A power system operation state prediction method and device, electronic equipment and storage medium

Abstract

The application discloses a power system operation state prediction method and device, electronic equipment and storage medium, and belongs to the technical field of power system operation state prediction, and the method comprises the steps of obtaining a physical equipment set, structured operation data and unstructured event data of a target power supply area; extracting a main frequency feature according to the structured operation data, reconstructing a main trend of an operation state; using a large language model to analyze the unstructured event data, generating an operation influencing factor, and obtaining an event-driven residual error through a residual mapping model; fusing the main trend of the operation state and the event-driven residual error to generate a target operation state prediction result; generating a scheduling control strategy according to the prediction result, and executing scheduling control on each physical equipment. Therefore, by implementing the application, the problem that the prediction result is easy to deviate from the real physical topology environment and difficult to provide fine control basis for the scheduling end in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the field of power system operation status prediction technology, specifically to a power system operation status prediction method, device, electronic equipment, and storage medium. Background Technology

[0002] Predicting the operating status of power systems is a crucial foundation for ensuring the safe, stable, and efficient dispatch of modern smart grids. With the increasing complexity of new power system architectures and the deepening coupling between nodes, accurately and proactively sensing the evolution trends of various system state variables can provide forward-looking decision support for grid dispatching.

[0003] However, existing state prediction methods suffer from technical shortcomings, such as insufficient perception of the impact of sudden events and inadequate depth of heterogeneous data fusion, leading to a decrease in prediction accuracy under complex external disturbances. This is because current prediction schemes primarily rely on structured historical runtime sequence data to fit basic numerical trends, but struggle to effectively handle and quantify the sudden impacts of unstructured text events such as equipment failure reports and extreme weather warnings on the power grid. Furthermore, due to the lack of deep information extraction and physical space mapping mechanisms, existing technologies cannot decompose and associate unstructured qualitative semantic events with specific affected physical devices. This makes it difficult to calculate the specific quantitative residuals of the events on various state variables, and consequently, to achieve vector superposition and fusion with structured main trends. Consequently, prediction results easily deviate from the actual physical device topology environment, making it difficult to provide the host scheduling terminal with precise control basis down to the specific device level. Summary of the Invention

[0004] This invention provides a method, device, electronic device, and storage medium for predicting the operating status of a power system. It can solve the problems in the prior art of insufficient perception of the impact of sudden events and insufficient depth of heterogeneous data fusion, making it difficult to decompose and quantify unstructured qualitative text events and associate them with specific physical devices, resulting in prediction results that are easily deviated from the real physical topology environment and are difficult to provide a fine control basis for the dispatching end.

[0005] An embodiment of the present invention provides a method for predicting the operating state of a power system, comprising: Acquire the set of physical devices in the target power supply area, the structured operational data of the target power supply area within the target time period, and the unstructured event data; The main frequency characteristics of each state variable are extracted from the structured operation data, and the main trend of the operation status of the target power supply area is reconstructed based on the main frequency characteristics. Unstructured event data is input into a large language model to perform information extraction, generating structured event knowledge units containing device description text of affected devices; the matching relationship between the device description text and the device identifier of each physical device in the physical device set is calculated to obtain the association weight; the association weight is aggregated and calculated based on the device identifier of each physical device and various state variables to generate the operation impact factor; the operation impact factor is processed according to the preset residual mapping model to output the event-driven residual; The main trend of the operating status and the event-driven residual are superimposed into a multi-dimensional vector to generate the target operating status prediction result of the target power supply area. Based on the predicted values ​​of various state variables in the target operating state prediction results, a scheduling control strategy is generated by matching the preset scheduling risk rules to execute scheduling control on each physical device.

[0006] Furthermore, the system acquires the physical device set of the target power supply area, structured operational data of the target power supply area within the target time period, and unstructured event data, including: Obtain the scheduling object ledger and equipment ledger of the target power supply area, perform object standardization processing on the scheduling object ledger and equipment ledger, and generate a physical equipment set; Acquire historical load data and meteorological characteristic data of the target power supply area within the target time period, unify the time interval and aggregate the granularity of the historical load data and meteorological characteristic data, and generate structured operation data; Receive alarm messages from the preset device fault alarm bus and generate device fault text; Extract warning information from preset weather warning information sources and generate weather warning text; Extract the main text of the pre-set scheduling operation announcement webpage to generate the scheduling announcement text; Time alignment is performed on equipment failure texts, weather warning texts, and dispatch announcement texts to generate unstructured event data.

[0007] Furthermore, the various state variables include electricity load, net load, and reserve requirement; the structured operating data includes historical operating data corresponding to the electricity load, net load, and reserve requirement, respectively. The dominant frequency characteristics of each state variable are extracted from the structured operational data, and the main trend of the operational state of the target power supply area is reconstructed based on the dominant frequency characteristics, including: For each state variable, historical operation data corresponding to the current state variable is extracted from the structured operation data; frequency domain decomposition is performed on the historical operation data to generate multiple frequency components and the amplitude parameters corresponding to each frequency component; The multiple frequency components are sorted in descending order of amplitude parameters, and a preset number of frequency components are extracted from the sorted multiple frequency components as the dominant frequency feature of the current state variable. Historical operating data is input into a pre-established trend regression prediction model to generate the target periodic component parameters of the current state variable; Based on the main frequency characteristics and target periodic component parameters, the time series signal is reconstructed to generate the basic numerical main trend of the current state variables; Extract the extreme values ​​of the basic numerical main trend within the target time period, and determine the extreme value time period corresponding to the extreme value; Extract the data sequence corresponding to the extreme value period from the basic numerical main trend to generate the extreme value period main trend of the current state variable; After generating the basic numerical main trend and extreme value period main trend of all state variables, the basic numerical main trend of each state variable is weighted and subjected to discrete interval mapping based on the preset risk weight and preset safety threshold to generate the risk level main trend of the target power supply area. The main trend of the basic values ​​of each state variable, the main trend of the extreme value period of each state variable, and the main trend of the risk level are combined to generate the main trend of the operating status of the target power supply area.

[0008] Furthermore, unstructured event data is input into a large language model to perform information extraction, generating structured event knowledge units containing device description text of affected devices; the matching relationship between the device description text and the device identifiers of each physical device in the physical device set is calculated to obtain the association weight, including: Based on the preset prompt word template, unstructured event data is input into a large language model for information extraction, generating structured event knowledge units; among them, the structured event knowledge units include event type information and device description text of multiple affected devices; Extract the device type attribute and device identifier of each physical device from the physical device set; Based on the preset type matching rules, the event type information is compared and verified with the device type attributes of each physical device to generate the type matching degree between the event type information and the device type attributes of each physical device. Based on the preset text representation model, feature extraction is performed on the device identifier of each physical device to generate a physical device feature vector corresponding to the device identifier of each physical device. For each device description text, features are extracted from the current device description text based on the text representation model to generate the affected device feature vector corresponding to the current device description text. The spatial distance between the feature vector of the affected device and the feature vector of each physical device is calculated one by one to generate the entity link similarity between the current device description text and the device identifier of each physical device. Numerical aggregation calculations are performed based on entity link similarity and type matching to generate association weights between the current device description text and the device identifiers of each physical device.

