Power grid section limit power dynamic adjustment method based on boundary condition correction model

Abstract

The invention relates to the technical field of safe operation and dispatching of a power system, and discloses a boundary condition correction model-based power grid section limit power dynamic adjustment method, which comprises the following steps of: acquiring and processing multi-source operation data to form a standardized operation boundary condition time sequence; acquiring the limit transmissible power of the target power transmission section at a historical moment or in an offline simulation scene, and obtaining a true limit correction value; training the supervised learning time series prediction model to obtain a boundary condition correction model; training an unsupervised learning model based on historical normal operation data to establish a reference model of a normal operation domain; and multiplying a quota correction predicted value output by the boundary condition correction model by the dynamic safety coefficient of the corresponding moment to obtain a final correction after safety adjustment, and adding the final correction and the static reference quota power to obtain the dynamic section quota power. According to the method, unsupervised learning is introduced to construct a normal operation domain, and automatic identification of extreme and unseen working conditions is realized.

Description

Technical Field

[0001] This invention relates to the field of power system safe operation and dispatch technology, specifically to a method for dynamic adjustment of grid cross-sectional limit power based on a boundary condition correction model. Background Technology

[0002] The stability limits for power transmission sections are key constraints for ensuring the safe and stable operation of the power system. Current methods primarily rely on the "Guidelines for Power System Safety and Stability" and related stability limit regulations. Under typical operating conditions and preset fault sets, static or tiered power limits for transmission sections are established through offline power flow calculations and stability analysis. These limits are applied to the dispatch and monitoring system to monitor power flow in real time and determine whether limits have been exceeded.

[0003] However, with the widespread integration of large-scale intermittent renewable energy sources and the increasing complexity of power grid operation, the actual transmission capacity of transmission sections is significantly affected by multi-dimensional operational boundary conditions such as load distribution, renewable energy output, equipment maintenance plans, and meteorological conditions. Traditional static quota systems and rule-based electronic quota management methods are unable to reflect the continuous changes of these boundary conditions in real time and accurately. This often leads to quota adjustments relying on manual experience or semi-automatic methods in actual operation, resulting in not only delayed response and heavy workload but also two unfavorable situations: first, under most conventional operating conditions, the quotas are too conservative, failing to fully utilize the actual transmission capacity of the power grid and reducing resource utilization efficiency and renewable energy absorption levels; second, under extreme or complex boundary conditions, the judgment of safety margins is not precise enough, potentially implying operational risks.

[0004] Therefore, there is an urgent need to study a new method that can automatically sense operating boundary conditions and realize intelligent, dynamic, and safe adjustment of cross-sectional limit power, so as to fully tap the transmission potential while ensuring grid security. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention aims to provide a method, device, and storage medium for dynamic adjustment of power grid cross-sectional limit power based on a boundary condition correction model. While fully compatible with the existing static limit system, this invention dynamically calculates the limit correction amount through a data-driven model and introduces an unsupervised safety adjustment mechanism. This aims to achieve forward-looking, adaptive, and risk-controllable dynamic fine-tuning of cross-sectional limits, thereby improving the utilization efficiency and safe operation level of the power grid transmission capacity.

[0006] This invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides a method for dynamic adjustment of the power limit limit of a power grid section based on a boundary condition correction model, comprising:

[0008] S1. For the target transmission section, collect and process multi-source operation data that affect the transmission capacity of the target transmission section to form a standardized operation boundary condition time series, wherein each time t in the operation boundary condition time series corresponds to a feature vector b(t) that characterizes the operation state at that time.

[0009] S2. Obtain the maximum transmittable power of the target transmission section in historical time or offline simulation scenario, and calculate the maximum transmittable power and the preset static reference limit power. The difference yields the true value of the limit correction amount for the corresponding time or scenario. The feature vector b(t) and the limit correction amount Construct the training sample set;

[0010] S3. Using the time series of the operation boundary conditions in the historical period as input, and the true value series of the limit correction amount at the same time and in the future period as the output target, train the supervised learning time series prediction model (boundary condition correction model) to obtain the boundary condition correction model; when applied online, input the real-time and historical operation boundary condition time series into the boundary condition correction model to predict the limit correction amount prediction series for the future period.

[0011] S4. Train an unsupervised learning model based on historical normal operation data to establish a reference model for the normal operation domain; input the current or predicted time series of operation boundary conditions into the reference model, calculate the degree to which the time series of operation boundary conditions deviates from the normal operation domain, and obtain an anomaly index. According to the aforementioned anomaly index Calculate a dynamic safety factor The safety factor The value varies with the degree of abnormality The value increases and then monotonically decreases;

[0012] S5. Multiply the predicted limit correction value output by the boundary condition correction model by the dynamic safety factor at the corresponding time. The final correction amount after safety adjustment is obtained, and the final correction amount is compared with the static reference limit power. The sums are used to obtain the dynamic cross-sectional power limit for power grid safety monitoring and dispatching.

[0013] Furthermore, the specific process of constructing the time series of the operational boundary conditions is as follows:

[0014] Normalize or standardize continuous operational data;

[0015] For periodic time data, periodic encoding is performed by converting between sine and cosine functions;

[0016] For discrete state data, binarization, one-hot coding, or embedded vector coding are used for processing.

[0017] The processed features from each dimension are concatenated into a unified feature vector b(t).

