denoted as , which represents all candidate semantic vectors traversing the physical manifold and is dimensionless. The projected semantic vector is the optimization result and is dimensionless. Let be the Euclidean norm of the vector, which is the arithmetic square root of the sum of the squares of the vector's components. The distance between the candidate vector and the vector before projection in the semantic space is dimensionless. argmin means in When traversing on a physical manifold, make The one that gets the minimum value and assign it to ; The physical manifold is defined as follows: in The power balance residual of node u is, i.e. Substituting into the power balance equation for node u, the difference between the right and left sides of the equation is expressed in MW; N is the total number of nodes in the power grid, which is a positive integer. Let L be the limit-crossing function for line l, and its function value is the aforementioned limit-crossing amount of line l: the portion of the active power actually transmitted by line l that exceeds its transmission power limit, in MW; L is the total number of power grid lines, which is a positive integer. That is, the active power transmitted by line l did not exceed its limit.

[0038] The reasons for using the above formula are as follows: First, every point on the physical manifold simultaneously satisfies all power balance equations and line transmission constraints, and optimization within the physical manifold guarantees that the projection result necessarily corresponds to a physically reachable power grid state; Second, in addition to the physically inconsistent components, the semantic vector also carries real and effective semantic features, with the minimum Euclidean distance as the optimization objective, that is, eliminating physical inconsistencies with the minimum correction amount and preserving the original semantic features to the maximum extent; moreover, the Euclidean norm treats each component equally, and the correction amount is evenly distributed across all semantic dimensions, without bias towards any specific semantic feature; Third, it is independent of... Optimization variables After traversing the manifold, the minimum point is then assigned a value. This makes the relationship between the solution objective and the solution variables clear, and the solution process executable.

[0039] Furthermore, it should be noted that the power balance residual... The physical meaning of the term is as follows: For any node u, its injected active power should be equal to the sum of the active power flowing out through all connected lines. The difference between the two is the power balance residual, measured in MW. A residual of zero indicates that the node's power input and output are perfectly balanced, corresponding to a physically possible operating state. A residual that is not zero indicates that the node has unbalanced power that cannot be disposed of—for example, a residual of +5MW means that the node has 5MW of injected power that is neither absorbed by loads nor transmitted through any lines, which is impossible to actually occur on the power grid. Therefore, whether the residuals of all nodes are zero is the equation basis for determining whether the semantic vector corresponds to a physically reachable state.

[0040] Beyond limited edition The physical meaning of "overload" is as follows: The actual active power transmitted by any line must not exceed its transmission power limit (determined by the line's thermal stability or stable operation requirements). The portion of the actual transmitted power exceeding the limit is called the overload, measured in MW. A overload greater than zero indicates that the line is in an overloaded state, which is physically unacceptable. An overload less than or equal to zero indicates that the line has a transmission margin. Therefore, whether all overloads are not greater than zero is the inequality basis for determining whether a semantic vector corresponds to a physically reachable state.

[0041] Furthermore, an iterative approach is adopted: in each iteration, the gradient of each power balance residual and the limit exceedance with respect to the current semantic vector is calculated, and the semantic vector is corrected along the gradient direction (i.e., the normal of the constraint surface at the current point), so that each residual and limit exceedance gradually decreases; the iteration continues until the absolute value of each power balance residual is less than the preset convergence accuracy (e.g., ...). (MW) and there is no stopping when the limit is exceeded, the result of the last iteration is taken as .

[0042] Based on the above, the following example can be used for further explanation: Taking a three-node power grid as an example: the injected active power of nodes 1, 2, and 3 are respectively , , If the sum of injected power is not zero, and substituting the electrical quantities of the corresponding semantic vector z into the nodal power balance equation results in a power balance residual, it is determined that z has a deviation. After the above iterative projection, the projected semantic vector... The corresponding injected active power is corrected to , , The power balance residuals at each node are all zero, and the active power transmitted on each connected line does not exceed the corresponding transmission power limit; therefore, there is no violation of the limit. If both equality and inequality relationships are satisfied, then determine... Located on the physical manifold, it proceeds to the subsequent semantic knowledge construction and analysis stage.

[0043] Queryable Semantic Data Construction: As an optional implementation, since time-series measurement data is a continuous numerical sequence, the amount of data stored point-by-point is large and lacks business meaning, making it difficult to directly support retrieval and analysis. Therefore, the semantic knowledge construction module identifies noteworthy changes in the measurement sequence as events and builds an index, transforming time-series data into knowledge that can be retrieved by time point and node. A time-series event index refers to a data structure that organizes identified events using the event's occurrence time as the search key; an event is a record of a noteworthy change in the measurement state within adjacent time periods.