[0009] Furthermore, the structured event knowledge unit also includes the credibility of the event source and the confidence level of information extraction; Based on the device identifiers and various state variables of each physical device, the associated weights are aggregated and calculated to generate operational impact factors; the operational impact factors are processed according to a pre-defined residual mapping model to output event-driven residuals, including: Based on the association weight and event type information, determine the conflict event characteristics corresponding to the device identifiers of each physical device; Based on the credibility of the event source and the confidence of information extraction, the conflict event features corresponding to the device identifiers of each physical device are harmonized and weighted to generate conflict handling coefficients corresponding to the device identifiers of each physical device. Based on the equipment identifiers of each physical device and various status variables, the associated weights are aggregated and calculated according to the conflict handling coefficients to generate operational impact factors corresponding to each status variable. Each operational influencing factor is input into a preset residual mapping model for feature mapping processing to generate initial event-driven residuals; Perform logic violation checks on the initial event-driven residuals and generate penalty bias values; Numerical correction operations are performed based on the initial event-driven residual and the penalty bias value to generate the event-driven residual.

[0010] Furthermore, the main trend of the operating state and the event-driven residuals are superimposed using a multi-dimensional vector operation to generate the target operating state prediction results for the target power supply area, including: Vector bias fusion calculation is performed on the main trend of the running state and the event-driven residuals to generate the initial running state prediction results; According to the preset operation state mapping rules, numerical boundary constraints and discrete dimension mapping processing are performed on the initial operation state prediction results to generate the target operation state prediction results for the target power supply area; wherein, the target operation state prediction results include the predicted values ​​corresponding to each state variable.

[0011] Furthermore, based on the predicted values ​​of various state variables in the target operating state prediction results, a pre-defined scheduling risk rule is matched to generate a scheduling control strategy to execute scheduling control on each physical device, including: According to the preset scoring mapping rules, the predicted values ​​of each state variable in the target operation state prediction results are normalized and mapped respectively to generate a single-dimensional risk score corresponding to each state variable. A weighted aggregation calculation is performed on the single-dimensional risk scores corresponding to each state variable to generate a comprehensive risk assessment value. Based on the preset scheduling risk rules, the comprehensive risk assessment values ​​are compared across intervals to determine the scheduling security risk level corresponding to the comprehensive risk assessment values. The preset scheduling risk rules include risk classification thresholds and candidate scheduling control strategies bound to each scheduling security risk level. Based on the scheduling safety risk level and the predicted values ​​of various state variables, candidate scheduling control strategies are extracted from the scheduling risk rules and used as scheduling control strategies to perform scheduling control on each physical device.

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

[0013] One embodiment of the present invention provides a power system operation status prediction device, comprising: a data acquisition module, a main trend reconstruction module, an event-driven processing module, a prediction result generation module, and a dispatch control module; The data acquisition module is used to acquire the set of physical devices in the target power supply area, the structured operation data of the target power supply area within the target time period, and the unstructured event data. The main trend reconstruction module is used to extract the main frequency characteristics of each state variable based on the structured operation data, and reconstruct the main trend of the operation status of the target power supply area based on the main frequency characteristics. The event-driven processing module is used to input unstructured event data into a large language model to perform information extraction, generate structured event knowledge units containing device description text of affected devices; calculate the matching relationship between the device description text and the device identifiers of each physical device in the physical device set to obtain the association weights; aggregate and calculate the association weights according to the device identifiers of each physical device and various state variables to generate the operation impact factor; process the operation impact factor according to the preset residual mapping model and output the event-driven residual. The prediction result generation module is used to perform multi-dimensional vector superposition operation on the main trend of the operating state and the event-driven residual to generate the target operating state prediction result of the target power supply area. The scheduling control module is used to match the predicted values ​​of various state variables in the target operating state prediction results with preset scheduling risk rules to generate a scheduling control strategy for executing scheduling control on each physical device.

[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 the power system operating state prediction method according to any one of 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 system operation state prediction 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 predicting the operating state of a power system. The method acquires a set of physical devices, structured operating data, and unstructured event data for a target power supply area; extracts the dominant frequency features of state variables from the structured operating data to reconstruct the main trend of the operating state; extracts information from the unstructured event data using a large language model to generate structured event knowledge units, and generates operating influencing factors by combining device identifier matching relationships, obtaining event-driven residuals through a residual mapping model; superimposes the main trend of the operating state with the event-driven residuals to generate a target operating state prediction result; matches scheduling risk rules based on the prediction result to generate a scheduling control strategy, and executes scheduling control on each physical device.

[0019] This invention solves the problem in existing technologies of decomposing, quantifying, and associating qualitative textual events with specific physical devices by inputting unstructured event data into a large language model to extract affected devices and generating event-driven residuals by combining device matching rules with a residual mapping model. Furthermore, this invention performs multi-dimensional vector superposition operations on the main trend of operating status reconstructed from structured data and the aforementioned event-driven residuals, overcoming the problem of insufficient depth in heterogeneous data fusion under sudden event impacts. This deep feature fusion prediction result can directly match scheduling risk rules, solving the problem of existing prediction results being detached from the real physical topology environment, and providing a refined and reliable basis for scheduling and control of physical devices within the power grid. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a method for predicting the operating status of a power system according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a power system operation status prediction 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 problems in existing technologies such as insufficient perception of the impact of sudden events and inadequate depth of heterogeneous data fusion, making it difficult to decompose and quantify unstructured qualitative text events and associate them with specific physical devices, resulting in prediction results that easily deviate from the real physical topology and fail to provide a basis for refined control at the dispatching end, an embodiment of the present invention provides a method for predicting the operating state of a power system, comprising at least the following steps: Step S1: Obtain the set of physical devices in the target power supply area, the structured operation data of the target power supply area within the target time period, and the unstructured event data. In a preferred embodiment, the acquisition includes a set of physical devices in the target power supply area, structured operational data of the target power supply area within a target time period, and unstructured event data, including: Obtain the scheduling object ledger and equipment ledger of the target power supply area, perform object standardization processing on the scheduling object ledger and equipment ledger, and generate a physical equipment set; Acquire historical load data and meteorological characteristic data of the target power supply area within the target time period, unify the time interval and aggregate the granularity of the historical load data and meteorological characteristic data, and generate structured operation data; Receive alarm messages from the preset device fault alarm bus and generate device fault text; Extract warning information from preset weather warning information sources and generate weather warning text; Extract the main text of the pre-set scheduling operation announcement webpage to generate the scheduling announcement text; Time alignment is performed on equipment failure texts, weather warning texts, and dispatch announcement texts to generate unstructured event data.

[0024] Specifically, regarding the execution steps for acquiring the physical equipment set of the target power supply area, the structured operational data of the target power supply area within the target time period, and the unstructured event data, the prediction task configuration parameters are first obtained. These parameters include the target prediction period, prediction granularity, rolling training window, and the target operational state vector to be predicted. The components of the target operational state vector cover electricity load, net load, reserve demand, peak load period, valley load period, and dispatch safety risk level. Electricity load represents the overall electricity demand level within the power supply area. Net load represents the remaining value of the electricity load after deducting the available output of new energy sources. Reserve demand represents the regulation capacity necessary to ensure the safe operation of the power grid. Peak load period and valley load period indicate the time intervals corresponding to the maximum and minimum values ​​of the load sequence, respectively. The dispatch safety risk level comprehensively reflects the safety margin of the external dispatch constraint verification dimension.

[0025] After clarifying the configuration parameters for the forecasting task, the scheduling object ledger and equipment ledger for the target power supply area are obtained. The scheduling object ledger and equipment ledger specifically include basic ledger information on dimensions such as load zoning, generator sets, new energy power plants, energy storage stations, and demand response resources. The scheduling object ledger and equipment ledger undergo object standardization processing to generate a physical equipment set. The physical equipment set consists of a standardized object table and an object association table. The standardized object table records structured fields such as unique object identifier, object type, standard name, alias list, region identifier, resource type, rated capacity, and availability status. The object association table records the association starting object identifier, association target object identifier, business association type, and association weight. The business association type is clearly divided into categories such as belonging to the same power supply area, participating in the same standby arrangement, affected by the same maintenance plan, and belonging to the same demand response resource group. Through the aforementioned object standardization processing, a power grid topology physical equipment index with a strict hierarchical relationship is constructed, providing a benchmark reference library for subsequently establishing a numerical mapping channel from unstructured semantic text to specific topology nodes.

