Electric power system steady state operation analysis method

Abstract

The invention discloses an electric power system steady-state operation analysis method, which is applied to the technical field of electric power equipment, and comprises the following steps: acquiring basic data of an electric power system, and constructing a standardized data set and a power grid topological structured model based on the basic data of the electric power system; historical operation data is obtained, a physical mechanism steady-state model is constructed based on the power grid topological structured model and a preset steady-state physical mechanism equation, the physical mechanism steady-state model is calibrated based on the historical operation data, and a target mechanism steady-state model is obtained; performing feature extraction on the standardized data set based on preset electric power system operation semantics to obtain a first feature set, and constructing a data driving model based on the first feature set, the target mechanism steady-state model and a preset machine learning model; and fusing the target mechanism steady-state model and the data driving model based on a preset fusion algorithm and a preset residual compensation mechanism to obtain a steady-state operation prediction model, and performing steady-state operation analysis on the power system based on the steady-state operation prediction model.

Description

Technical Field

[0001] This invention relates to the field of power grid stability analysis technology, and in particular to a method for analyzing the steady-state operation of a power system. Background Technology

[0002] Steady-state operation of a power system is fundamental to ensuring the safe, economical, and high-quality power supply of the power grid. Under steady-state conditions, the system achieves real-time balance between active and reactive power, and the frequency and voltage at each node are maintained within allowable fluctuation ranges, thereby ensuring that power quality meets the requirements of users and sensitive loads. Simultaneously, steady-state operation provides reliable references for relay protection, automation devices, and power flow calculations, ensuring the accuracy and coordination of fault diagnosis and protection actions. Good steady-state operation also reduces thermal and mechanical stresses on equipment caused by overload or frequent fluctuations, extends the service life of generators, transformers, and switchgear, and reduces operation and maintenance costs. In the context of the electricity market, stable steady-state operation is also a prerequisite for economic dispatch, optimal power flow, and reserve capacity allocation, contributing to achieving the lowest operating costs and controllable risks. With the high proportion of renewable energy and distributed energy integration, steady-state operation analysis must also consider the long-term impact of inverter characteristics and output fluctuations in order to formulate effective dispatch, energy storage, and voltage and frequency support strategies. In conclusion, maintaining and monitoring steady-state operation is not only a necessary condition for grid reliability but also a key link in achieving the goals of safe, economical, and sustainable operation.

[0003] Existing steady-state operation analysis methods for power systems, primarily based on physical mechanisms, heavily rely on precise equipment parameters and network topology. They are insufficient in characterizing fluctuations in renewable energy output, time-varying loads, and uncertainties, and struggle to quantify model errors and uncertainty propagation. Furthermore, these methods do not adequately consider the real-time impact of dynamic events such as topology switching and fault isolation, and parameter calibration often relies on experience, resulting in low adaptability and automation. While data-driven methods offer advantages in computational speed and pattern recognition, they typically neglect hard physical constraints such as power balance and node voltage constraints. Moreover, they are prone to producing predictions that do not conform to physical laws when training data distributions drift or extreme operating conditions are scarce. Therefore, as the scale and operational complexity of the power grid increase, the computational burden of purely mechanistic methods rises significantly, while purely data-driven methods suffer from insufficient reliability and interpretability, making it difficult to achieve a balance between computational accuracy, efficiency, and operational reliability. Summary of the Invention

[0004] This invention provides a method for analyzing the steady-state operation of a power system, which reduces computational complexity while ensuring the accuracy of the steady-state operation analysis.

[0005] To address the aforementioned technical problems, this invention provides a method for analyzing the steady-state operation of a power system, comprising:

[0006] Collect basic power system data, and construct a standardized dataset and a grid topology structured model based on the power system basic data;

[0007] Historical operating data is acquired, and a physical mechanism steady-state model is constructed based on the power grid topology structured model and the preset steady-state physical mechanism equation. The physical mechanism steady-state model is then calibrated based on the historical operating data to obtain the target mechanism steady-state model.

[0008] Based on the preset power system operation semantics, feature extraction is performed on the standardized dataset to obtain a first feature set, and a data-driven model is constructed based on the first feature set, the target mechanism steady-state model, and the preset machine learning model.

[0009] The target mechanism steady-state model and the data-driven model are fused based on a preset fusion algorithm and a preset residual compensation mechanism to obtain a steady-state operation prediction model. The steady-state operation prediction model is then used to perform steady-state operation analysis on the power system.