[0044] Specifically, in S401, the semantic knowledge construction module arranges the semantic vectors corresponding to the constrained projected temporal measurement data into a vector sequence according to the order of occurrence of the measurements corresponding to each semantic vector; denoted as the time window sequence number. , Let m be the total number of time windows, and let T be the preset window length in minutes. In this embodiment, T is 15 minutes, consistent with the settlement cycle of the electricity market. The vector sequence is divided into m non-overlapping time windows according to the window length T. The mean of the semantic vectors within each time window is used as the representative vector of that time window, denoted as... ; Furthermore, in S402, the representative vectors of adjacent time windows are compared. and The distance between two representative vectors in the semantic space is used as the change between adjacent time windows. When the change meets the preset event conditions, the next time window is identified as an event. The preset event conditions are that the change exceeds the preset event judgment benchmark. The preset event judgment benchmark is set according to the historical fluctuation level of each physical node. For example, the preset multiple (such as 3 times) of the standard deviation of the change of the representative vectors of the historical adjacent time windows of the node is taken as the preset event judgment benchmark of the node. After an event is identified, the event elements are determined as follows: the occurrence time is the start time of the time window; the event type is determined based on the types of measurements involved in the change, including price mutations, output mutations, and congestion; the associated physical nodes are the graph nodes attached to the semantic vectors within the time window; the associated measurements are the identifiers of the original measurement data within the time window; if there is no comparable preceding time window at the beginning of the vector sequence, the first time window is not judged as an event, and its representative vector is directly retained for subsequent comparisons.

[0045] In S403, the identified events are arranged and a time-series event index is established according to the order of their occurrence. The time-series event index uses the occurrence time as the search key and the event element as the search content, and supports searching events by time range, by event type, and by associated physical node.

[0046] In S404, events in the time-series event index are divided into corresponding event views according to the trading rhythm of the electricity market, which is day-ahead, intraday, real-time, and settlement, forming four hierarchical event views. The event views corresponding to different trading rhythms are linked through a mapping chain of planned value, deviation amount, and settlement amount. Specifically, the planned value in the clearing result of the day-ahead rhythm is used as the comparison benchmark for intraday and real-time rhythms. The deviation amount between the real-time rhythm and the day-ahead rhythm is used to locate deviation events. The settlement amount of the settlement rhythm is used to verify the events of the aforementioned rhythms, so that any event can be queried along the mapping chain to find its corresponding record under other trading rhythms.

[0047] Thus, taking the output force measurement of node 1 as an example: after its semantic vector sequence is divided into 15-minute time windows, comparison reveals that the first... The first time window and the first If the distance between the representative vectors of each time window exceeds the preset event judgment benchmark of the node, the 9th time window is identified as an event. The event elements are determined as follows: the occurrence time is the start time of the 9th time window, the event type is output change, the associated physical node is node 1, and the associated measurement is the output measurement data identifier within the time window. The event is written into the time-series event index according to the occurrence time, and is classified into the real-time event view according to the real-time beat to which the occurrence time belongs. In subsequent analysis, the settlement volume record of the event under the settlement beat can be queried along the mapping chain.

[0048] As an optional embodiment, the time-series event index organizes events only by time, and the correspondence between events and the physical structure of the power grid is not explicitly expressed. When analyzing the anomaly of a certain node, it is impossible to directly know the association range of that node in the power grid.

[0049] Specifically, the semantic knowledge construction module constructs a semantic graph based on the power grid topology data. A semantic graph is a graph structure with physical nodes and lines as nodes and the relationships between graph elements as edges. Graph nodes are divided into two categories: graph nodes corresponding to physical nodes, which carry the electrical attributes of the node; and graph nodes corresponding to lines, which carry the voltage level and transmission parameters of the line.

[0050] Edges are divided into two categories: those where physical nodes are directly electrically connected by lines, establishing electrical connection edges between corresponding graph nodes; and those where market members to which physical nodes belong have transaction records, establishing transaction relationship edges between corresponding graph nodes. After construction, the semantic vectors after constraint projection are attached to the node representations of the corresponding graph nodes. The node representation refers to the semantic vector attached to the graph node, which represents the semantic features of the multimodal data corresponding to the graph node, and is used for subsequent semantic retrieval and comparison.