[0026] Subsequently, historical load data and meteorological characteristic data for the target power supply area within the target time period are acquired. In the specific execution process, historical electricity load, actual output of new energy sources, predicted output of new energy sources, standby arrangement execution records, equipment availability status, energy storage charging and discharging status, and calendar attributes are obtained from the preset dispatch automation control terminal, energy management business platform, and load forecast data terminal. Temperature, humidity, wind speed, irradiance, and rainfall are obtained as meteorological characteristic data from the preset meteorological numerical forecast business platform. Missing value imputation and outlier removal operations are performed on the historical load data and meteorological characteristic data, and time interval unification and granularity aggregation are performed on the historical load data and meteorological characteristic data to generate structured operational data. In specific implementation, the above missing value imputation operation can be implemented using linear interpolation, mean imputation, or polynomial fitting imputation; the above outlier removal operation can be implemented using the 3σ criterion (Laida criterion), box plot method, or isolated forest algorithm, thereby effectively ensuring the quality of the input data of the underlying time series model. For data with hourly accuracy, numerical aggregation is strictly performed according to the configured prediction granularity on a daily, weekly, or monthly basis. When processing data spanning multiple years, an automatic alignment operation is performed to directly remove the additional hourly data related to February 29th in leap years, or to uniformly convert it into a standard sequence of 8760 hours per year through a linear interpolation algorithm, ensuring that the input dimensions of the subsequent cycle main trend analysis model remain absolutely consistent.

[0027] The system synchronously performs unstructured text acquisition, employing distributed message subscription and web crawling technologies to obtain multi-source heterogeneous unstructured events. Specific acquisition paths include: using an application programming interface (API) to subscribe in real-time to a pre-set equipment fault alarm bus within the power grid business network, receiving alarm messages from the pre-set equipment fault alarm bus, and generating equipment fault text; using automated scripts to periodically crawl simplified information aggregation source data format warning information released by pre-set meteorological warning information sources, generating meteorological warning text; and using regular expressions in natural language processing technology to extract text slices from the hypertext markup language format pages generated by pre-set scheduling operation announcement web pages, generating scheduling announcement text. Based on the same distributed acquisition architecture, heterogeneous event texts from maintenance notification databases, natural disaster information databases, and public energy supply event reporting databases are collected in parallel.

[0028] The original collected texts, such as equipment failure texts, weather warning texts, and dispatch announcement texts, undergo deduplication, language conversion, text content extraction, and noise removal. Then, time alignment is performed on these texts to generate unstructured event data. In the time alignment process, a dedicated candidate event time window is constructed for each target time. If the extracted text record represents an event that has already occurred and whose impact is ongoing, the time alignment algorithm forces the event occurrence time to be no later than the corresponding target time. If the extracted text record represents a future maintenance event that has already been announced, a weather warning event that has already been issued, or a dispatch operation announcement event that has already been issued, the impact range of the corresponding event must overlap with the corresponding target time. The specific length of the candidate event time window is strictly limited by preset window parameters. In a specific embodiment, the candidate event time window length for weather warning texts is configured to be 1 to 7 days; the candidate event time window length for equipment failure texts is configured to be 1 to 14 days; and the candidate event time window length for planned maintenance notification texts is configured to be from 30 days before the maintenance start date to the maintenance end date.

[0029] After completing the above data acquisition and alignment steps, the feature data from each dimension are combined to construct a standard model training sample set. The specific mathematical expression for the model training sample set is as follows: In the above mathematical expression, Representing the target moment The corresponding target running state vector; Represents electrical load; Represents net load; Represents backup demand; Represents peak load periods; Represents the period of lowest load; This represents the level of safety risk in dispatching; The representative model training sample set; Representing the target moment The corresponding structured operational feature vector after granular aggregation processing; Representing the target moment The corresponding unstructured event data set generated after time alignment; This represents the verification result of external scheduling constraints as known input conditions; This represents the set of training times.

[0030] Through the above-mentioned comprehensive, multi-granular data acquisition and standardized spatiotemporal alignment steps with strict topological constraints, the dimensional barriers between multimodal heterogeneous data have been completely eliminated. This provides a high-quality underlying data foundation for the subsequent accurate decomposition, quantification, and anchoring of qualitative unstructured text events to the real physical power grid equipment topology environment, effectively overcoming the technical bottleneck of the lack of depth in heterogeneous data fusion under the impact of sudden events in traditional prediction schemes.

[0031] Step S2: Extract the main frequency characteristics of each state variable from the structured operation data, and reconstruct the main trend of the operation status of the target power supply area based on the main frequency characteristics; In a preferred embodiment, the state variables include electricity load, net load, and reserve requirement; the structured operating data includes historical operating data corresponding to the electricity load, net load, and reserve requirement, respectively. The dominant frequency characteristics of each state variable are extracted from the structured operational data, and the main trend of the operational state of the target power supply area is reconstructed based on the dominant frequency characteristics, including: For each state variable, historical operation data corresponding to the current state variable is extracted from the structured operation data; frequency domain decomposition is performed on the historical operation data to generate multiple frequency components and the amplitude parameters corresponding to each frequency component; The multiple frequency components are sorted in descending order of amplitude parameters, and a preset number of frequency components are extracted from the sorted multiple frequency components as the dominant frequency feature of the current state variable. Historical operating data is input into a pre-established trend regression prediction model to generate the target periodic component parameters of the current state variable; Based on the main frequency characteristics and target periodic component parameters, the time series signal is reconstructed to generate the basic numerical main trend of the current state variables; Extract the extreme values ​​of the basic numerical main trend within the target time period, and determine the extreme value time period corresponding to the extreme value; Extract the data sequence corresponding to the extreme value period from the basic numerical main trend to generate the extreme value period main trend of the current state variable; After generating the basic numerical main trend and extreme value period main trend of all state variables, the basic numerical main trend of each state variable is weighted and subjected to discrete interval mapping based on the preset risk weight and preset safety threshold to generate the risk level main trend of the target power supply area. The main trend of the basic values ​​of each state variable, the main trend of the extreme value period of each state variable, and the main trend of the risk level are combined to generate the main trend of the operating status of the target power supply area.

[0032] Specifically, regarding the execution steps of extracting the dominant frequency characteristics of various state variables from structured operational data and reconstructing the main trend of the target power supply area's operational status based on these characteristics, each state variable is explicitly defined to include electricity load, net load, and reserve demand. The corresponding structured operational data consists of historical sequence data corresponding to each of the electricity load, net load, and reserve demand. By performing cross-analysis of the time and frequency dimensions of the state variables in different dimensions, the complex non-stationary operational sequence can be decoupled into a periodic fluctuation baseline with clear physical meaning.

[0033] For each state variable, historical operational data corresponding to the currently traversed state variable is extracted from the structured operational data. A Discrete Fourier Transform is performed on the extracted historical operational data to convert the time-domain signal to the frequency domain for frequency domain decomposition. After frequency domain decomposition, multiple frequency components and their corresponding amplitude parameters are generated. The specific mathematical expressions for calculating the amplitude parameters are as follows: In the above mathematical expression, The identifier represents the state variable being traversed. The state variable identifier includes the power load identifier, the net load identifier, and the reserve requirement identifier. A yearly index representing the historical operational data; This represents the total number of time series sampling points within a single year. Under the condition of a non-leap year and sampling granularity at the hour level, the total number of time series sampling points is 8760. Represents the sampling time step variable; A historical data sequence of specific state variables corresponding to a time year index; Represents a discrete frequency index; Represents the complex coefficients generated by the Discrete Fourier Transform; This indicates the operation of taking the real part of a complex number; This represents the operation of taking the imaginary part of a complex number; Represents the cosine component coefficients; Represents the coefficients of the sinusoidal component; This represents the frequency domain amplitude parameter corresponding to the discrete frequency index.