[0010] The technical features of this invention in collecting basic power system data and constructing standardized datasets and structured power grid topology models ensure the consistency, traceability, and topological semantic integrity of input data, thus providing a reliable data foundation for subsequent modeling. Furthermore, by constructing a target mechanism steady-state model based on power grid topology and preset steady-state physical mechanism equations, and calibrating it with historical data, the invention ensures that the model conforms to physical laws and fits actual operating conditions, improving the accuracy and interpretability of mechanism calculations. Moreover, feature extraction based on power system operating semantics and data-driven model construction enhances the model's comprehensive learning ability for temporal, static, and topological correlation information, thereby improving its ability to capture complex nonlinear relationships and implicit patterns. Finally, a preset fusion algorithm and residual compensation mechanism are used to fuse the mechanism model and the data model, effectively balancing the physical constraints of the mechanism model with the predictive capabilities of the data model, improving computational accuracy, robustness, and engineering applicability. Ultimately, the obtained steady-state operation prediction model achieves a balance between accuracy, reliability, and real-time performance, which is beneficial for actual scheduling and operational decision support.

[0011] Furthermore, the collection of basic power system data, and the construction of a standardized dataset and a grid topology structured model based on the basic power system data, include:

[0012] Collect basic data of the power system, including raw data, power grid topology diagrams, and uncertainty statistics.

[0013] Data cleaning is performed on the raw data and uncertain statistical data to obtain a standard dataset;

[0014] The power grid topology drawing is analyzed using the standard dataset to obtain a structured model of the power grid topology.

[0015] This invention collects technical features including raw time-series data, topology maps, and uncertainty statistics, comprehensively covering static and dynamic factors affecting steady-state calculations and improving the completeness of subsequent analysis from the source. By cleaning the raw data and uncertainty statistics to obtain the technical features of a standard dataset, data quality is improved, and the impact of noise and anomalies on model training and mechanism calculations is reduced. Furthermore, by combining the standard dataset with topology analysis of the topology maps to obtain the technical features of a structured model, the power grid topology is automated, structured, and can be directly called by the model, thereby reducing errors from manual analysis and improving model construction efficiency and consistency.

[0016] Furthermore, the step of cleaning the raw data and uncertainty statistics to obtain a standard dataset includes:

[0017] The uncertain statistical data is quantified to obtain statistical quantification data;

[0018] Anomaly removal and missing data completion are performed on the original data and the statistical quantification data to obtain a standard dataset.

[0019] This invention quantifies uncertain statistical data, enabling uncertainties such as wind power, photovoltaics, load, and topology events to be formally represented and incorporated into subsequent uncertainty modeling and robustness analysis, thereby enhancing the model's adaptability to operational fluctuations. Furthermore, by implementing anomaly removal and missing data completion on raw and statistically quantified data, the negative impact of dirty data on mechanistic model calibration and data-driven model training can be significantly reduced, ensuring the representativeness and reliability of training samples and mechanistic calibration samples, thereby improving overall prediction accuracy and stability.

[0020] Furthermore, the step of acquiring historical operating data, constructing a physical mechanism steady-state model based on the power grid topology structured model and a preset steady-state physical mechanism equation, and calibrating the physical mechanism steady-state model based on the historical operating data to obtain a target mechanism steady-state model includes:

[0021] Obtain historical operational data;

[0022] Based on the aforementioned grid topology structured model, a steady-state physical mechanism equation for the power system is constructed, and based on the steady-state physical mechanism equation, a steady-state physical mechanism model is constructed.

[0023] The physical mechanism steady-state model is calibrated based on the historical operating data to obtain the target mechanism steady-state model.

[0024] This invention provides sufficient measured samples and extreme operating condition samples for the mechanism model by acquiring historical operating data, thus providing statistical support for model calibration. It constructs steady-state physical mechanism equations based on a grid topology structured model, ensuring that the model strictly adheres to the topological and physical constraints of the power system, thereby improving interpretability and engineering usability. Furthermore, calibrating the mechanism model based on historical operating data to obtain the target mechanism steady-state model helps eliminate parameter biases, correct model errors, and enable the mechanism model to better reflect actual operating conditions, thereby improving the baseline accuracy and reliability of subsequent fusion results.

[0025] Furthermore, the step of constructing the steady-state physical mechanism equations of the power system based on the grid topology structured model, and constructing a steady-state physical mechanism model based on the steady-state physical mechanism equations, includes:

[0026] Based on the power grid topology structured model, the connection relationships of nodes, branches, and power grid equipment in the power system are obtained;

[0027] A steady-state physical mechanism equation is constructed based on the connection relationship between the nodes, branches, and power grid equipment.