[0018] Furthermore, the specific process for obtaining the maximum transmittable power of the target transmission section at a historical moment or in an offline simulation scenario is as follows:

[0019] From the historical operation records of the power grid dispatch automation system and the safety verification system, extract the transmittable capacity assessment value of the target section obtained by safety constraint unit combination, safety constraint economic dispatch or total transmission capacity assessment at the corresponding historical moment and under the predetermined N-1 safety and stability constraints, and use it as the limit transmittable power at that historical moment.

[0020] And / or,

[0021] On the offline simulation platform, power flow calculation and safety and stability analysis are performed for the set operating boundary conditions. Under the premise of meeting the power system safety and stability guidelines and equipment thermal stability constraints, the maximum power that the target section can safely transmit is solved as the limit transmittable power under the simulation scenario.

[0022] Furthermore, the boundary condition correction model employs a deep neural network model with an encoding / decoding structure. The input to this deep neural network model is a sequence of operational boundary condition features within a historical time window, and the output is a predicted sequence of limit correction amounts for a future time window. The deep neural network model uses the static baseline limit power... Residual modeling is performed for reference.

[0023] Furthermore, the boundary condition correction model is configured to simultaneously output multiple probability quantile values ​​of the limit correction amount for future periods, wherein the probability quantiles include a median value and at least one conservative estimate greater than the median; and the training loss function of the boundary condition correction model combines the mean squared error loss for the median value and the quantile regression loss for each quantile.

[0024] Furthermore, when training the boundary condition correction model, a weight is assigned to each training sample. The weight is negatively correlated with the safety margin of the operating condition corresponding to the training sample, so that the training sample with a smaller safety margin contributes more to the training loss function during the training process.

[0025] Furthermore, the unsupervised learning model is a reference model constructed using an autoencoder or a variational autoencoder; the anomaly index The error norm is measured by calculating the output vector reconstructed from the reference model of the current running boundary condition feature vector and comparing it with the original input vector.

[0026] Furthermore, the safety factor The calculation satisfies: ,in, The minimum safety factor is preset; when the anomaly index When the preset threshold is exceeded, a fallback mechanism is activated to reduce the safety factor. Set as Alternatively, the final correction amount may be truncated to a preset conservative upper limit.

[0027] In a second aspect, the present invention discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned method for dynamic adjustment of grid cross-sectional limit power based on a boundary condition correction model.

[0028] Thirdly, the present invention discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned method for dynamic adjustment of grid cross-sectional limit power based on a boundary condition correction model.

[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0030] This invention adopts a residual modeling framework of static benchmark and dynamic correction, which fully inherits the existing static quota system. The model output is an easily understandable and acceptable correction amount, which is convenient to connect with the existing scheduling procedures and has strong engineering feasibility.

[0031] This invention uses a deep time-series model to accurately learn the complex mapping relationship between operating boundary conditions and the actual transmission capacity of a section. Under normal and favorable operating conditions, it can dynamically relax the limit, fully explore the potential transmission capacity of the section, and improve the economic efficiency of the power grid and the level of new energy consumption.

[0032] This invention innovatively introduces unsupervised learning to construct the normal operating domain, enabling automatic identification of extreme and unseen operating conditions. It also dynamically and quantitatively tightens the safety factor through anomaly indices, achieving adaptive management of the safety margin. Combined with the multi-quantile risk output of the supervised model, it forms a dual safety barrier of prediction and fallback.

[0033] The sample-weighted training mechanism of this invention makes the model pay more attention to high-risk extreme conditions; the safety fallback mechanism ensures that the system can automatically retreat to a conservative state in extreme cases outside the model's cognitive boundaries. The entire system has good environmental adaptability and operational robustness.

[0034] This invention automates the entire process from operational data perception and intelligent model prediction to automatic quota adjustment. It can also make forward-looking predictions for multiple future scheduling periods, providing continuous and dynamic quota trajectories for scheduling decisions and improving the level of scheduling intelligence. Attached Figure Description

[0035] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0036] Figure 1 This is a flowchart of a method for dynamically adjusting the power limit of a power grid section based on a boundary condition correction model, as described in this invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0038] This embodiment 1 provides a method for dynamically adjusting the power limit of a power grid section based on a boundary condition correction model, such as... Figure 1 As shown, steps S1-S5 are included. Next, the implementation process of each step of the present invention will be described in detail.

[0039] S1. For the target transmission section, collect and process multi-source operation data that affect the transmission capacity of the target transmission section to form a standardized operation boundary condition time series, wherein each time t in the operation boundary condition time series corresponds to a feature vector b(t) that characterizes the operation state at that time.

[0040] The first step in implementing this invention is to collect, process, and standardize the operational boundary condition data for selected key transmission sections, constructing a feature sequence b(t) that can be directly used by subsequent models. The specific implementation process is as follows:

[0041] S1.1 Target Determination and Benchmark Setting:

[0042] First, transmission sections within the power grid that have a critical impact on inter-regional interconnection, power transmission, and system security are selected as the implementation targets of this invention. Based on the current "Guidelines for Power System Safety and Stability," stability limit regulations, and historical stability verification reports, the static benchmark power limit for each section is determined. This serves as the starting point and hard boundary for the dynamic adjustment of this invention. This serves as the reference starting point and safety baseline for dynamic adjustments in this invention. Simultaneously, the physical composition of this section needs to be identified and recorded, including its constituent lines, main transformer, and related DC pole groups, clarifying its topological connections to provide a structural foundation for subsequent boundary condition feature construction and safety verification.