[0034] After obtaining the frequency domain amplitude parameters of all frequency components, the frequency components are arranged in descending order of their amplitude parameters. A predetermined number of frequency components at the top of the sorted list are extracted as the dominant frequency feature of the current state variable. This predetermined number is strictly set to a natural number between 5 and 10. The extracted dominant frequency feature comprehensively covers intraday fluctuation frequencies, weekly fluctuation frequencies, annual cycle frequencies, and associated harmonic overtones, effectively eliminating high-frequency random noise and aperiodic abnormal disturbances.

[0035] Historical operating data is input into a pre-established trend regression prediction model to generate the target periodic component parameters of the current state variable. The trend regression prediction model is constructed using a multi-output ridge regression network architecture. The multi-output ridge regression network includes an input mapping layer that receives and standardizes historical periodic features, and a linear mapping output layer with L2 regularization penalty constraints. The mathematical expression of the loss function for parameter optimization training of the trend regression prediction model using historical operating data is as follows: In the above mathematical expression, This represents the optimal network weight matrix corresponding to the current state variable; This represents the parameter-solving operator that minimizes the subsequent cost function; Represents the dynamic weight variables in the iterative optimization process; This represents the historical structured periodic feature matrix after normalization using the pre-statistical training set mean and standard deviation. This represents the target fitting coefficient matrix, which is formed by concatenating the cosine and sine component coefficients generated from historical data. The Frobenius norm of a matrix; This represents a pre-configured regularization penalty constant, specifically designed to control network architecture complexity and prevent overfitting.

[0036] After determining the optimal network weight matrix through the above optimization training, the newly added structured periodic feature vector corresponding to the target time period is input into the multi-output ridge regression network. The linear mapping output layer multiplies the input feature vector with the optimal network weight matrix, directly outputting the target periodic component parameters of the current state variable. The target periodic component parameters include the predicted average baseline term for the target time period, the predicted future cosine coefficient, and the predicted future sine coefficient.

[0037] Based on the dominant frequency characteristics and target periodic component parameters, inverse time-series signal reconstruction is performed to generate the basic numerical trend of the current state variables. The specific mathematical expression for inverse time-series signal reconstruction is as follows: In the above mathematical expression, Represents the target time Reconstruct the main trend of the underlying numerical values ​​in the output; This represents the predicted average baseline term in the target periodic component parameters; The set of core frequency indices defined by the main frequency characteristics; Represents the traversed elements in the core frequency index set; Represents the cosine coefficient for predicting the future; This represents the sine coefficient for predicting the future.

[0038] Extract the extreme values ​​of the basic numerical main trend within the target time period, and determine the extreme value periods corresponding to these extreme values. Specifically, the execution logic involves a global search for the maximum and minimum evolutionary peaks and troughs of the basic numerical main trend within the target time period, and locking in the time scale intervals corresponding to the occurrence of these peaks and troughs. Data sequence fragments corresponding to these time scale intervals are extracted from the basic numerical main trend to generate the extreme value period main trend of the current state variable. The extreme value period main trend is accurately divided into peak load period main trend sequences and trough load period main trend sequences.

[0039] After generating the basic numerical trends and extreme value periods for all state variables, the basic numerical trends of each state variable are weighted and mapped using a discrete interval method based on preset risk weights and safety thresholds to generate the risk level trend for the target power supply area. The specific mathematical expressions for the weighted summation and discrete interval mapping are as follows: In the above mathematical expression, Representing the target moment The generated risk level main trend output value; This indicates that the calculation result is rounded down. This represents a set of continuous state variable attributes, which consists of electrical load, net load, and reserve demand. The severity influence weight parameter in the preset risk weights is matched with a specific continuous state quantity; This represents a non-negative truncation filter operation, retaining values ​​greater than or equal to zero; This represents the critical safety threshold that matches a specific continuous state quantity within the preset safety threshold. This represents the strong constraint penalty weight parameter specifically configured for external scheduling over-limit indicators in the preset risk weights; The conditional indicator function takes a value of 1 when the external scheduling constraint verification result at the target time determines that the power grid safety standard is not met, and a value of 0 otherwise. The output of the floor operation is ultimately restricted by the mapping rule to a continuous integer space from 1 to 5, representing five discrete risk baseline levels from low to high.

[0040] By combining the fundamental numerical trends of various state variables, the extreme value periods of various state variables, and the risk level trends across all dimensions, the main trend of the operating status of the target power supply area is generated. Through this multi-stage decomposition and reconstruction process, not only are the intrinsic periodic evolution patterns hidden behind massive amounts of power grid operation data deeply extracted, but a risk assessment baseline with solid electrical fundamentals and physical constraints is also constructed. This significantly improves the predictive model's ability to withstand disturbances and the global physical consistency of its output results when facing changes in non-stationary power grid operating conditions.

[0041] Step S3: Input unstructured event data into the large language model to perform information extraction and generate structured event knowledge units containing device description text of affected devices; calculate the matching relationship between the device description text and the device identifier of each physical device in the physical device set to obtain the association weight; aggregate the association weights according to the device identifier of each physical device and various state variables to generate the operation impact factor; process the operation impact factor according to the preset residual mapping model and output the event-driven residual; In a preferred embodiment, unstructured event data is input into a large language model to perform information extraction, generating structured event knowledge units containing device description text of the affected devices; the matching relationship between the device description text and the device identifiers of each physical device in the physical device set is calculated to obtain the association weight, including: Based on the preset prompt word template, unstructured event data is input into a large language model for information extraction, generating structured event knowledge units; among them, the structured event knowledge units include event type information and device description text of multiple affected devices; Extract the device type attribute and device identifier of each physical device from the physical device set; Based on the preset type matching rules, the event type information is compared and verified with the device type attributes of each physical device to generate the type matching degree between the event type information and the device type attributes of each physical device. Based on the preset text representation model, feature extraction is performed on the device identifier of each physical device to generate a physical device feature vector corresponding to the device identifier of each physical device. For each device description text, features are extracted from the current device description text based on the text representation model to generate the affected device feature vector corresponding to the current device description text. The spatial distance between the feature vector of the affected device and the feature vector of each physical device is calculated one by one to generate the entity link similarity between the current device description text and the device identifier of each physical device. Numerical aggregation calculations are performed based on entity link similarity and type matching to generate association weights between the current device description text and the device identifiers of each physical device.

[0042] In a preferred embodiment, the structured event knowledge unit further includes event source credibility and information extraction confidence; Based on the device identifiers and various state variables of each physical device, the associated weights are aggregated and calculated to generate operational impact factors; the operational impact factors are processed according to a pre-defined residual mapping model to output event-driven residuals, including: Based on the association weight and event type information, determine the conflict event characteristics corresponding to the device identifiers of each physical device; Based on the credibility of the event source and the confidence of information extraction, the conflict event features corresponding to the device identifiers of each physical device are harmonized and weighted to generate conflict handling coefficients corresponding to the device identifiers of each physical device. Based on the equipment identifiers of each physical device and various status variables, the associated weights are aggregated and calculated according to the conflict handling coefficients to generate operational impact factors corresponding to each status variable. Each operational influencing factor is input into a preset residual mapping model for feature mapping processing to generate initial event-driven residuals; Perform logic violation checks on the initial event-driven residuals and generate penalty bias values; Numerical correction operations are performed based on the initial event-driven residual and the penalty bias value to generate the event-driven residual.