[0028] The node voltages and branch power of the power system are calculated based on the steady-state physical mechanism equations, and a steady-state physical mechanism model is constructed based on the node voltages and branch power.

[0029] This invention obtains the connection relationships of nodes, branches, and equipment based on the topology, which makes the subsequent equation construction have clear entity mapping and facilitates location. Furthermore, it constructs steady-state physical mechanism equations based on these connection relationships, ensuring that the established mathematical model corresponds one-to-one with the actual power grid structure, thereby avoiding calculation deviations caused by structural mismatches. It further calculates node voltages and branch power based on the equations and constructs a steady-state physical mechanism model, which enables the mechanism model to directly output the key quantities required for scheduling and monitoring, facilitating point-by-point comparison and fusion with data-driven results, thereby improving system-level analysis capabilities and fault sensitivity.

[0030] Furthermore, the calibration of the physical mechanism steady-state model based on the historical operating data to obtain the target mechanism steady-state model includes:

[0031] Based on the historical operating data, historical node voltage data and historical branch power data are obtained;

[0032] Based on the historical node voltage data, historical branch power data, and the least squares method, the parameters of the power grid equipment in the physical mechanism steady-state model are calibrated to obtain the target mechanism steady-state model.

[0033] This invention uses historical measurement data as the basis for calibration, which can align model parameters with actual operating characteristics and reduce deviations caused by equipment aging, parameter labeling errors, or model simplification. Thus, by adjusting parameters according to the least squares method, it provides a quantifiable, stable, and easy-to-implement parameter optimization method, thereby improving the numerical accuracy of the mechanistic model and the consistency of steady-state calculations. This provides a more reliable baseline reference for data-driven models and improves the accuracy of the final fusion prediction.

[0034] Furthermore, the step of extracting features from the standardized dataset based on preset power system operation semantics to obtain a first feature set, and constructing a data-driven model based on the first feature set, the target mechanism steady-state model, and a preset machine learning model, includes:

[0035] Based on the preset power system operation semantics, feature extraction is performed on the standardized dataset to obtain time series features, static parameter features and original field features of topology structure, and to generate a candidate feature set;

[0036] The candidate feature set is normalized based on preset standardization rules to obtain the first feature set;

[0037] A data-driven model is constructed based on the first feature set, the target mechanism steady-state model, and the preset machine learning model.

[0038] This invention extracts features based on the operational semantics of the power system, ensuring that the selected features have physical meaning and engineering interpretability, thereby improving the model's ability to understand the power grid operation mechanism. The candidate features are then normalized according to a unified rule to obtain a first feature set, ensuring consistency in numerical scale during training and online inference, which facilitates faster convergence and improves model generalization. This feature set, the target mechanism model, and a pre-set machine learning model are used to construct a data-driven model, enabling the data-driven model to absorb both topological and temporal information, while also using the mechanism output as a reference, thus enhancing its physical perception and predictive stability.

[0039] Furthermore, the step of constructing a data-driven model based on the first feature set, the target mechanism steady-state model, and a preset machine learning model includes:

[0040] The first feature set is input into the target mechanism steady-state model to obtain the prediction result corresponding to the first feature set, and training samples are generated based on the first feature set and the prediction result.

[0041] An initial data-driven model is constructed based on a preset learning model, and the initial data-driven model is trained based on the training samples to obtain a data-driven model.

[0042] This invention utilizes the technical feature of predicting the first feature set and generating training results using a target mechanism model, providing rich and physically-background labels for the data-driven model, which helps to alleviate the problem of insufficient measurement data or the influence of noise; and uses this to construct training samples and train the initial data-driven model, enabling the data model to learn system nonlinearities and empirical correction terms that are difficult for the mechanism model to characterize, thereby improving its adaptability to extreme working conditions and unseen situations.

[0043] Furthermore, the process of fusing the target mechanism steady-state model and the data-driven model based on a preset fusion algorithm and a preset residual compensation mechanism to obtain a steady-state operation prediction model, and performing steady-state operation analysis of the power system based on the steady-state operation prediction model, includes:

[0044] The first prediction result of the steady-state model of the target mechanism is obtained based on a preset time window, and the first error is calculated based on the first prediction result;

[0045] The second prediction result of the data-driven model is obtained based on the time window, and the second error is calculated based on the second prediction result;

[0046] The residual values ​​of the target mechanism steady-state model and the data-driven model are determined based on the first prediction result and the second prediction result;

[0047] The fusion weights are determined based on the first error and the second error combined with a preset weight allocation function;

[0048] Based on the residual value, the preset residual compensation mechanism, and the fusion weight, the target mechanism steady-state model and the data-driven model are fused to obtain a steady-state operation prediction model. Based on the steady-state operation prediction model, the power system is subjected to steady-state operation analysis.