[0051] In this invention, semantic representation is a general term for semantic vectors and their attachment forms in semantic graphs and temporal event indexes, including the node representation of graph nodes and the semantic vectors corresponding to events.

[0052] Based on this, events in the time-series event index are attached to corresponding graph nodes in the semantic graph according to their associated physical nodes. Attachment refers to the process of establishing a correspondence between events and graph nodes: reading the associated physical node of an event, searching for the corresponding graph node in the semantic graph, and establishing the attachment relationship between the event and that graph node; multiple events can be attached to the same graph node, and the events attached to the same graph node are arranged in chronological order of their occurrence. If an event's associated physical node does not have a corresponding graph node in the semantic graph, the event is marked as an unattached event and stored separately, to be re-attached after the semantic graph is updated.

[0053] Furthermore, the semantic knowledge construction module supports retrieving associated events and associated graph nodes within a preset neighborhood range, starting from any graph node in the semantic graph and traversing the edges of the semantic graph. The neighborhood retrieval process includes: starting from the graph node being queried, visiting the graph nodes connected to the starting point hop by hop along the electrical connection edges and transaction relationship edges, with the visiting range not exceeding the preset neighborhood range; the preset neighborhood range is denoted as d hops, where d is a positive integer, meaning that the graph nodes reachable from the starting point along at most d edges belong to the neighborhood range, and in this embodiment, d is 2; each graph node within the neighborhood range is obtained as an associated graph node, and the events attached to each associated graph node are obtained as associated events; the retrieval results are organized and output according to the length of the connection path between each associated graph node and the starting point, from nearest to farthest. If there are no graph nodes with attached events within the neighborhood range, an empty retrieval result is output, and the node representation of the starting point is retained for subsequent analysis.

[0054] Ontology Construction and Relationship Modeling in the Electricity Market Sector: It should be noted that electricity market operations involve a large number of concepts such as market participants, trading instruments, clearing mechanisms, and constraints, and there are complex subordinate and restrictive relationships between these concepts; business rules are usually published in the form of natural language clauses, which cannot be directly understood and executed by computers.

[0055] Therefore, as an embodiment of the present invention, the ontology and rule base module organizes the ontology of the power market domain according to the concept hierarchy, and transforms the business rules into structured conditions and actions, so that market knowledge can be directly invoked by semantic retrieval and rule reasoning. The ontology of the power market domain refers to the formal representation of concepts, attributes of concepts and relationships between concepts in the power market domain; a concept node refers to a node in the domain ontology that represents a concept.

[0056] Specifically, the electricity market entity is organized according to the conceptual hierarchy of market participants, trading instruments, clearing mechanisms, and constraints: The market participant layer includes conceptual nodes such as power generation companies, electricity sales companies, and electricity users, with attributes for each conceptual node including member identification, member type, and affiliated node; the trading instrument layer includes conceptual nodes such as energy trading, ancillary service trading, and capacity trading, with attributes for each conceptual node including instrument identification, trading period, and unit of measurement; the clearing mechanism layer includes conceptual nodes such as day-ahead clearing, intraday clearing, and real-time clearing, with attributes for each conceptual node including clearing method, clearing price formation method, and clearing period; and the constraint layer includes conceptual nodes such as line transmission constraints, node power constraints, and price limit constraints, with attributes for each conceptual node including constraint object, constraint direction, and source of limit.

[0057] The conceptual nodes are connected by attributes and relationships: these relationships include subordination and constraint. For example, a subordination edge is established between the energy trading conceptual node and the day-ahead clearing conceptual node, indicating that day-ahead clearing applies to energy trading; a constraint edge is established between the line transmission constraint conceptual node and the physical node, indicating that the constraint applies to the corresponding physical node. Through these relationships, the conceptual nodes in the domain ontology correspond to the graph nodes in the semantic graph, enabling the connection between market business knowledge and power grid physical structure knowledge.

[0058] Building upon this, business rules in natural language form are transformed into structured forms of rule conditions and rule actions, and then attached to the corresponding concept nodes in the domain ontology. Rule conditions refer to the conditions required to trigger the business rule, and are composed of attributes of the attached concept node. Rule actions refer to the processing performed when the rule conditions are met, acting on the data objects associated with the attached concept node. These data objects include graph nodes in the semantic graph, events in the time-series event index, and their corresponding semantic vectors.