[0043] Specifically, for the execution steps of inputting unstructured event data into a large language model to extract information and generate structured event knowledge units containing device description text of affected devices, the first step is to obtain a preset prompt word template. The preset prompt word template includes task background instructions, extraction target field constraint instructions, and input text placeholders.

[0044] After filling unstructured event data into input text placeholders, the data is fed into a pre-deployed large language model based on a self-attention mechanism. For example, this large language model can be a pre-trained large model based on the Transformer architecture (such as the GPT series, BERT series, or ChatGLM). Before pre-deployment, this large language model has been fine-tuned based on massive amounts of power system professional corpora (such as historical dispatch operation records, equipment fault analysis reports, power grid operation specifications, etc.), enabling it to possess deep semantic understanding and entity feature extraction capabilities for power field-specific terminology. Specifically, this fine-tuning training process is supervised fine-tuning. During the training phase, the input of the model training samples is historically collected unstructured event text records, and the corresponding training labels are manually verified and labeled standardized event knowledge units (i.e., clearly defined event type information and descriptions of the actual affected physical equipment). The model iteratively updates the network weight parameters by minimizing the cross-entropy loss function, thereby achieving a precise mapping from unstructured text to structured knowledge. For example, the training sample data pairs used for supervised fine-tuning mentioned above include the following specific form: An input unstructured text sample is "Due to the influence of strong cold air, extreme low temperatures are expected in the city tomorrow. Please take precautions. Additionally, Unit 2 of the West District Thermal Power Plant has been shut down due to unplanned maintenance." The corresponding manually labeled structured training labels are formatted JSON data: "{Event 1:{Event Type: Low Temperature Warning, Affected Equipment:[]},Event 2:{Event Type: Unit Failure, Affected Equipment:[West District Thermal Power Plant Unit 2]}}". Through the input of massive amounts of similar sample pairs, the model accurately learns professional power semantic information.

[0045] The large language model outputs serialized structured data based on task context instructions, generating structured event knowledge units. Each structured event knowledge unit explicitly includes event type information, equipment description text for multiple affected devices, event source credibility, and information extraction confidence. Event type information is strictly divided into four categories: high-temperature warnings, low-temperature warnings, unit failures, and maintenance operations. The event source credibility is assigned a floating-point value between 0.1 and 1.0 based on a preset matching table of the authority level of the issuing organization. For example, this matching table can be specifically defined as follows: when the event source is a national meteorological bureau or a provincial power grid dispatch center, its authority level coefficient is assigned 1.0; when the source is an internal notice issued by a municipal power supply bureau, the coefficient is assigned 0.8; and when the source is a report from public online news media, the coefficient is assigned 0.5. This achieves a quantitative assessment of the reliability of different information sources. The information extraction confidence is directly calculated from the mean of the normalized probability distribution generated by the large language model when generating the corresponding equipment description text.

[0046] The system centrally extracts the device type attributes and device identifiers of each physical device from the physical device set. Based on preset type matching rules, it compares and verifies the event type information with the device type attributes of each physical device. The preset type matching rules are represented by a two-dimensional matrix logical mapping table. When an event type information and a specific physical device's device type attribute have a corresponding mapping relationship in the two-dimensional matrix logical mapping table, the matching state is considered valid, and a type matching degree of 1 is generated; when no mapping relationship exists in the two-dimensional matrix logical mapping table, the matching state is considered invalid, and a type matching degree of 0 is generated. For example, the row dimension of this two-dimensional matrix logical mapping table is the event type set, and the column dimension is the device type attribute set. In actual configuration, for example, the event type "unit failure" is only configured with a value of 1 at the intersection point with the device type "generator set" in the matrix (there is a mapping relationship), while the value at the intersection point of "unit failure" with "transformer" or "transmission line" is configured with 0; similarly, the event type "high temperature warning" is configured to have a mapping relationship with the device type representing the load node. This filters out invalid illusions that may arise from large language models at the level of common sense in power grid physics.

[0047] A pre-defined text representation model is synchronously introduced to encode and transform the text to be tested. This model employs a deep bidirectional encoding network architecture to extract features from the device identifiers of each physical device, generating physical device feature vectors corresponding to those identifiers. As a feasible implementation, this text representation model can specifically use BERT (Bidirectional Encoder Representations from Transformers) or its variants, leveraging its deep bidirectional attention mechanism to map discrete text characters into high-dimensional dense vectors containing rich contextual semantics. Furthermore, to adapt to the device matching task of this invention, the text representation model undergoes supervised contrastive learning training based on the aforementioned pre-training. Specifically, a training dataset is constructed containing positive sample pairs (i.e., non-standard event descriptions and standard device identifiers pointing to the same physical device) and negative sample pairs (i.e., descriptions and identifiers pointing to different physical devices); the binary classification result (1 or 0) of whether the sample pairs match is used as the training label; during training, positive and negative sample pairs are input into the model respectively, and the model parameters are optimized by minimizing the contrastive loss function (such as InfoNCE loss), so that the distance between text feature vectors pointing to the same physical device in the vector space is continuously reduced, while the distance between vectors pointing to different devices is increased.

[0048] For each extracted device description text, deep feature extraction is performed on the current device description text based on the text representation model to generate the affected device feature vector corresponding to the current device description text. The spatial distance between the affected device feature vector and the feature vector of each physical device is calculated one by one. The specific mathematical expression for spatial distance calculation is as follows: In the above mathematical expression, An event index representing the traversed unstructured event data; The device index representing the physical devices being traversed; This represents the similarity of entity links between the current device description text and the device identifier of a specific physical device. This represents the device description text corresponding to the current event index; This represents the device identifier corresponding to the current device index; This represents the mapping function of the text representation model for describing the text; This represents the mapping function of the text representation model for device identifiers; This represents vector norm operations.

[0049] In the above mathematical expression, An event index representing the traversed unstructured event data; This index strictly represents the physical device entity node index (e.g., a specific transformer or generator set) in the power grid topology model. This index establishes a physical mapping mechanism for converting unstructured text into power grid topology node attributes. After obtaining the entity link similarity, numerical aggregation calculations are performed based on the entity link similarity and type matching degree to generate the association weights between the current device description text and the device identifiers of each physical device. The specific mathematical expression for the association weights is as follows: In the above mathematical expression, This represents the association weight between the current device description text and the device identifiers of each physical device; This represents the type matching degree between event type information and the device type attributes of each physical device.

[0050] Based on the association weights and event type information, conflict event features corresponding to the device identifiers of each physical device are determined. A conflict event feature refers to a situation where, within the same target prediction time period, the device identifier of the same physical device is simultaneously associated with multiple event types pointing to the same state variable but with diametrically opposed directions. Based on the event source confidence and information extraction confidence, the conflict event features corresponding to the device identifiers of each physical device are harmonicly weighted to generate conflict resolution coefficients corresponding to each physical device identifier. The specific mathematical expression for the conflict resolution coefficients is as follows: In the above mathematical expression, Represents the generated conflict resolution coefficient; This represents the credibility of the event source corresponding to the current event index; This represents the confidence level of information extraction corresponding to the current event index; A set of indexes representing conflicting events with opposite directions of impact currently faced by the physical device; The elements to be traversed in the set of conflict event indices; This represents the credibility of other related events that are in conflict. This represents the confidence level of extracting other related information that is in a conflict state. If a specific physical device does not have corresponding conflict event characteristics, the current conflict handling coefficient is uniformly set to a constant of 1.