[0049] This invention acquires predictions from two types of models and calculates short-term errors within a time window, enabling the fusion decision to have a "short-term performance awareness" capability, thus dynamically responding to the time-varying nature of model performance. Furthermore, by determining the fusion weights based on the first and second errors combined with a preset weight allocation function, it helps to allocate confidence levels according to real-time performance, avoiding the bias caused by long-term fixed weights and making full use of the model that performs better in the short term. Thus, by fusing based on residuals, residual compensation mechanisms, and fusion weights, the difference information between the two can be explicitly processed during fusion, preserving the physical consistency of the mechanistic model while introducing correction information from the data model, thereby improving overall prediction accuracy and reducing the frequency of extreme biases.

[0050] Furthermore, the process of fusing the target mechanism steady-state model and the data-driven model based on the residual value, a preset residual compensation mechanism, and fusion weights to obtain a steady-state operation prediction model, and performing steady-state operation analysis of the power system based on the steady-state operation prediction model, includes:

[0051] The residual value is compared with a preset residual threshold.

[0052] When the residual value is less than or equal to the preset residual threshold, the target mechanism steady-state model and the data-driven model are fused based on the fusion weight to form an initial fusion model, and the initial fusion model is used as the steady-state operation prediction model.

[0053] When the residual value is greater than the residual threshold, the target mechanism steady-state model and the data-driven model are fused based on the fusion weight to obtain an initial fusion model, and the initial fusion model is corrected based on the residual value and the preset residual compensation coefficient to obtain a steady-state operation prediction model.

[0054] The steady-state operation prediction model is used to perform steady-state operation analysis on the power system.

[0055] This invention compares the residuals with a preset threshold and processes them according to the threshold branch, which can trigger a more cautious correction process to prevent error propagation when the residuals are abnormal. When the residuals are small, the initial fusion model is obtained directly based on the fusion weights, which simplifies the calculation and improves real-time performance when the model is consistent. When the residuals are large, the initial fusion model is corrected by applying the residual compensation coefficient. The useful correction driven by data is incorporated into the final result through the controllable correction amount. At the same time, the adjustable compensation intensity avoids the destruction of power balance or voltage constraints. Therefore, the reliability of the fused model to sudden changes and abnormal situations is significantly improved while ensuring that physical constraints are met. Attached Figure Description

[0056] Figure 1 A flowchart illustrating a power system steady-state operation analysis method provided in an embodiment of the present invention;

[0057] Figure 2 This invention provides a schematic diagram of the structure of a power system steady-state operation analysis system. Detailed Implementation

[0058] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0059] The terms "first" and "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0060] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0061] Example 1

[0062] See Figure 1 , Figure 1 This is a flowchart illustrating a power system steady-state operation analysis method provided by an embodiment of the present invention. The embodiment of the present invention provides a power system steady-state operation analysis method, including steps 101 to 104, as detailed below:

[0063] Step 101: Collect basic power system data, and construct a standardized dataset and a grid topology structured model based on the basic power system data;

[0064] In this embodiment, the collection of basic power system data and the construction of a standardized dataset and a grid topology structured model based on the basic power system data include:

[0065] Collect basic data of the power system, including raw data, power grid topology diagrams, and uncertainty statistics.

[0066] Data cleaning is performed on the raw data and uncertain statistical data to obtain a standard dataset;

[0067] The power grid topology drawing is analyzed using the standard dataset to obtain a structured model of the power grid topology.

[0068] In this embodiment, the raw power system data includes parameters of new energy power plants, historical load data, and parameters of power grid equipment. The collection frequency of parameters of new energy power plants (such as wind speed and photovoltaic irradiance) is set to 1 minute / time, the collection frequency of historical load data is set to 5 minutes / time, and the collection frequency of parameters of power grid equipment (such as transformer impedance and line resistance) is once per quarter.

[0069] In this embodiment, the statistical data on uncertain factors include the fluctuation range of renewable energy output, load change patterns, and topology switching records. The fluctuation range of renewable energy output and load change patterns are updated once a day, while topology switching records are collected in real time (recorded immediately when a topology switching event is triggered).