[0059] Taking the price limit rule as an example: the natural language clause "the declared price in the electricity market shall not exceed the price ceiling" is structured as follows: Rule condition: the commodity identifier of the trading commodity concept node is electricity trading, and the declared price is greater than the limit of the price limit constraint concept node; Rule action: mark the corresponding declared data as exceeding the limit data, and generate an exceeding the limit event and write it to the time sequence event index; this business rule is attached to the electricity trading concept node in the trading commodity layer and the price limit constraint concept node in the constraint condition layer. When the rule reasoning atomic service is called, it reads the attribute of the concept node referenced by the rule condition to perform condition judgment, and executes the rule action when the condition is met.

[0060] Specifically, this can be explained using the following example: If a market participant's declared price is 1.8 yuan / kWh, the rule inference atomic service reads the limit attribute of the price limit constraint concept node (1.5 yuan / kWh in this example), determines that the rule condition is met, executes the rule action, marks the declared data as exceeding the limit, and generates an exceeding limit event; this exceeding limit event is attached to the graph node corresponding to the physical node associated with the market participant through the time-series event index, and can be directly called by the subsequent neighborhood retrieval and analysis pipeline, realizing the linkage between business rule knowledge, semantic graph, and time-series event index.

[0061] It should be noted that the rules for electricity market operations are revised periodically as the market is developed. After the rules are revised, the semantic vectors generated by semantic mapping based on the old rules are inconsistent with the new rules. If all are remapped, the computational cost is high; if no action is taken, the analysis conclusions will use both old and new definitions.

[0062] Therefore, as an embodiment of the present invention, the ontology and rule base module locates the affected semantic vectors and performs incremental remapping when business rules change, while retaining the correspondence between versions to support historical backtracking. The effective version refers to the version of the business rule that has been effective since a certain implementation date. The effective version of each business rule is recorded by the ontology and rule base module when the rule is entered into the database. The recorded content includes the version number, implementation date and clause content.

[0063] Specifically, the old effective version of the business rule is recorded as follows: The new version is When business rules change, and The terms and conditions are broken down into clauses: let clause number j, and the resulting clauses be denoted as clause j. j=1,2, ,k, where k is the total number of clauses; and The terms are matched and compared according to their content, and the correspondence between the two versions of the terms is determined by the similarity of the text. Unable to be with Any clause that matches another clause is considered a newly added clause. Unable to be with Any clause that matches another clause will be deleted; clauses that match each other but have different content will be modified; and clauses with the same content but different numbering will only have their numbering updated.

[0064] In some implementations, based on the concept node attached to the modified clause, the following location is performed: The concept node attached to the modified clause is read; the semantic graph node associated with the concept node is determined along the association relationship between the concept node and the semantic graph node in the domain ontology; and the event generated by the rule action of the business rule in the time-series event index is determined based on the data object associated with the concept node; the node representation of the determined graph node and the semantic vector corresponding to the determined event are determined as the affected semantic vector. That is, "affected" is determined by the association path of "modified clause → attached concept node → associated graph node and event → corresponding semantic vector," without subjective judgment.

[0065] Furthermore, the scope of incremental remapping is limited to the identified affected semantic vectors, while the remaining semantic vectors remain unchanged: the semantic mapping module is triggered, and the new effective version is applied. The rules and semantics are redefined, and the affected semantic vectors are remapped and constrained. The remapped semantic vectors replace the original semantic vectors, and their corresponding effective versions are recorded. After the incremental remapping is completed, the dynamic orchestration service module re-executes the corresponding analysis pipeline based on the semantic representation after incremental remapping, so that the subsequent analysis conclusions are consistent with the current rule standards.

[0066] As an embodiment of the present invention, the ontology and rule base module records the correspondence between clauses in different effective versions, and the correspondence includes: the modified clauses in and The clause number corresponds to the newly added clause, and the new clause is marked in [the relevant section]. There is no corresponding clause in the document; the deleted clause is marked as such. There is no corresponding clause in this document.

[0067] Based on the business rule version that was in effect when the historical data was generated, a retrospective query is performed on the semantic representation corresponding to the historical data: when querying historical data, the business rule version that was in effect when the data was generated is determined according to the time of generation, the clause content and the clause correspondence of that version are retrieved, and the semantic representation corresponding to the historical data is interpreted according to the rule semantics of that version, so that the historical analysis conclusions can be reproduced according to the rule caliber that was effective at that time.