[0051] Based on the device identifiers of each physical device and various state variables, the association weights are aggregated using the conflict resolution coefficients corresponding to each device identifier, generating operational impact factors corresponding to each state variable. The specific mathematical expression for the aggregation calculation is as follows: In the above mathematical expression, This represents the target prediction time period; The dimension index representing the state variable; These represent the operational influencing factors of specific state variables within the target prediction time period; This represents the set of all event knowledge units within the target prediction time period; Represents a collection of physical devices; The influence direction coefficient of the event type information on a specific state variable is represented by a value of 1 or -1. This represents the severity value contained in unstructured event data, which is normalized to the range of 0 to 1.

[0052] The operational influencing factors corresponding to each state variable are input into a pre-defined residual mapping model for feature mapping processing to generate initial event-driven residuals. The pre-defined residual mapping model employs a multilayer perceptron network architecture, which includes one input layer, two hidden layers, and one output layer. The first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons, all using a linear rectified function as the activation function. During the training phase, the pre-defined residual mapping model uses the operational influencing factors of each state variable corresponding to historical events as network input features. The target residual label is obtained by subtracting the historical true values ​​of the operational states from the historical main trend prediction values. The mean squared error loss function between the forward output residual value and the target residual label is minimized using the backpropagation algorithm, thus solidifying the weight parameters of the multilayer perceptron network architecture. After forward feature propagation, the output layer directly outputs the initial event-driven residuals that are perfectly aligned with the dimensions of each state variable.

[0053] Logical violation checks are performed on the initial event-driven residuals to generate a penalty bias value. The logical violation check aims to verify whether the network architecture output violates fundamental physical principles of power systems. The specific mathematical expression for the penalty bias value is as follows: In the above mathematical expression, This represents the total penalty bias value; Residual components representing the backup demand dimension in the initial event-driven residual; This represents the unit fault indication function, which takes a value of 1 when the event type information indicates a unit fault, and a value of 0 otherwise. Represents the residual component in the power load dimension of the initial event-driven residual; This represents a low-temperature event indicator function, which takes a value of 1 when the event type information indicates a low-temperature warning and it is in a winter month; otherwise, it takes a value of 0. This represents the non-negative truncation operation function. The logical violation verification mechanism forces the neural network model to follow predetermined electrical laws. If a unit failure event occurs, the reserve demand residual component must not be significantly negative; if an extreme low-temperature event occurs, the electricity load residual component must not be significantly negative. Numerical correction calculations are performed based on the initial event-driven residual and the penalty bias value. The penalty bias value is multiplied by a preset descent step size constant and then subtracted from the initial event-driven residual to finally generate the event-driven residual.

[0054] By performing deep feature extraction of unstructured text and feature mapping calculation steps that follow the red lines of physical rules, a quantization conversion channel from semantic space to topological numerical space is opened up. Furthermore, relying on the rigorous bottom line of electrical logic, unreasonable prediction biases of pure data-driven models are eliminated, thereby improving the physical interpretability and output reliability of the final prediction residual results.

[0055] Step S4: Perform a multi-dimensional vector superposition operation on the main trend of the operating status and the event-driven residual to generate the target operating status prediction result of the target power supply area. In a preferred embodiment, the main trend of the operating state and the event-driven residual are superimposed using a multi-dimensional vector operation to generate a target operating state prediction result for the target power supply area, including: Vector bias fusion calculation is performed on the main trend of the running state and the event-driven residuals to generate the initial running state prediction results; According to the preset operation state mapping rules, numerical boundary constraints and discrete dimension mapping processing are performed on the initial operation state prediction results to generate the target operation state prediction results for the target power supply area; wherein, the target operation state prediction results include the predicted values ​​corresponding to each state variable.

[0056] Specifically, after obtaining the main trend of the operating state and the event-driven residual, the main trend of the operating state and the event-driven residual are superimposed by multi-dimensional vectors to generate the target operating state prediction result of the target power supply area, which is clearly divided into two execution steps: vector bias fusion and state mapping processing.

[0057] Vector bias fusion is performed on the main trend of the running state and the event-driven residuals to generate the initial running state prediction result. The specific mathematical expression of the vector bias fusion operation is as follows: In the above mathematical expression, Representing the target moment The corresponding initial running state prediction results; Representing the target moment The corresponding main trend of the operating state, which includes the state baseline vector without considering non-periodic events; Representing the target moment The corresponding event-driven residual; The index represents the target time. By performing multi-dimensional vector addition and accumulation operations, the initial running state prediction result, while retaining the historical fundamental cyclical fluctuation characteristics, directly superimposes the quantized perturbation bias value brought about by external unstructured events.

[0058] Based on the preset operating state mapping rules, numerical boundary constraints and discrete dimension mapping processing are applied to the initial operating state prediction results to generate the target operating state prediction results for the target power supply area. The target operating state prediction results include the predicted values ​​corresponding to each state variable. The specific mapping processing mathematical expressions are as follows: In the above mathematical expression, Representing the target moment The corresponding final output is the target running state prediction result; This represents the mapping processing function corresponding to the preset running state mapping rule; This represents the set of mapping parameters included in the preset running state mapping rules.

[0059] To ensure that the predicted output conforms to the operational standards of the actual power grid operation, the preset operation state mapping rules include differentiated constraint mechanisms for different data dimensions. Each state variable explicitly includes electricity load, net load, and reserve demand. For continuous state variables in the dimensions of electricity load, net load, and reserve demand, the numerical boundary constraints in the preset operation state mapping rules forcibly truncate and reset any unreasonable negative values ​​in the initial operation state prediction results to a constant of 0, ensuring that the output power and capacity prediction values ​​strictly conform to the non-negative physical constraints of the actual power grid operating environment.

[0060] For discrete state variables in the extreme value period dimension, including peak load periods and valley load periods, a discrete dimension mapping processing mechanism is executed. Specifically, according to the standard scheduling scale of 24-hour level operating intervals or 96 15-minute level operating intervals per day, the consecutive extreme value period prediction time nodes in the initial operating state prediction results are accurately projected into the corresponding valid discrete integer numbers in the intervals of 1 to 24 or 1 to 96 based on the nearest rounding algorithm.

[0061] By performing multidimensional vector superposition operations and differential mapping constraint processing, the generated target operating state prediction results not only integrate the main trend and event disturbances, but also strictly follow the physical boundaries and discrete-time scheduling specifications of the power grid's underlying structure, providing quantitative data support with absolute physical feasibility for subsequent multidimensional risk scoring and scheduling strategy matching.

[0062] Step S5: Based on the predicted values ​​of each state variable in the target operating state prediction results, match the preset scheduling risk rules to generate a scheduling control strategy to execute scheduling control on each physical device.

[0063] In a preferred embodiment, based on the predicted values ​​of various state variables in the target operating state prediction results, a preset scheduling risk rule is matched to generate a scheduling control strategy to perform scheduling control on each physical device, including: According to the preset scoring mapping rules, the predicted values ​​of each state variable in the target operation state prediction results are normalized and mapped respectively to generate a single-dimensional risk score corresponding to each state variable. A weighted aggregation calculation is performed on the single-dimensional risk scores corresponding to each state variable to generate a comprehensive risk assessment value. Based on the preset scheduling risk rules, the comprehensive risk assessment values ​​are compared across intervals to determine the scheduling security risk level corresponding to the comprehensive risk assessment values. The preset scheduling risk rules include risk classification thresholds and candidate scheduling control strategies bound to each scheduling security risk level. Based on the scheduling safety risk level and the predicted values ​​of various state variables, candidate scheduling control strategies are extracted from the scheduling risk rules and used as scheduling control strategies to perform scheduling control on each physical device.