[0070] In this embodiment, the parameters of the new energy power plants and the historical load data come from the power dispatch automation system (SCADA system), the parameters of the power grid equipment come from the equipment management system (PMS system), and the power grid topology drawings come from the power grid planning and design platform; the data interface is compatible with the power industry standard IEC61850, the original data format is uniformly CSV format, and the power grid topology drawing format is CAD format (AutoCAD 2020 and above).

[0071] In this embodiment, parameters of new energy power plants and historical load data are collected from the power dispatch automation system (SCADA), power grid equipment parameters are obtained from the power management system (PMS), and power grid topology drawings are obtained from the power grid planning and design platform. Simultaneously, a real-time connection is established with the event / topology recording system to receive topology switching events. To ensure data format and interface consistency, the acquisition module supports the IEC61850 protocol and converts all time-series raw data into a unified CSV format for storage. The topology drawings are archived in CAD (AutoCAD 2020 and above) format.

[0072] In this embodiment, parameters of new energy power plants (such as wind speed and photovoltaic irradiance) are collected every minute (e.g., once every 1 minute); historical load data are collected every 5 minutes; and grid equipment parameters (such as transformer impedance and line resistance) are updated quarterly. The explicit definition of the above collection frequency and data sources ensures the temporal resolution, coverage, and data integrity for subsequent modeling.

[0073] In this embodiment, the technical features collected, including raw time-series data, topology maps, and uncertainty statistics, can comprehensively cover the static and dynamic factors affecting steady-state calculations, thereby improving the completeness of subsequent analysis from the source. By cleaning the raw data and uncertainty statistics to obtain the technical features of the standard dataset, data quality is improved, and the impact of noise and anomalies on model training and mechanism calculation is reduced. Furthermore, the technical features of the structured model are obtained by combining the standard dataset with the topology map for topology analysis, making the power grid topology automated, structured, and directly callable by the model, thereby reducing manual analysis errors and improving model construction efficiency and consistency.

[0074] In this embodiment, the step of cleaning the raw data and uncertain statistical data to obtain a standard dataset includes:

[0075] The uncertain statistical data is quantified to obtain statistical quantification data;

[0076] Anomaly removal and missing data completion are performed on the original data and the statistical quantification data to obtain a standard dataset.

[0077] In this embodiment, the uncertainty statistics include wind power output, photovoltaic power output, load time series statistics, and topology switching records. Different methods are used to quantify wind power output, photovoltaic power output, load time series statistics, and topology switching records to obtain statistical quantification data.

[0078] In this embodiment, the collected uncertainty statistics are first formatted and divided into time windows. Then, based on the selected analysis window, the parameters of various uncertainty quantification models are estimated and updated. Wind power output is modeled using a two-parameter Weibull distribution, and shape and scale parameters are obtained through maximum likelihood estimation (MLE) or least squares fitting methods. Photovoltaic output is fitted with a normalized Beta distribution to estimate its α / β parameters. Load fluctuations are described using a normal distribution, and the mean μ and standard deviation σ after amplification are calculated using historical data from the same period. Topology switching records are quantized using binary 0 / 1 sequences, and the trigger time and duration of each switch are recorded to calculate the switching percentage. The quantized outputs are saved in the form of a structured parameter table (including distribution type, parameter estimates, estimation time window, and confidence interval), and are archived in versions by day or by analysis trigger frequency for model retrieval.

[0079] In this embodiment, wind power output is quantitatively analyzed using a Weibull distribution. The two-parameter Weibull distribution is used to describe the probabilistic characteristics of wind power output, and the probability density function is:

[0080]

[0081] Where v is the actual wind power output (unit: MW), k is the shape parameter (fitted based on wind power output data of the past year, with a value range of 1.8-2.5), and c is the scale parameter (fitted based on wind power output data of the past year, with a value of 0.7-0.9 times the rated wind power output).

[0082] In this embodiment, for wind farms, a two-parameter Weibull distribution is used to describe their power output probability characteristics. When estimating parameters, historical power output data with a resolution of minutes / five minutes over the past year is preferred. The recommended range for the shape parameter k is 1.8–2.5. The initial estimate of the scale parameter c can be set at 0.7–0.9 times the rated power output of the wind farm and fine-tuned by historical fitting.