[0068] This can be further explained with the following example: An embodiment of the aforementioned price limit rule: The effective version of this rule is... Revised to The price ceiling has been adjusted from 1.5 yuan / kWh to 1.2 yuan / kWh; and By breaking down and comparing the clauses, the clause containing the price cap can be determined. For the revised terms; Read The concept node for the price limit constraint is linked, and the semantic graph nodes related to electricity trading are determined along the correlation, and the over-limit events previously generated by the rule actions are identified; the node representations of the above graph nodes and the semantic vectors corresponding to the over-limit events are determined as the affected semantic vectors; according to... The rule semantics perform incremental remapping on the above semantic vectors, updating the limit attribute in the node representation from 1.5 yuan / kWh to 1.2 yuan / kWh, while the remaining semantic vectors remain unchanged; records exist and The corresponding clauses in the document; when querying historical analysis conclusions generated before the limit adjustment, the conclusions should be based on the ones in effect when the data was generated. The interpretation is to reproduce the basis for the over-limit judgment at that time using the limit of 1.5 yuan / kWh.

[0069] Multimodal semantic modeling and fusion: It should be noted that different business analysis tasks, such as anomaly detection, electricity price forecasting, and settlement verification, require different data processing procedures. If a fixed processing procedure is built for each task, the development and maintenance costs will be high and it will not be reusable as the rules evolve.

[0070] Therefore, as an embodiment of the present invention, the dynamic orchestration service module encapsulates semantic retrieval, feature fusion, and rule reasoning into atomic services, and dynamically combines them according to task type to generate an analysis pipeline. Here, an atomic service refers to the smallest functional service with definite input and output that can be independently invoked; an analysis pipeline refers to an end-to-end processing flow composed of several atomic services connected in execution order.

[0071] In S801, specifically, after receiving a business analysis task, the dynamic orchestration service module parses the task content: identifying the task type, which includes anomaly detection, operation monitoring, electricity price forecasting, and settlement verification; extracting the analysis objects, which include target physical nodes, target time periods, and target data ranges; the parsing method includes word segmentation and keyword matching of the task text, determining the task type based on a preset task type keyword table, and extracting the analysis objects based on the graph node identifiers and time-series event indexes in the semantic graph. If the task content cannot match any task type, the business analysis task is marked as an unparseable task and returned for manual confirmation, and does not proceed to the pipeline generation stage.

[0072] In S802, the atomic services to be invoked and their execution order are determined according to the pre-configured correspondence between task types and atomic service combinations, and an analysis pipeline is generated. The correspondence is pre-stored in the dynamic orchestration service module in configuration form. The corresponding relationships include: for anomaly detection tasks, the semantic retrieval atomic service, feature fusion atomic service, and rule reasoning atomic service are executed sequentially; for operation monitoring tasks, the semantic retrieval atomic service and rule reasoning atomic service are executed sequentially; for electricity price prediction tasks, the semantic retrieval atomic service and feature fusion atomic service are executed sequentially; and for settlement verification tasks, the semantic retrieval atomic service and rule reasoning atomic service are executed sequentially. During pipeline generation, cataloged data is retrieved from metadata records based on the analysis object, serving as input data for each atomic service.

[0073] Furthermore, the analysis pipeline is executed, and the atomic services are called sequentially according to the execution order. When the feature fusion atomic service is called, the following processing is performed: intermodal attention interaction is performed on the semantic vectors of different modalities after constraint projection. Intermodal attention interaction refers to calculating the relevance score between the semantic vector of any modality and the semantic vectors of other modalities, so that the semantic vector of each modality is fused with the relevant information of other modalities. Let the modal number be , F represents the total number of modes participating in the fusion; The task text of the business analysis task is mapped to the semantic space using the same semantic encoding method as the rule text to obtain the task semantic vector. The cosine similarity between the semantic vector of each modality and the task semantic vector is calculated as the relevance score of that modality relative to the current business analysis task. The higher the score, the stronger the relevance of that modality to the current task. The natural exponent of the relevance score of each modality is then divided by the sum of the natural exponents of the relevance scores of all modalities to obtain the contribution weight of each modality. Each contribution weight is positive and the sum is 1. The semantic vectors of different modalities are weighted and summed according to their contribution weights to obtain the fused intermediate vector. The fused intermediate vector is then projected onto the physical manifold to obtain the fused semantic vector.

[0074] The reason for using cosine similarity as the relevance score is that cosine similarity measures the relevance of vectors by angle, eliminating the influence of differences in the magnitude of vectors of different modalities, and making modal scores of different magnitudes comparable. The reason for normalization after taking the natural exponent is that each contribution weight is positive and the sum is 1. The weighted sum does not change the overall scale of the semantic vector. The fused intermediate vector is still a vector in the semantic space and can be directly projected back onto the physical manifold. At the same time, the exponential operation amplifies the score difference, so that the task-related dominant modalities play a leading role in the fusion result.