[0064] Specifically, after obtaining the target operating state prediction results, a series of actions are performed on the predicted data contained in the target operating state prediction results, including dimensional unification, risk assessment, and underlying equipment control and scheduling. First, according to preset scoring mapping rules, the predicted values ​​of each state variable covered in the target operating state prediction results are subjected to feature transformation, i.e., normalization mapping processing is performed separately. The preset scoring mapping rules include a linear interpolation normalization function for continuous state variables and a step mapping function for discrete state variables. The linear interpolation normalization function projects the predicted values ​​of electricity load, net load, and reserve demand to a standard percentage score range of 0 to 100. The step mapping function maps the predicted values ​​of peak load periods and valley load periods to standard scores representing the severity of time offset. After normalization mapping processing, a one-dimensional risk score corresponding to each state variable is generated.

[0065] After obtaining the individual risk scores, a weighted aggregation calculation is performed on all individual risk scores to generate a comprehensive risk assessment value. The specific mathematical expression for the weighted aggregation calculation is as follows: In the formula, This represents a comprehensive risk assessment value; The influence weighting parameter represents the corresponding electricity load forecast data; The influence weighting parameter represents the net load forecast data; The influence weighting parameter represents the data corresponding to the reserve demand forecast. The influence weight parameter represents the predicted data for extreme value periods; The influence weight parameter represents the result of external scheduling constraint verification. This represents a one-dimensional risk score generated by performing a normalization operation on the predicted electricity load values. This represents a one-dimensional risk score generated by performing a normalization operation on the net load forecast values. This represents a one-dimensional risk score generated by performing a normalization operation on the predicted reserve demand values. This represents a one-dimensional risk score generated by performing a normalization operation on the predicted values ​​during extreme value periods. This represents a one-dimensional risk score generated by performing a normalization operation on the verification results of external scheduling constraints.

[0066] The values ​​of each influencing weight parameter are preset by the prior historical data of the scheduling business after calibration with the validation set, and the sum of all influencing weight parameters is strictly equal to the constant 1.

[0067] After generating the comprehensive risk assessment value, an interval comparison process is performed on the value according to the preset scheduling risk rules to determine the corresponding scheduling security risk level. The preset scheduling risk rules contain multiple discrete risk grading thresholds, which divide the effective space of the comprehensive risk assessment value into multiple independent judgment intervals. The mathematical expression for the interval comparison is as follows: when Upon establishment, the dispatching security risk level was determined to be: The numerical value corresponding to the variable.

[0068] In the formula, This represents the lower boundary value of a specific judgment interval within the risk classification threshold. This represents the upper boundary value of a specific judgment interval within the risk classification threshold. This represents the level of safety risk in scheduling.

[0069] Dispatch safety risk levels are strictly limited to consecutive natural numbers from 1 to 5. In the implementation logic of the preset dispatch risk rules, constants 20, 40, 60, and 80 are explicitly set as core risk classification thresholds. When the comprehensive risk assessment value is below constant 20, the dispatch safety risk level obtained from the interval comparison is level 1. When the comprehensive risk assessment value is greater than or equal to constant 80, the dispatch safety risk level obtained from the interval comparison is level 5. The higher the level value, the more serious the operational safety hazards faced by the target power supply area.

[0070] The pre-defined scheduling risk rules not only include risk classification thresholds but also candidate scheduling control strategies deeply bound to each scheduling safety risk level. After determining the scheduling safety risk level, the candidate scheduling control strategies bound to the pre-defined scheduling risk rules are precisely extracted as the target scheduling control strategies, based on the scheduling safety risk level and the predicted values ​​of various state variables. The extraction and control execution actions are divided into multiple dimensions. When the predicted value of reserve demand triggers a high-risk level judgment, the extracted scheduling control strategies specifically include increasing spinning reserve capacity, pre-scheduling the start-up of adjustable generator units, and calling demand response resources. When the predicted value of net load shows a rapid increase in a high-risk level judgment, the extracted scheduling control strategies specifically include adjusting the generator unit output plan and optimizing the charging and discharging time of the energy storage station. When the difference between the predicted value of the extreme period and the baseline period is too large, causing the extreme period to deviate from the high-risk judgment, the extracted scheduling control strategies specifically include adjusting the start-up and shutdown time of the units and adjusting the adjustable load plan. After obtaining the scheduling control strategy, entity control commands containing the above control actions are issued to execute command-level scheduling control actions on various physical devices within the power supply area. Each physical device clearly encompasses generator sets, electrochemical energy storage power station equipment, and transformer node equipment within the actual physical space.

[0071] By performing the aforementioned multidimensional state variable feature normalization, weighted aggregation evaluation, and control strategy mapping actions based on discrete risk levels, the execution link for transforming upstream virtual prediction values ​​into front-end physical equipment control actions has been completely opened up. This not only enables the early assessment of potential power grid safety hazards based on multi-source heterogeneous fusion prediction results, but also provides highly forward-looking and quantitatively accurate scheduling and control basis for underlying physical entities such as generator sets and energy storage power stations.

[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 system operation status prediction device, including: a data acquisition module 101, a main trend reconstruction module 102, an event-driven processing module 103, a prediction result generation module 104, and a scheduling control module 105; The data acquisition module 101 is used to acquire the set of physical devices in the target power supply area, the structured operation data of the target power supply area in the target time period, and the unstructured event data. The main trend reconstruction module 102 is used to extract the main frequency characteristics of each state variable based on the structured operation data, and reconstruct the main trend of the operation status of the target power supply area based on the main frequency characteristics. The event-driven processing module 103 is used to input unstructured event data into a large language model to perform information extraction, generate structured event knowledge units containing device description text of affected devices; calculate the matching relationship between the device description text and the device identifiers of each physical device in the physical device set to obtain the association weights; aggregate and calculate the association weights according to the device identifiers of each physical device and various state variables to generate the operation impact factor; process the operation impact factor according to the preset residual mapping model and output the event-driven residual. The prediction result generation module 104 is used to perform multi-dimensional vector superposition operation on the main trend of the operating state and the event-driven residual to generate the target operating state prediction result of the target power supply area. The scheduling control module 105 is used to generate a scheduling control strategy by matching the predicted values ​​of various state variables in the target operating state prediction results with preset scheduling risk rules, so as to perform scheduling control on each physical device.

[0074] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the power system operation state prediction method described in any one of the present invention. Furthermore, the embodiments of the device 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. In addition, in the accompanying drawings of the device 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 system operation state prediction method according to any one of the present invention, or the processor executes the computer program to implement 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 controls the execution of any of the above-described power system operation state prediction 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 predicting the operating state of a power system, characterized in that, include: Acquire the set of physical devices in the target power supply area, the structured operational data of the target power supply area within the target time period, and the unstructured event data; The main frequency characteristics of each state variable are extracted from the structured operation data, and the main trend of the operation status of the target power supply area is reconstructed based on the main frequency characteristics. Unstructured event data is input into a large language model to perform information extraction, generating structured event knowledge units containing device description text of the affected devices; The matching relationship between the device description text and the device identifier of each physical device in the physical device set is calculated to obtain the association weight; based on the device identifier of each physical device and various state variables, the association weight is aggregated and calculated to generate the operation impact factor; The system processes operational influencing factors based on a pre-defined residual mapping model and outputs event-driven residuals. The main trend of the operating status and the event-driven residual are superimposed into a multi-dimensional vector to generate the target operating status prediction result of the target power supply area. Based on the predicted values ​​of various state variables in the target operating state prediction results, a scheduling control strategy is generated by matching the preset scheduling risk rules to execute scheduling control on each physical device.