[0043] S1.2 Multi-source operational data acquisition:

[0044] Based on the selection of cross-sections and the determination of benchmark limits, and adhering to the principle of "operational information that can significantly affect the safety boundary and transmission capacity of the cross-section," multi-dimensional and multi-time-scale operational data are collected from various production and management systems of the power grid to form the original set of operational boundary conditions. The data to be collected must include at least the following categories:

[0045] Source and load data: active load levels of the entire network and various regions, output of conventional units and new energy power plants, unit start-up and shutdown status, system reserve capacity and constraints.

[0046] Network status data: commissioning / outage / maintenance status of critical lines, main transformers, and DC pole groups; topology connection of the power grid (such as ring network opening / closing, tie line commissioning / deactivation).

[0047] Market and environmental data: inter-provincial planned electricity volume, spot market clearing results, and external power transmission and reception constraints that are closely related to cross-sectional power flow; as well as meteorological parameters that affect load and renewable energy output, such as temperature, wind speed, wind direction, and solar radiation.

[0048] The above data is acquired through interfaces of dispatch automation systems (EMS / SCADA), energy management systems, electricity market technical support systems, and professional meteorological service systems, forming a multi-source heterogeneous raw data stream.

[0049] S1.3 Data preprocessing and time series alignment:

[0050] The collected raw data is cleaned and integrated to prepare for feature encoding:

[0051] Time alignment: unify all data to the same sampling time step (e.g., 15 minutes, 1 hour).

[0052] Missing value handling: For data missing due to communication interruption or other reasons, reasonable completion is carried out by using methods such as interpolation of previous and next points and filling with historical data from the same period.

[0053] Outlier removal: Identify and correct outlier data points that significantly deviate from the normal range.

[0054] By preprocessing, the continuity and consistency of the data are ensured, forming a multivariate time series that strictly corresponds to the time index t.

[0055] S1.4, Encoding of operational boundary conditions:

[0056] The preprocessed data are encoded into dimensionless numerical features that the model can process, based on their physical properties and information types. This is the core step in constructing a unified feature vector b(t).

[0057] S1.4.1 Encoding of Continuous Operating Data. For continuously changing physical quantities such as load level, unit output, renewable energy output, temperature, and wind speed, denoted as the original quantity x(t). To eliminate differences in dimensions and numerical scales among different physical quantities, this invention, based on interpolation alignment and missing data completion at a unified time step, uses normalization or standardization methods to map them to a unified interval, for example:

[0058] Normalization: linearly scales values ​​to the [0,1] interval. For example, using min-max normalization:

[0059] ;

[0060] in, and These are the minimum and maximum values ​​of the physical quantity within the historical statistical interval.

[0061] Standardization: Transforming numerical values ​​into a distribution with a mean of 0 and a standard deviation of 1. For example:

[0062] ;

[0063] in, and These are the historical mean and standard deviation of the physical quantity, respectively. Through the above transformation, different dimensions and numerical scales...

[0064] All physical quantities are uniformly mapped to real-valued features, which facilitates joint modeling within the same time series model.

[0065] S1.4.2 Encoding of Periodic Time Data. For time information with obvious periodicity, such as hours, days of the week, and seasons, a sine-cosine function is used to encode the time index h(t) to explicitly depict the daily cycle, weekly cycle, and other patterns.

[0066] ;

[0067] ;

[0068] Where T is the length of the period (e.g., daily period T=24, weekly period T=168). This encoding transforms a single time scalar into a two-dimensional periodic feature to explicitly characterize the daily cycles and seasonal patterns of load and renewable energy output.

[0069] S1.4.3. Encoding of Discrete State Data. For categories or state information such as power grid topology, equipment maintenance, and operating mode, this invention adopts discrete encoding technology to uniformly represent them as classification features or binary indication features.

[0070] Binary coding: Applicable to two clearly defined states. For example, encoding the operational status of a critical railway line:

[0071] ;

[0072] Similarly, corresponding binary indication quantities can be constructed for main transformer shutdown, DC pole group operation mode, etc.

[0073] One-hot encoding: suitable for categorical variables with K possible values ​​(such as operating mode). Encode it as a K-dimensional binary vector e(t):

[0074] ;

[0075] Among them, when hour, ,otherwise .

[0076] Embedding Encoding: For categorical variables with high cardinality or inherent semantic association, the one-hot encoded vector e(t) can be mapped to a low-dimensional real-valued vector through a learnable embedding matrix E. This aims to improve the ability of time series models to represent complex categorical information.

[0077] S1.5 Construction of the feature vector b(t). All features obtained from the three types of encoding processes described above are concatenated at each time step t to form a unified high-dimensional feature vector b(t). Its structure can be represented as:

[0078] ;

[0079] in: The encoded values ​​(normalized values) represent M continuous physical quantities. Indicates periodic time encoding; This represents L binary state indicators; Represents the one-hot or embedding encoding vector of R categorical variables (which can be expanded into multiple dimensions).

[0080] Through the above steps, this invention transforms the multi-source, heterogeneous operational data affecting the cross-sectional power transmission capacity into a unified, numerically standardized, and easily processed operational boundary condition time series {b(1),b(2),...,b(T)}, laying a solid data foundation for subsequent model training and online applications. It contains not only physical quantity information, but also topology, maintenance and mode category information, forming a high-dimensional feature vector sequence with a unified structure that is easy to process by subsequent deep time series models, thus realizing standardized modeling and serialized expression of the boundary conditions of cross-section operation.

[0081] S2. Obtain the maximum transmittable power of the target transmission section in historical time or offline simulation scenario, and calculate the maximum transmittable power and the preset static reference limit power. The difference yields the true value of the limit correction amount for the corresponding time or scenario. The unified feature vector b(t) and the limit correction amount This constitutes the training sample set.