[0075] The purpose of reprojection is to ensure that the fused semantic vector still satisfies the physical consistency of the power grid, as weighted summation may cause the fusion result to deviate from the physical manifold. The fused semantic vector serves as the input for subsequent rule-based reasoning or prediction processing.

[0076] In some implementations, during the execution of the analysis pipeline, the dynamic orchestration service module records the source data, mapping path, and the version of the business rules used as traceability information for the data being called. The traceability information includes: source data identifier (corresponding to the data source field in the metadata record), mapping path (the names and order of the atomic services through which the data is called), and the version number of the business rules used. The analysis results are output along with this traceability information; upon receiving a traceability request for the analysis results, the traceability information is traced backwards: the input data of each processing stage is determined in reverse along the mapping path until the original data is reached, and the clause content of the business rule version used is retrieved to verify the data source and rule basis of the analysis conclusions.

[0077] This can be further explained using the following example, taking the "Node 1 Price Abnormal Fluctuation Verification" task as an example: S801 Parses to obtain the task type as anomaly detection, the analysis object as Node 1 and the target time period; S802 Checks the configuration table to determine the atomic service combination as semantic retrieval → feature fusion → rule reasoning, and retrieves the data associated with the corresponding metadata record based on the analysis object; S803 Executes the pipeline—the semantic retrieval atomic service performs neighborhood retrieval starting from the graph node corresponding to Node 1, obtaining the output mutation event and the blocking event; the feature fusion atomic service performs attention interaction and weighted summation on the semantic vectors of the three modalities of transaction, measurement, and topology, and inputs the fused semantic vector obtained after reprojection into the rule reasoning atomic service; the rule reasoning atomic service determines that the price exceeds the limit according to the price limit rule and outputs the anomaly verification conclusion; S804 Records the source data identifier of the conclusion, the sequence of atomic services passed through, and the effective version number of the price limit rule. The analysis conclusion carries the above-mentioned traceability information and can be traced back to the original declaration data and rule clauses.

[0078] In another embodiment of the present invention, a method for constructing a semantic mapping architecture for a large-scale electricity market model is proposed, which implements the above-mentioned semantic mapping architecture for a large-scale electricity market model, including the following steps: S1 accesses structured transaction data, time-series measurement data, grid topology data, and rule texts from the electricity market, and catalogs the accessed data in a unified manner according to a pre-established metadata system; S2, map the various types of data into the same semantic space to obtain semantic vectors, and based on the physical manifold defined by the power grid physical equations, perform constrained projection on the semantic vectors carrying electrical quantity semantics (including semantic vectors corresponding to time series measurement data and semantic vectors generated by fusion), so that the constrained projected semantic vectors are located on the physical manifold. S3. Construct a semantic graph based on the power grid topology data, with physical nodes and lines as graph nodes and electrical connection relationships and transaction relationships as edges. Attach the semantic vectors after constraint projection to the node representations of the corresponding graph nodes, and generate a time-series event index from the semantic vectors corresponding to the time-series measurement data after constraint projection. S4, construct the ontology and business rule library for the power market field, and record the effective versions of each business rule; S5 parses business analysis tasks, dynamically orchestrates and calls atomic services of semantic retrieval, feature fusion and rule reasoning to generate analysis pipelines, and writes the analysis results back to the semantic graph and time-series event index; S6, when business rules change, compare the differences between the terms of different effective versions, locate the semantic vectors affected by the differences, perform incremental remapping on the affected semantic vectors, and re-execute the corresponding analysis pipeline based on the semantic representation after incremental remapping.

[0079] S2 constrains the projection of semantic vectors carrying electrical quantities onto the physical manifold defined by the power grid physical equations, including: determining the physical manifold by the equality and inequality relationships between the node power balance equations and the line transmission constraints; calculating the deviation of the semantic vectors relative to the physical manifold; and projecting the semantic vectors with deviations onto the physical manifold along the normal of the physical manifold.