2. The power system operation status prediction method as described in claim 1, characterized in that, Acquire the physical device set of the target power supply area, the structured operational data of the target power supply area within the target time period, and the unstructured event data, including: Obtain the scheduling object ledger and equipment ledger of the target power supply area, perform object standardization processing on the scheduling object ledger and equipment ledger, and generate a physical equipment set; Acquire historical load data and meteorological characteristic data of the target power supply area within the target time period, unify the time interval and aggregate the granularity of the historical load data and meteorological characteristic data, and generate structured operation data; Receive alarm messages from the preset device fault alarm bus and generate device fault text; Extract warning information from preset weather warning information sources and generate weather warning text; Extract the main text of the pre-set scheduling operation announcement webpage to generate the scheduling announcement text; Time alignment is performed on equipment failure texts, weather warning texts, and dispatch announcement texts to generate unstructured event data.

3. The power system operation status prediction method as described in claim 2, characterized in that, The various state variables include electricity load, net load, and reserve requirement; the structured operation data includes historical operation data corresponding to the electricity load, net load, and reserve requirement, respectively. The dominant frequency characteristics of each state variable are extracted from the structured operational data, and the main trend of the operational state of the target power supply area is reconstructed based on the dominant frequency characteristics, including: For each state variable, historical operating data corresponding to the current state variable is extracted from the structured operating data; frequency domain decomposition is performed on the historical operating data to generate multiple frequency components and the amplitude parameters corresponding to each frequency component; The multiple frequency components are sorted in descending order of amplitude parameters, and a preset number of frequency components are extracted from the sorted multiple frequency components as the dominant frequency feature of the current state variable. Historical operating data is input into a pre-established trend regression prediction model to generate the target periodic component parameters of the current state variable; Based on the main frequency characteristics and target periodic component parameters, the time series signal is reconstructed to generate the basic numerical main trend of the current state variables; Extract the extreme values ​​of the basic numerical main trend within the target time period, and determine the extreme value time period corresponding to the extreme value; Extract the data sequence corresponding to the extreme value period from the basic numerical main trend to generate the extreme value period main trend of the current state variable; After generating the basic numerical main trend and extreme value period main trend of all state variables, the basic numerical main trend of each state variable is weighted and subjected to discrete interval mapping based on the preset risk weight and preset safety threshold to generate the risk level main trend of the target power supply area. The main trend of the basic values ​​of each state variable, the main trend of the extreme value period of each state variable, and the main trend of the risk level are combined to generate the main trend of the operating status of the target power supply area.

4. The power system operation status prediction method as described in claim 3, characterized in that, Unstructured event data is input into a large language model to perform information extraction, generating structured event knowledge units containing device description text of the affected devices; The matching relationship between the device description text and the device identifiers of each physical device in the physical device set is used to obtain the association weight, including: Based on the preset prompt word template, unstructured event data is input into a large language model for information extraction, generating structured event knowledge units; among them, the structured event knowledge units include event type information and device description text of multiple affected devices; Extract the device type attribute and device identifier of each physical device from the physical device set; Based on the preset type matching rules, the event type information is compared and verified with the device type attributes of each physical device to generate the type matching degree between the event type information and the device type attributes of each physical device. Based on the preset text representation model, feature extraction is performed on the device identifier of each physical device to generate a physical device feature vector corresponding to the device identifier of each physical device. For each device description text, features are extracted from the current device description text based on the text representation model to generate the affected device feature vector corresponding to the current device description text. The spatial distance between the feature vector of the affected device and the feature vector of each physical device is calculated one by one to generate the entity link similarity between the current device description text and the device identifier of each physical device. Numerical aggregation calculations are performed based on entity link similarity and type matching to generate association weights between the current device description text and the device identifiers of each physical device.

5. The power system operating state prediction method as described in claim 4, characterized in that, The structured event knowledge unit also includes the credibility of the event source and the confidence level of information extraction; Based on the device identifiers of each physical device and various status variables, the associated weights are aggregated and calculated to generate operational impact factors; The system processes operational influencing factors based on a pre-defined residual mapping model, outputting event-driven residuals, including: Based on the association weight and event type information, determine the conflict event characteristics corresponding to the device identifiers of each physical device; Based on the credibility of the event source and the confidence of information extraction, the conflict event features corresponding to the device identifiers of each physical device are harmonized and weighted to generate conflict handling coefficients corresponding to the device identifiers of each physical device. Based on the equipment identifiers of each physical device and various status variables, the associated weights are aggregated and calculated according to the conflict handling coefficients to generate operational impact factors corresponding to each status variable. Each operational influencing factor is input into a preset residual mapping model for feature mapping processing to generate initial event-driven residuals; Perform logic violation checks on the initial event-driven residuals and generate penalty bias values; Numerical correction operations are performed based on the initial event-driven residual and the penalty bias value to generate the event-driven residual.

6. The power system operation status prediction method as described in claim 5, characterized in that, The main trend of the operating status is superimposed with the event-driven residuals using a multi-dimensional vector operation to generate the target operating status prediction result for the target power supply area, including: Vector bias fusion calculation is performed on the main trend of the running state and the event-driven residuals to generate the initial running state prediction results; According to the preset operation state mapping rules, numerical boundary constraints and discrete dimension mapping processing are performed on the initial operation state prediction results to generate the target operation state prediction results for the target power supply area; wherein, the target operation state prediction results include the predicted values ​​corresponding to each state variable.

7. The power system operating state prediction method as described in claim 6, characterized in that, Based on the predicted values ​​of various state variables in the target operating state prediction results, a pre-set scheduling risk rule is matched to generate a scheduling control strategy to execute scheduling control on each physical device, including: According to the preset scoring mapping rules, the predicted values ​​of each state variable in the target operation state prediction results are normalized and mapped respectively to generate a single-dimensional risk score corresponding to each state variable. A weighted aggregation calculation is performed on the single-dimensional risk scores corresponding to each state variable to generate a comprehensive risk assessment value. Based on the preset scheduling risk rules, the comprehensive risk assessment values ​​are compared across intervals to determine the scheduling security risk level corresponding to the comprehensive risk assessment values. The preset scheduling risk rules include risk classification thresholds and candidate scheduling control strategies bound to each scheduling security risk level. Based on the scheduling safety risk level and the predicted values ​​of various state variables, candidate scheduling control strategies are extracted from the scheduling risk rules and used as scheduling control strategies to perform scheduling control on each physical device.

8. A power system operating status prediction device, characterized in that, include: The module consists of a data acquisition module, a main trend reconstruction module, an event-driven processing module, a prediction result generation module, and a scheduling and control module. The data acquisition module is used to acquire the set of physical devices in the target power supply area, the structured operation data of the target power supply area within the target time period, and the unstructured event data. The main trend reconstruction module is used to extract the main frequency characteristics of each state variable based on the structured operation data, and reconstruct the main trend of the operation status of the target power supply area based on the main frequency characteristics. The event-driven processing module is used to input unstructured event data into a large language model to perform information extraction and generate structured event knowledge units containing device description text of the affected devices. The matching relationship between the device description text and the device identifier of each physical device in the physical device set is calculated to obtain the association weight; based on the device identifier of each physical device and various state variables, the association weight is aggregated and calculated to generate the operation impact factor; The system processes operational influencing factors based on a pre-defined residual mapping model and outputs event-driven residuals. The prediction result generation module is used to perform multi-dimensional vector superposition operation on the main trend of the operating state and the event-driven residual to generate the target operating state prediction result of the target power supply area. The scheduling control module is used to match the predicted values ​​of various state variables in the target operating state prediction results with preset scheduling risk rules to generate a scheduling control strategy for executing scheduling control on each physical device.

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 system operating state prediction 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 executed, it controls the device where the storage medium is located to perform the power system operating state prediction method as described in any one of claims 1 to 7.

Images