[0082] After modeling the boundary condition feature vector b(t), this invention constructs a supervised learning sample set for training the boundary condition correction model by combining historical data with offline simulation, and determines the true value of the limit correction amount corresponding to each sample. The specific process is as follows:

[0083] S2.1 Obtaining the maximum transmittable power.

[0084] To obtain the maximum power that a transmission section can actually safely transmit under different operating boundary conditions (i.e., the ultimate transmittable power) This invention employs a combination of the following two approaches to ensure the comprehensiveness and representativeness of sample coverage:

[0085] Approach 1: Engineering approximation based on historical operation records.

[0086] Records of the target section during past operations are extracted from the historical databases of the Power Grid Dispatch Automation System (EMS) and the Safety Verification System. For each historical moment... The following information will be retrieved simultaneously:

[0087] The eigenvector of the running boundary conditions at this moment (Obtained from step S1);

[0088] The cross-sectional transmittance capacity assessment value is obtained at or near this time through scheduling plans or safety verification modules (such as Safety Constrained Unit Combination (SCUC) or Safety Constrained Economic Scheduling (SCED), or through total transmittance capacity (TTC) assessment under preset safety and stability constraints such as N-1. .

[0089] This invention will Considered as being under this specific boundary condition Below, the maximum transmittable power of the cross section This is a reliable engineering approximation. The samples obtained through this approach are derived from actual operation and have a high degree of realism and credibility.

[0090] Approach 2: Supplementary analysis based on offline simulation.

[0091] To cover extreme loads, extreme renewable energy output, and complex maintenance methods that are less frequently encountered in historical data, and to compensate for the lack of historical data on a few or no observed operating conditions such as extreme loads, extreme renewable energy output, and complex equipment maintenance combinations, this invention constructs a supplementary scenario library on the power system offline simulation platform. The specific process is as follows:

[0092] Based on the characteristics of power grid operation and risk analysis, a series of representative boundary condition scenarios are set. Perform trend and stability analysis, paying particular attention to sparse corner areas in historical data.

[0093] For each set scenario Detailed power flow calculations and security and stability analyses were performed on a simulation platform. The constraints of the analysis strictly followed the "Guidelines for the Security and Stability of Power Systems" and the thermal stability limits of the equipment.

[0094] Under the premise of satisfying all safety constraints, the maximum power that the cross section can safely transmit in this scenario is determined by optimization algorithm or stepwise trial method, denoted as . This approach can systematically obtain the "true value" under extreme or special boundary conditions, greatly expanding the coverage of training samples in the boundary condition space.

[0095] S2.2 Calculation of the true value of the limit correction amount.

[0096] Achieving the maximum transmittable power Therefore, this invention does not directly use it as the model fitting target. From or To seamlessly integrate with the existing static quota system and constrain the output range of the model, this invention does not directly fit the absolute quota of the cross section. Instead, this invention uses the static benchmark quota power determined in step S1.1. To establish a unified reference point, calculate the true value of the limit correction amount. .

[0097] The calculation formula is:

[0098] ;

[0099] Among them: when From historical records ( (or offline simulation) This is the true value of the correction amount for that sample.

[0100] In some cases, if only a relatively conservative lower bound on the available transmission capacity can be obtained... (For example, considering a more stringent safety margin), the lower bound correction can be calculated. And use it as a conservative side constraint label in the model output.

[0101] This step converts the absolute power value into a correction value relative to the baseline, thus completing the preparation work for residual modeling. This indicates that power can be safely increased under the current operating conditions. This indicates that the quota needs to be maintained or tightened.

[0102] S2.3 Construction of supervised learning sample set.

[0103] In order to adapt to the deep time series prediction models (such as TiDE) that are capable of long sequence modeling, this invention adopts the sliding time window technique to organize the "feature-label" pairs at a single time point into continuous sequence samples.

[0104] Using a fixed historical window length L as the unit, a continuous segment is extracted from the running boundary condition time series as the model input. For example, at time t, the model input sequence is: .

[0105] Using the true value sequence of the limit correction amount at the current time and the next H times as the prediction target of the model, a time series sample set of boundary conditions and correction amounts covering normal and extreme operating conditions is formed, providing a physically meaningful and traceable training standard state basis for the boundary condition correction model. Corresponding to the above input... The output label sequence is: .

[0106] By sliding the time window forward from the beginning of the sequence and repeating the above truncation process, a large number of sequences like... The sample pairs of the boundary condition sequence and the correction sequence.

[0107] Finally, by fusing historical data and simulation supplementary data, and through the aforementioned serialization process, this invention constructs a high-quality supervised learning sample set that is comprehensive, has clear physical meaning, and includes temporal dependencies. This sample set provides a solid "truth" foundation for the next step of training a boundary condition correction model capable of accurately predicting future multi-step corrections.

[0108] S3. Using the time series of the operation boundary conditions within the historical period as input, and the true value series of the limit correction amount at the same time and in the future period as the output target, train the supervised learning time series prediction model (boundary condition correction model) to obtain the boundary condition correction model; when applied online, input the real-time and historical operation boundary condition time series into the boundary condition correction model to predict the limit correction amount prediction series for the future period.

[0109] After obtaining a structured training sample set, this invention constructs and trains a supervised learning time series prediction model, namely, a boundary condition correction model. The core task of this model is to learn a complex nonlinear mapping from historical boundary condition sequences to future limit correction sequence, as detailed below:

[0110] S3.1 Model Architecture Selection and Improvement.