[0080] In summary, this invention uses a metadata system to uniformly catalog multi-source heterogeneous data in the power market, enabling data to have a unified basis for association and to be automatically retrieved by business analysis tasks. In the semantic mapping stage, the semantic vector is constrained and projected based on the physical manifold defined by the power grid physical equations, and the original semantic features are preserved with the minimum correction amount, so that the semantic representations entering the downstream analysis correspond to the physically reachable power grid state, ensuring that the analysis conclusions have a physical basis. By using semantic graphs and time-series event indexes organized hierarchically according to transaction beats, structured organization of event knowledge and cross-beat traceability are achieved. By versioning business rules, affected semantic vectors are located along the associated paths and incremental remapping is performed when rules change, avoiding the computational overhead of full remapping, ensuring consistency of analysis conclusions and that historical conclusions can be reproduced according to the rules at that time. Through dynamic orchestration of atomic services and reprojection of fused vectors, the analysis process is reusable and the results meet physical consistency, and full traceability is supported.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A semantic mapping architecture for a large-scale electricity market model, characterized in that, The architecture includes: The data access module is used to access structured transaction data, time-series measurement data, power grid topology data and rule texts from the electricity market, and to uniformly catalog all types of data according to a pre-established metadata system. The semantic mapping module is used to map various types of data to the same semantic space to obtain semantic vectors, and based on the physical manifold defined by the power grid physical equations, to constrain the semantic vectors carrying electrical quantity semantics so that the constrained and projected semantic vectors lie on the physical manifold. The semantic knowledge construction module is used to construct a semantic graph based on the power grid topology data, with physical nodes and lines as graph nodes and electrical connection relationships and transaction relationships as edges. It attaches the semantic vectors after constraint projection to the node representations of the corresponding graph nodes, and generates a time-series event index from the semantic vectors corresponding to the time-series measurement data after constraint projection. The ontology and rule base module is used to store the ontology and business rules in the power market field, record the effective versions of each business rule, and when the business rules are changed, compare the differences in the terms of different effective versions to locate the semantic vectors affected by them, and trigger the semantic mapping module to perform incremental remapping on the affected semantic vectors. The dynamic orchestration service module is used to parse business analysis tasks, orchestrate and call atomic services of semantic retrieval, feature fusion and rule reasoning to generate analysis pipelines, write the analysis results back to the semantic graph and time-series event index, and re-execute the corresponding analysis pipelines based on the semantic representation after incremental remapping.

2. The semantic mapping architecture for a large-scale electricity market model according to claim 1, characterized in that: The establishment of the metadata system includes: sorting out the data requirements of typical anomaly analysis scenarios in the power market, including abnormal price fluctuations, transaction manipulation and settlement anomalies; extracting data types, data sources, related business rules and analysis scenario identifiers from the data requirements as metadata fields; establishing the relationship between the metadata fields to form the metadata system. The data access module catalogs various types of data in a unified manner according to a pre-established metadata system, including: parsing the data type and data source of the access data, generating metadata records containing the metadata fields for the access data, and associating the access data with the corresponding business analysis tasks according to the analysis scenario identifier in the metadata records.

3. The semantic mapping architecture for a large-scale electricity market model according to claim 1, characterized in that: The semantic mapping module performs constrained projection on the semantic vector carrying the semantics of electrical quantities, including: The physical manifold is determined by the equality and inequality relationships between the node power balance equation and the line transmission constraints; the deviation of the semantic vector relative to the physical manifold is calculated, and the semantic vector with deviation is projected onto the physical manifold along the normal of the physical manifold.

4. The semantic mapping architecture for a large-scale electricity market model according to claim 1, characterized in that: The semantic knowledge construction module generates a time-series event index by including the following steps: S401, the semantic vectors corresponding to the time series measurement data after constraint projection are arranged into a vector sequence according to the occurrence time, and the vector sequence is divided according to a preset time window; S402, compare the semantic vectors in adjacent time windows, identify the time windows in which the changes in semantic vectors meet the preset event conditions as events, and determine the event elements of the event, the event elements including the time of occurrence, event type, associated physical node and associated measurement; S403, Establish a time-series event index for the identified events according to the time of occurrence; S404, the time-series event index is organized hierarchically according to the day-ahead, intraday, real-time and settlement transaction rhythm of the electricity market, and the event views corresponding to different transaction rhythms are associated with the mapping chain of planned value, deviation and settlement amount.

5. The semantic mapping architecture for a large-scale electricity market model according to claim 4, characterized in that: The semantic knowledge construction module constructs a semantic graph, including: Using physical nodes and lines as graph nodes, if physical nodes are directly electrically connected by lines, establish electrical connection edges between the corresponding graph nodes. If there are transaction records between market members to which physical nodes belong, establish transaction relationship edges between the corresponding graph nodes. And attach the semantic vector after constraint projection to the node representation of the corresponding graph node. The semantic knowledge construction module is also used to attach events in the time-series event index to the corresponding graph nodes in the semantic graph according to the associated physical nodes, and supports retrieving associated events and associated graph nodes in a preset neighborhood range along the edges of the semantic graph, starting from any graph node in the semantic graph.