[0083] In this embodiment, photovoltaic output is quantitatively analyzed (Beta distribution). The Beta distribution is used to describe the fluctuation of photovoltaic output, and the probability density function is:

[0084]

[0085] Where p is the normalized value of photovoltaic output (p = actual output / rated output, ranging from 0 to 1); α and β are shape parameters (fitted based on photovoltaic output data from the past year, α ∈ [2.0, 3.0], β ∈ [1.5, 2.5]); B(α, β) is the Beta function. Γ(·) is the gamma function.

[0086] In this embodiment, for a photovoltaic field, the actual power output is first normalized to p∈[0,1] (p=actual power output / rated power output), and then the Beta distribution is used to fit the α and β parameters. The recommended initial range is α∈[2.0,3.0] and β∈[1.5,2.5]. The fitting can use moment estimation or maximum likelihood estimation. After fitting, the goodness-of-fit index (such as KS test or residual sum of squares) for both types of distribution fitting is calculated and saved for subsequent model selection and anomaly alarm.

[0087] In this embodiment, the load variation pattern is quantified (normal distribution). The normal distribution is used to describe load fluctuations, and the probability density function is:

[0088]

[0089] Where L is the actual load value (unit: MW); μ is the average load value (take the average load value of the same period in the past year); σ is the standard deviation of the load (take 1.2 times the standard deviation of the load in the same period in the past year).

[0090] In this embodiment, when quantifying load and topology switching, the load often exhibits obvious patterns in intraday and seasonal structures. Quantification is first performed by calculating the average load μ based on the same historical period (e.g., a short period of the same month and day), and then multiplying the standard deviation of the same period sample over the past year by an amplification factor (recommended factor 1.2) to obtain σ. A normal distribution is used to characterize short-term load fluctuations. For topology switching, an event-driven log is used to record the type, start and end times, affected area, and duration of each switching. The event sequence is then binarized (1 indicates a switching occurred within the analysis time window, and 0 indicates no switching occurred). At the same time, the proportion of switching to the analysis period is statistically analyzed as a quantitative indicator of topology uncertainty, so that topology changes can be included as conditional variables in uncertainty scanning or scenario analysis during steady-state analysis.

[0091] In this embodiment, the original data and data with uncertainties are cleaned, including removing outliers and filling in missing data, to obtain a standardized dataset (including subsets of new energy, load, and equipment parameters).

[0092] In this embodiment, preliminary statistics are first performed on each time series field, and candidate outliers exceeding the [μ-3σ, μ+3σ] interval (3σ principle) are marked. Then, physical constraints are verified on these candidate outliers: for example, the load must not exceed the rated capacity of the transformer, the output of new energy from a single station must not exceed the installed capacity of the power plant, and the node voltage must not exceed the allowable amplitude range. Samples that violate both statistical criteria and physical constraints are directly marked as outliers and removed. Samples that only violate statistical criteria but meet physical probability are retained as extreme cases and marked as "extreme event samples" for subsequent extreme case training. The outlier removal process generates an outlier log, recording the sample ID, the reason for removal, and the reviewer or the basis for the automatic strategy decision, ensuring auditability.

[0093] In this embodiment, different completion methods and criteria are defined for short-term and long-term missing data. For short-term missing data (missing duration ≤ 30 minutes) at the minute or hour level, linear interpolation or neighbor-value forward or backward filling is used to restore the continuity of the time series; simultaneously, the completion method and confidence level are marked in the data element after filling. For long-term missing data (missing duration > 30 minutes), historical similar working condition mapping filling is used: data of similar time periods are retrieved from the historical database based on meteorological conditions (e.g., similar irradiance / wind speed), similar load levels, and holiday or workday labels, and then replaced. All completion actions record the filling source, time window, and confidence level so that weight or uncertainty adjustments can be made during model training or online inference.

[0094] In this embodiment, when cleaning statistical quantitative data, if the instantaneous output of a certain new energy power station exceeds the specified confidence level (e.g., 99%) of its fitted distribution and does not violate physical constraints, it is considered a statistical anomaly but may be a real extreme. This sample is retained in the "extreme sample pool" for subsequent training. If the sample is missing, alternative samples generated by random sampling from its distribution model or quantiles of the distribution can be used as filler values, and the weight of the sample is adjusted after filling to reflect its synthetic attributes.

[0095] In this embodiment, the final output standard dataset includes a subset of new energy sources, a subset of loads, a subset of equipment parameters, and a subset of uncertainty quantification. Each record is accompanied by data quality metadata: original source, whether interpolation or padding is used, anomaly identifier, confidence level, and version number.