[0111] In a preferred embodiment, the present invention uses TiDE (Time-series Dense Encoder) as the basic model architecture. TiDE is an encoding and decoding structure model based on multilayer perceptron (MLP). By densely encoding historical sequences and covariates and then decoding them in combination with known future covariates, it can efficiently and accurately complete long-range multi-step prediction without using self-attention or recurrent neural networks, and has advantages in computational efficiency and accuracy.

[0112] For the specific task of dynamic correction of cross-sectional limit power, this invention makes three key improvements to the basic TiDE model, forming the final boundary condition correction model.

[0113] S3.2 Residual Modeling

[0114] This invention does not directly allow the model to predict the absolute limit power of the cross section, but instead uses the residual modeling concept to predict the correction amount time series.

[0115] Given a sequence of operational boundary conditions within a historical time window (of length L): [b(t-L+1),...,b(t)]. Optionally, a sequence of known future covariates c(t:t+H) (such as a confirmed maintenance schedule, weather forecast, etc.) can be added.

[0116] Model (parameter is) The output of ) is a predicted sequence of quota adjustments for the next H+1 time periods: Its training objective is to approximate the true sequence of the limit correction amount obtained in step S2. .

[0117] When applied online, the predicted correction values ​​output by the model are used. With static benchmark limit Adding them together gives the predicted value of the dynamic limit power: .

[0118] This structure ensures that the model output is always a "correction" based on the existing procedural baseline value, which facilitates integration with the existing scheduling system and also allows for the use of ground truth data during the training phase. The natural constraint correction range is kept within the physical limit given by the offline stability check. ).

[0119] In other words, with the amount of correction To achieve this goal, let the improved TiDE be denoted as... Given a sequence of historical boundary conditions and future covariate sequences In the case of outputting the future Forecast of corrections for each time period:

[0120] ; and through Obtaining the predicted dynamic limit power ensures structurally that the model output is always a "correction" based on the existing standard benchmark value, which facilitates integration with the current limit system and also helps to control the correction magnitude during the training phase to not exceed the limit value given by the offline stability check.

[0121] S3.3, Multiquantile Risk Prediction.

[0122] Given that the cross-sectional limit is essentially an upper bound with a safety margin, this invention improves the model's output layer by introducing multiquantile prediction and risk-weighted loss into the TiDE output layer to enable the model to quantify the uncertainty of prediction and provide decision options under different risk levels. The specific improvements are as follows:

[0123] Let the model simultaneously output the predicted values ​​of the limit correction amount at multiple quantiles at each prediction time t+k, denoted as Typically, the output quantiles should include at least:

[0124] Median (q=0.5): Represents the "most likely" correction amount.

[0125] One or more conservative quantiles (e.g., q=0.9, 0.95): represent conservative adjustments that will not exceed the limit at higher confidence levels (e.g., 90%, 95%), and can be used to set risk-averse limits.

[0126] Risk-weighted loss function: To train this multiquantile output model, a combined loss function L is designed:

[0127] ;

[0128] Where: the first term is the mean squared error (MSE) loss for the median output, with a weight of . The second term is the quantile regression loss (Pinball Loss) for each quantile output. The weight is H represents the time step for the future period. Pinball loss ensures that approximately q% of the predicted q quantile values ​​in the model output are lower than the actual values. and The weighting coefficients used to balance the various losses are denoted by m, which represents the set quantile. For example, if q = 0.5, 0.8, 0.9, 0.95 (a total of 4 values), then m ranges from 1 to M, representing these M quantiles respectively.

[0129] This design allows the model to provide not only a "point estimate" but also a prediction range. Dispatchers can select appropriate correction amounts based on different risk preferences (such as using the median for daily optimization and a conservative quantile for critical power supply periods), thereby achieving dynamic limit adjustments with controllable risks.

[0130] In other words, in this way, the improved TiDE not only gives the most likely correction amount, but also the correction amount range under different risk levels of exceeding the limit. This is consistent with the TiDE practice of combining probability loss to quantify uncertainty in engineering practice. However, the output object is changed from general time series value to cross-sectional correction amount, so that the limit calculation can select the corresponding quantile as the final limit under given risk constraints.

[0131] Furthermore, considering that the closer the cross-section is to the limit operation boundary, the greater the impact of its correction error on safety, this invention introduces a boundary condition weighting function during TiDE training to give higher weights to samples that are close to the limit state.

[0132] S3.4, Sample safety margin weighted training.

[0133] Considering that the prediction error of the correction amount has a more significant impact on the system safety when the cross section is running near the limit boundary, this invention introduces a sample weighting mechanism in the model training stage, so that the model pays more attention to those "critical" samples with small safety margins.

[0134] For each sample in the training sample set, there exists an actual cross-sectional power flow for its corresponding operating condition. and limiting power The safety margin M of this sample is defined as: The smaller the M value, the closer the operating condition is to the safety limit.

[0135] Each sample is assigned a weight w, which is a monotonically decreasing function of the safety margin M. One specific implementation is as follows:

[0136] ;

[0137] in, and This is an adjustable parameter. This function ensures that the smaller the safety margin, the larger the weight w.

[0138] During training, the loss term for each sample is multiplied by its corresponding weight w to form the weighted total loss. By optimizing this weighted loss, the model is driven to achieve higher fitting accuracy near extreme boundary conditions with small safety margins, thereby improving its predictive reliability in high-risk regions.

[0139] S3.5, Model Training and Online Application.