6. The semantic mapping architecture for a large-scale electricity market model according to claim 1, characterized in that: The electricity market ontology is organized according to the conceptual hierarchy of market participants, trading instruments, clearing mechanisms and constraints, with each conceptual node connected by attributes and relationships. The business rules are represented in a structured form of rule conditions and rule actions, and are attached to the corresponding concept nodes in the domain ontology. The rule conditions reference the attributes of the attached concept nodes, and the rule actions act on the data objects associated with the attached concept nodes.

7. The semantic mapping architecture for a large-scale electricity market model according to claim 6, characterized in that: The ontology and rule base module includes: The business rules of different effective versions are broken down into clauses and compared one by one to determine the added, deleted and modified clauses; based on the concept nodes attached to the modified clauses, the semantic graph nodes associated with the concept nodes and the events in the time sequence event index are determined, and the node representation of the corresponding graph node and the semantic vector corresponding to the event are determined as the affected semantic vectors. The execution scope of the incremental remapping is the determined affected semantic vectors, while the remaining semantic vectors remain unchanged; the ontology and rule base module records the correspondence between the terms in the different effective versions, and performs a backtracking query on the semantic representation corresponding to the historical data according to the business rule version that took effect when the historical data was generated.

8. The semantic mapping architecture for a large-scale electricity market model according to claim 1, characterized in that: The dynamic orchestration service module generates the analysis pipeline by including the following steps: S801, parse the business analysis task to obtain the task type and analysis object of the business analysis task; S802, according to the pre-configured correspondence between task types and atomic service combinations, determine the semantic retrieval, feature fusion and rule reasoning atomic services to be invoked and their execution order, and generate the analysis pipeline; S803, execute the analysis pipeline, wherein when the feature fusion atomic service is invoked, perform intermodal attention interaction on the different modal semantic vectors after constraint projection, calculate the contribution weight of each modal semantic vector relative to the current business analysis task, perform weighted summation on the different modal semantic vectors according to the contribution weight, and project the weighted summation result back onto the physical manifold to obtain the fused semantic vector; S804, during the execution of the analysis pipeline, the source data, mapping path and the version of the business rule on which the called data are recorded as traceability information, so that the output analysis results carry the traceability information and support tracing back to the original data and business rule terms along the traceability information.

9. A method for constructing a semantic mapping architecture for a large-scale electricity market model, the method being applicable to the semantic mapping architecture for a large-scale electricity market model as described in any one of claims 1-8, characterized in that, Includes the following steps: S1 accesses structured transaction data, time-series measurement data, grid topology data, and rule texts from the electricity market, and catalogs the accessed data in a unified manner according to a pre-established metadata system; S2, map the various types of data into the same semantic space to obtain semantic vectors, and perform constrained projection on the semantic vectors carrying electrical quantity semantics based on the physical manifold defined by the power grid physical equations, so that the constrained and projected semantic vectors are located on the physical manifold; S3. Construct a semantic graph based on the power grid topology data, with physical nodes and lines as graph nodes and electrical connection relationships and transaction relationships as edges. Attach the constrained projected semantic vectors as node representations of the corresponding graph nodes, and generate a time-series event index from the semantic vectors corresponding to the constrained projected time-series measurement data. S4, construct the ontology and business rule library for the power market field, and record the effective versions of each business rule; S5, parse the business analysis task, dynamically orchestrate and call the atomic services of semantic retrieval, feature fusion and rule reasoning to generate the analysis pipeline, and write the analysis results back to the semantic graph and time-series event index; S6. When business rules change, compare the differences between the terms of different effective versions, locate the semantic vectors affected by the differences in the terms, perform incremental remapping on the affected semantic vectors, and after the incremental remapping is completed, re-execute the corresponding analysis pipeline based on the semantic representation after incremental remapping.

10. The method for constructing a semantic mapping architecture for a large-scale electricity market model according to claim 9, characterized in that, The constraint projection of the semantic vector carrying electrical quantity semantics based on the physical manifold defined by the power grid physical equations in S2 includes: determining the physical manifold by the equality and inequality relationship between the node power balance equation and the line transmission constraint; calculating the deviation of the semantic vector relative to the physical manifold; and projecting the semantic vector with deviation onto the physical manifold along the normal of the physical manifold.

Images