[0096] In this embodiment, the power grid topology drawing is analyzed using a standardized dataset to construct a power grid topology model, and the relationships between nodes, branches, and equipment are marked to obtain a structured power grid topology model.

[0097] In this embodiment, the power grid topology drawing is first matched with the node, equipment, and measurement point information in the standardized dataset. Based on the equipment identifier, coordinates, and IEC61850 / CSV interface fields, the power grid components such as buses (nodes), branches (lines, bus couplers, tie lines), transformers, circuit breakers or disconnectors, instrument transformers, and parallel compensation devices are automatically identified. Then, based on the graphical elements and standardized attributes, connectivity relationships are constructed to generate a directed or undirected graph representation with nodes as vertices and branches as edges (while exporting the adjacency matrix and branch-node association table). Each node, branch, and equipment is labeled with structured attribute fields (e.g., node type: balancing node, generating node, load node; voltage level, rated capacity, transformer impedance, line resistance, susceptance, length, heat flow limit, switch status, measurement point ID, and timestamp, etc.), thereby obtaining a structured model of the power grid topology.

[0098] In this embodiment, by quantifying the statistical data of uncertainty, uncertainties such as wind power, photovoltaics, load, and topology events can be formally represented and incorporated into subsequent uncertainty modeling and robustness analysis, thereby enhancing the model's adaptability to operational fluctuations. Furthermore, by implementing anomaly removal and missing data completion on the original data and statistically quantified data, the negative impact of dirty data on mechanistic model calibration and data-driven model training can be significantly reduced, ensuring the representativeness and reliability of training samples and mechanistic calibration samples, thereby improving the overall prediction accuracy and stability.

[0099] Step 102: Obtain historical operating data, construct a physical mechanism steady-state model based on the power grid topology structured model and the preset steady-state physical mechanism equation, and calibrate the physical mechanism steady-state model based on the historical operating data to obtain the target mechanism steady-state model;

[0100] In this embodiment, the steps of acquiring historical operating data, constructing a physical mechanism steady-state model based on the power grid topology structured model and a preset steady-state physical mechanism equation, and calibrating the physical mechanism steady-state model based on the historical operating data to obtain a target mechanism steady-state model include:

[0101] Obtain historical operational data;

[0102] Based on the aforementioned grid topology structured model, a steady-state physical mechanism equation for the power system is constructed, and based on the steady-state physical mechanism equation, a steady-state physical mechanism model is constructed.

[0103] The physical mechanism steady-state model is calibrated based on the historical operating data to obtain the target mechanism steady-state model.

[0104] In this embodiment, when constructing the physical mechanism steady-state model, based on the grid topology structured model, Kirchhoff's laws and power balance equations are first used to establish the nodal power balance equations of the power system. These equations include active power balance equations and reactive power balance equations, where the power balance equation for each node describes the relationship between the injected power and the node voltage. Then, the Newton-Raphson method is used to solve the problem, iteratively calculating the voltage magnitude and branch power of each node.

[0105] In this embodiment, the technical characteristics of acquiring historical operating data provide sufficient measured samples and extreme operating condition samples for the mechanism model, enabling statistical support for model calibration. Steady-state physical mechanism equations are constructed based on the grid topology structured model to ensure that the model strictly follows the topological and physical constraints of the power system, thereby improving interpretability and engineering usability. Furthermore, calibrating the mechanism model based on historical operating data to obtain the target mechanism steady-state model helps to eliminate parameter deviations, correct model errors, and enable the mechanism model to better reflect the actual operating conditions, thereby improving the baseline accuracy and reliability of subsequent fusion results.

[0106] In this embodiment, the step of constructing the steady-state physical mechanism equations of the power system based on the grid topology structured model, and constructing a steady-state physical mechanism model based on the steady-state physical mechanism equations, includes:

[0107] Based on the power grid topology structured model, the connection relationships of nodes, branches, and power grid equipment in the power system are obtained;

[0108] A steady-state physical mechanism equation is constructed based on the connection relationship between the nodes, branches, and power grid equipment.

[0109] The node voltages and branch power of the power system are calculated based on the steady-state physical mechanism equations, and a steady-state physical mechanism model is constructed based on the node voltages and branch power.