[0140] Using the large-scale sample set constructed in step S2, with the weighted combined loss function as the objective, the improved boundary condition correction model is trained end-to-end using stochastic gradient descent or its variants (such as the Adam optimizer) until the model converges.

[0141] After model training, it is deployed to the production environment. During online runtime, the system aggregates the latest operational data in real time and generates feature vector sequences b(t) for the current and historical moments through the process described in S1. These sequences are then input into the trained boundary condition correction model. The model then outputs a predicted sequence of limit correction amounts for multiple future scheduling periods (e.g., the next 24 hours). This predicted sequence forms the basis for subsequent safety adjustments and final limit calculations. The improved TiDE achieves higher fitting accuracy near extreme boundary conditions, making it more suitable for driving dynamic adjustments to cross-sectional limits.

[0142] S4. Train an unsupervised learning model based on historical normal operation data to establish a reference model for the normal operation domain; input the current or predicted time series of operation boundary conditions into the reference model, calculate the degree to which the time series of operation boundary conditions deviates from the normal operation domain, and obtain an anomaly index. According to the aforementioned anomaly index Calculate a dynamic safety factor The safety factor The value varies with the degree of abnormality It increases and then decreases monotonically.

[0143] In obtaining the future correction sequence predicted by the boundary condition correction model Subsequently, to ensure safety under extreme or unprecedented operating conditions, this invention introduces unsupervised learning technology to establish a reference model of the normal operating domain. The deviation of real-time operating conditions from this normal domain is quantified into a dynamic safety factor, thereby automatically tightening limits under rare or unseen boundary conditions, enhancing safety margin management capabilities, and adaptively adjusting the predicted correction amount. The specific process is as follows:

[0144] S4.1 Construction of the normal operation domain reference model.

[0145] This invention utilizes a large number of historical data samples marked as safe operating periods to train an unsupervised learning model, which implicitly learns the normal data distribution or manifold structure of the operating boundary conditions.

[0146] Autoencoders or variational autoencoders are preferred as the underlying architecture. Autoencoders learn a compressed representation of the data through an "encode-decode" process; variational autoencoders further learn the probability distribution of the data.

[0147] Select a historical period that has been confirmed as safe and stable in operation, and extract the corresponding feature vector sequences of operational boundary conditions for these periods. This serves as the training set. The key point is that this training set should exclude known faults, extremes, or abnormal runtime periods to ensure that the model learns purely normal patterns.

[0148] For autoencoders, the model learns a mapping. ,in It is an encoder that compresses the high-dimensional input b(t) into a low-dimensional latent vector; It is a decoder that attempts to reconstruct the original input from the latent vectors. These are the model parameters. The training objective is to minimize the reconstruction error on all normal samples.

[0149] ;

[0150] By minimizing this loss, the model learns how to efficiently reconstruct the normal b(t) it has seen, enabling the model to have good reconstruction capabilities only within the "normal operation domain," thereby implicitly characterizing the normal manifold of the cross-sectional operation boundary conditions in the latent space.

[0151] For variational autoencoders, in addition to the reconstruction error, a KL divergence regularization term is added to the distribution of latent variables, forcing the latent variables to approximate a simple prior distribution (such as the standard normal distribution), thereby more smoothly characterizing the distribution features of the normal operating domain in a probabilistic sense.

[0152] Once trained, the model possesses the ability to describe "what the normal operating boundary conditions should look like." For normal operating conditions similar to the training set, the model can reconstruct them relatively accurately; however, for abnormal operating conditions that deviate from the normal pattern, the reconstruction error will increase significantly.

[0153] S4.2 Calculation of anomaly index.

[0154] During online operation, anomaly detection is performed on the current real-time or predicted future operating boundary condition sequence {b(t),...,b(t+H)} using a pre-trained unsupervised reference model.

[0155] The feature vector b(t+k) at each time step is input into the reference model to obtain its reconstruction output. Calculate the original reconstruction error at this moment:

[0156] ;

[0157] The squared value of the Euclidean distance (L2 norm) of this error directly measures the degree to which the current operating condition b(t+k) deviates from the "normal mode" identified by the model.

[0158] To make the anomaly index more robust and comparable, the original reconstruction error can be standardized. This can be achieved by using the mean reconstruction error obtained during the training phase on a historical set of normal samples. and standard deviation Calculate the standardized outlier:

[0159] ;

[0160] Standardized This can be intuitively understood as how many times the current operating condition deviates from the normal range compared to the historical normal fluctuation range.

[0161] This is the final anomaly index. The larger the value, the rarer the current or future operating boundary conditions are, the further they deviate from the historical normal operating range, and the more likely they belong to extreme operating conditions, complex maintenance combinations, or unknown operating modes that are not fully covered in historical data, with higher potential safety risks.

[0162] S4.3 Generation of dynamic safety factor.

[0163] Based on the calculated anomaly index This invention utilizes a preset monotonically decreasing function. This is mapped to a dynamic safety factor used for adjusting the correction amount. , .

[0164] Safety factor Must meet ,in, This is the preset minimum safety factor, representing the maximum possible tightening. (Function) Must satisfy: When When very small (close to normal), , representing the prediction of the basic trust model;

[0165] When it increases, It should decrease smoothly and monotonically, indicating a decrease in confidence in the forecast and a tightening of the safety margin.

[0166] For example, a typical safety factor function takes the form of exponential decay:

[0167] ;

[0168] in, It is an adjustable parameter that controls the degree of anomaly. Safety factor Sensitivity to influence. Parameters and It can be adjusted according to the system's risk tolerance.

[0169] To ensure absolute safety, a high anomaly threshold is set. .when When this condition is deemed an extreme abnormal situation, the rollback mechanism is triggered directly. At this point, the following options are available:

[0170] Set the safety factor to the lowest value directly: .

[0171] Alternatively, the final correction amount can be directly truncated to a preset conservative upper limit value.

[0172] In extreme cases, the correction amount can be completely ignored, and the system can directly revert to the static baseline limit. This mechanism is the last line of defense against out-of-distribution samples.

[0173] S5. Multiply the predicted limit correction value output by the boundary condition correction model by the dynamic safety factor at the corresponding time. The final correction amount after safety adjustment is obtained, and the final correction amount is compared with the static reference limit power. The sums are used to obtain the dynamic cross-sectional power limit for power grid safety monitoring and dispatching.

[0174] The specific process is as follows:

[0175] S5.1 Safety adjustment of correction amount and calculation of final limit.

[0176] After obtaining the original correction sequence predicted by the boundary condition correction model and the security factor sequence generated by the unsupervised security adjustment module Then, the present invention performs the final quota synthesis step.

[0177] For each future time t+k (k=0,1,...,H), the original correction value is predicted. Multiply by the dynamic safety factor at the corresponding time point This yields the final correction amount after safe contraction. : ;in, It is the final correction amount after safety adjustments. This operation enables the quantification and adaptive scaling of the prediction results.

[0178] Final correction amount Compared with static reference limit power Adding them together yields the dynamic cross-sectional power limit for power grid safety monitoring and dispatch at that moment. :

[0179] This formula is the core output expression of the method of this invention. It clearly embodies the residual idea of ​​"static benchmark + dynamic correction" and embeds a safety adjustment mechanism.

[0180] As a safety net, when anomalies When the preset extreme threshold is exceeded, a fallback mechanism is triggered. At this point, the system will no longer use the conventional calculation formula mentioned above, but will instead execute a more conservative strategy, such as:

[0181] Safety factor Forced to set to the minimum value Or directly use the final correction amount Cut off to a preset conservative upper limit. At the highest alert level, the correction amount can be completely ignored, making... The amount will be directly refunded to the static limit.

[0182] This mechanism ensures that, under extremely unknown conditions beyond the model's cognitive scope, the system's output always remains on the absolutely safe and conservative side.

[0183] S5.2 Collaborative Working Mechanism and System Value.

[0184] By leveraging the synergistic operation of the supervised prediction model in S3 and the unsupervised security adjustment module in S4, this invention achieves hierarchical and intelligent dynamic limit adjustment.

[0185] Normal optimization (trust prediction): When the running boundary conditions are within the historically common range (anomaly degree) Small, safety factor When the system fully trusts the accurate predictive capabilities of the monitoring model, it dynamically relaxes the cross-sectional limits, thereby maximizing the utilization efficiency of transmission capacity and improving the economic efficiency of the power grid and the level of new energy consumption.

[0186] Abnormal tightening (adaptive protection): When the operating boundary conditions begin to deviate from the normal state (abnormality) Increase the safety factor When the risk is reduced, the system automatically and quantitatively tightens the safety margin, "discounting" the predicted correction amount. This enables proactive risk prevention for rare or complex operating conditions, making the limit more conservative and enhancing system safety.

[0187] Extreme rollback (safety fallback): When encountering unprecedented extreme operating conditions (abnormality) When the threshold is exceeded, the system triggers a fallback mechanism to maintain the current state. To ensure a safe baseline, we must completely avoid radical and unsafe modifications caused by model extrapolation.

[0188] In summary, the collaborative closed loop constructed in this invention, combining primary intelligent prediction (S3) with secondary safety fallback (S4), is not a simple model stacking but rather forms an intrinsic, data-driven adaptive safety protection system. It provides a forward-looking dynamic limit trajectory in the time dimension and achieves a smooth transition from optimization to protection and then to fallback in the risk dimension. This significantly improves the robustness, adaptability, and overall controllability of the dynamic adjustment method for cross-sectional limits in the modern power grid environment characterized by high uncertainty in source loads and diverse operating modes, providing dispatching and operation personnel with intelligent decision support that combines efficiency and safety.

[0189] This invention, while maintaining the existing stable regulations and static cross-sectional quota system, introduces a structured modeling approach of "static benchmark quota + operational boundary condition correction amount". It adopts a deep time series prediction model with long-sequence modeling capabilities (such as TiDE / Autoformer-type structures with trend-seasonal decomposition and multi-step output capabilities) to directly learn the mapping from "boundary condition sequence → cross-sectional quota power correction amount sequence", thereby achieving continuous and forward-looking dynamic adjustment of cross-sectional quota power for multiple future scheduling periods. At the same time, by using quantile output to quantitatively correlate the correction amount with the risk of exceeding the limit, it can automatically provide a dynamic cross-sectional quota trajectory with controllable risk within safety constraints, which is different from the existing fixed-value configuration method based on static quota or rule tree tiered quota.

[0190] This invention introduces unsupervised learning to construct a "normal operating condition domain" for cross-section operation. By using models such as autoencoders to learn the boundary condition manifold on normal operating samples, the degree of deviation of real-time operating conditions from this manifold is transformed into an anomaly index, which is further mapped to a dynamic safety factor of the cross-section's limit power. The correction amount output by the deep time series model is adaptively contracted, thereby automatically tightening the limit and enhancing the safety margin under extreme boundary conditions or unseen conditions. This "unsupervised extreme operating condition identification + adaptive adjustment of safety factor" mechanism forms a collaborative closed loop with the supervised time series correction model. Compared with the existing approach that only relies on a fixed safety reduction ratio or manual experience correction, it can release the available transmission capacity of the cross-section as much as possible while ensuring safety.

[0191] Example 2 discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the dynamic adjustment method for grid cross-section limit power based on the boundary condition correction model as described in Example 1.

[0192] Example 3 discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for dynamic adjustment of grid cross-sectional limit power based on a boundary condition correction model as described in Example 1.

[0193] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamically adjusting the power limit of a power grid section based on a boundary condition correction model, characterized in that, include: For the target transmission section, multi-source operation data affecting the transmission capacity of the target transmission section are collected and processed to form a standardized operation boundary condition time series. Each time t in the operation boundary condition time series corresponds to a feature vector b(t) that characterizes the operation state at that time. Obtain the maximum transmittable power of the target transmission section at a historical time or in an offline simulation scenario, and calculate the maximum transmittable power and the preset static reference limit power. The difference yields the true value of the limit correction amount for the corresponding time or scenario. The feature vector b(t) and the limit correction amount Construct the training sample set; Using the time series of the operational boundary conditions within a historical period as input, and the true value series of the quota correction amount at the same time and in future periods as output targets, a supervised learning time series prediction model is trained to obtain a boundary condition correction model. When applied online, the real-time and historical operational boundary condition time series are input into the boundary condition correction model to predict the quota correction amount prediction series for future periods. An unsupervised learning model is trained based on historical normal operation data to establish a reference model for the normal operation domain. The current or predicted time series of operational boundary conditions are input into the reference model to calculate the degree to which the time series deviates from the normal operation domain, thus obtaining an anomaly index. According to the aforementioned anomaly index Calculate a dynamic safety factor The safety factor The value varies with the degree of abnormality The value increases and then monotonically decreases; The predicted limit correction value output by the boundary condition correction model is multiplied by the dynamic safety factor at the corresponding time. The final correction amount after safety adjustment is obtained, and the final correction amount is compared with the static reference limit power. The sums are used to obtain the dynamic cross-sectional power limit for power grid safety monitoring and dispatching.

2. The method for dynamic adjustment of power grid cross-sectional limit based on a boundary condition correction model according to claim 1, characterized in that, The specific process of constructing the time series of the running boundary conditions is as follows: Normalize or standardize continuous operational data; For periodic time data, periodic encoding is performed by converting between sine and cosine functions; For discrete state data, binarization, one-hot coding, or embedded vector coding are used for processing. The processed features from each dimension are concatenated into a unified feature vector b(t).

3. The method for dynamic adjustment of power grid cross-sectional limit based on a boundary condition correction model according to claim 1, characterized in that, The specific process for obtaining the maximum transmittable power of the target transmission section at a historical moment or in an offline simulation scenario is as follows: From the historical operation records of the power grid dispatch automation system and the safety verification system, extract the transmittable capacity assessment value of the target section obtained by safety constraint unit combination, safety constraint economic dispatch or total transmission capacity assessment at the corresponding historical moment and under the predetermined N-1 safety and stability constraints, and use it as the limit transmittable power at that historical moment. And / or, On the offline simulation platform, power flow calculation and safety and stability analysis are performed for the set operating boundary conditions. Under the premise of meeting the power system safety and stability guidelines and equipment thermal stability constraints, the maximum power that the target section can safely transmit is solved as the limit transmittable power under the simulation scenario.

4. The method for dynamic adjustment of power grid cross-sectional limit based on a boundary condition correction model according to claim 1, characterized in that, The boundary condition correction model employs a deep neural network model with an encoding / decoding structure. The input to this deep neural network model is a sequence of operational boundary condition features within a historical time window, and the output is a predicted sequence of limit correction amounts for a future time window. Furthermore, the deep neural network model uses the static baseline limit power... Residual modeling is performed for reference.

5. The method for dynamic adjustment of power grid cross-sectional limit based on a boundary condition correction model according to claim 1, characterized in that, The boundary condition correction model is configured to simultaneously output multiple probability quantile values ​​for the limit correction amount in future periods, wherein the probability quantiles include a median value and at least one conservative estimate greater than the median; and the training loss function of the boundary condition correction model combines the mean squared error loss for the median value and the quantile regression loss for each quantile.

6. A method for dynamic adjustment of power grid cross-sectional limit based on a boundary condition correction model according to claim 4 or 5, characterized in that, When training the boundary condition correction model, a weight is assigned to each training sample. The weight is negatively correlated with the safety margin of the operating condition corresponding to the training sample, so that the training sample with a smaller safety margin contributes more to the training loss function during the training process.

7. The method for dynamic adjustment of power grid cross-sectional limit based on a boundary condition correction model according to claim 1, characterized in that, The unsupervised learning model is a reference model constructed using an autoencoder or a variational autoencoder; the anomaly index The error norm is measured by calculating the output vector reconstructed from the reference model of the current running boundary condition feature vector and comparing it with the original input vector.

8. A method for dynamic adjustment of power grid cross-sectional limit based on a boundary condition correction model according to claim 1 or 7, characterized in that, The safety factor The calculation satisfies: ,in, The minimum safety factor is preset; when the anomaly index When the preset threshold is exceeded, a fallback mechanism is activated to reduce the safety factor. Set as Alternatively, the final correction amount may be truncated to a preset conservative upper limit.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for dynamic adjustment of power grid cross-section limit power based on the boundary condition correction model as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for dynamic adjustment of grid cross-sectional power limit based on the boundary condition correction model as described in any one of claims 1 to 8.

Images