[0110] In this embodiment, the physical mechanism steady-state model can be constructed based on Kirchhoff's laws and power balance equations, using classical steady-state calculation methods such as the Newton-Raphson method and the fast decomposition method. Furthermore, the Newton-Raphson method is preferentially used to construct the physical mechanism steady-state model because this algorithm has high convergence accuracy in power system power flow calculations (node ​​voltage calculation error ≤ 0.5%), making it suitable for complex power grids with renewable energy integration. When the number of power grid nodes exceeds 500, the fast decomposition method (based on PQ decomposition, improving convergence speed by 30% to meet the computational efficiency requirements of large power grids) is switched to. Specifically:

[0111] Construct the nodal power balance equations. For the i-th node in the power system, the relationship between injected power and node voltage is as follows:

[0112] Active power balance:

[0113]

[0114] Reactive power balance:

[0115]

[0116] Among them, P i Q i P represents the injected active power and injected reactive power at node i (units: MW, MVar), respectively; Gi Q Gi P represents the active power output and reactive power output of the generator at node i (units: MW, MVar), respectively; Li Q Li V represents the active power and reactive power of the load at node i (units: MW, MVar); |V i |、|V j | represents the voltage magnitude (in kV) at nodes i and j, respectively; δ ij =δ i -δ j ,δ i δ j G represents the voltage phase angles (in rad) at nodes i and j, respectively; ij B ij These are the conductance and susceptance elements (unit: S) in the i-th row and j-th column of the node admittance matrix; n is the total number of nodes in the power grid.

[0117] In this embodiment, the connection relationships of nodes, branches, and equipment are obtained based on the topology, which makes the subsequent equation construction have clear entity mapping and facilitates location. Then, steady-state physical mechanism equations are constructed based on these connection relationships to ensure that the established mathematical model corresponds one-to-one with the actual power grid structure, thereby avoiding calculation deviations caused by structural mismatches. Furthermore, node voltages and branch power are calculated based on the equations and a steady-state physical mechanism model is constructed, so that the mechanism model can directly output the key quantities required for scheduling and monitoring, which facilitates point-by-point comparison and fusion with data-driven results, thereby improving system-level analysis capabilities and fault sensitivity.

[0118] In this embodiment, calibrating the physical mechanism steady-state model based on the historical operating data to obtain the target mechanism steady-state model includes:

[0119] Based on the historical operating data, historical node voltage data and historical branch power data are obtained;

[0120] Based on the historical node voltage data, historical branch power data, and the least squares method, the parameters of the power grid equipment in the physical mechanism steady-state model are calibrated to obtain the target mechanism steady-state model.

[0121] In this embodiment, during the iterative solution process, the power balance equation is first linearized to obtain the Jacobian matrix and the iterative correction formula. The calculation of the elements of the Jacobian matrix includes the partial derivatives of each node with respect to the voltages of adjacent nodes, with key terms such as the partial derivatives of active power with respect to voltage phase angle and reactive power with respect to voltage magnitude. These partial derivatives are used to update the calculation of node voltages and branch power, ensuring that each iteration can get closer to the actual steady-state solution of the power grid.

[0122] In this embodiment, the power balance equation is linearized to obtain the iterative correction formula:

[0123] [ΔS]=[J][ΔX], that is, [ΔX]=[J] -1 [ΔS] (6)

[0124] Among them: [ΔS]=[ΔP1,ΔP2,...,ΔP n ,ΔQ1,ΔQ2,...,ΔQ n ] T ,ΔP i =P i calc -P i spec , These represent the deviations between the calculated active and reactive power values ​​and the given values, respectively; [ΔX] = [Δδ1, Δδ2, ..., Δδ] n ,Δ|V1|,Δ|V2|,...,Δ|V n |] T , represents the correction amount for voltage phase angle and amplitude; [J] is the Jacobian matrix, and its elements are calculated as follows (taking a 2×2 sub-block as an example):

[0125] The partial derivative of the active power with respect to the phase angle is:

[0126]

[0127] J Pδ,ii =∑ j≠i |V i ||V j |(G ij sinδ ij -B ij cosδ ij (8)

[0128] The partial derivative of active power with respect to voltage amplitude is:

[0129]

[0130] J P|V| ,ii=2|V i|G ii +∑ j≠i |V i |(G ij cosδ ij +B ij sinδ ij (10)

[0131] The partial derivative of reactive power with respect to phase angle is:

[0132]

[0133] J Qδ,ii =∑ j≠i |V i ||V j |(G ij cosδ ij +B ij sinδ ij (12)

[0134] The partial derivative of reactive power with respect to voltage amplitude is:

[0135]

[0136] J Q|V| ,ii=-2|V i |B ii +∑ j≠i |V i |(G ij sinδ ij -B ij cosδ ij (14)

[0137] In this embodiment, the iterative correction formula calculates the power deviation (ΔP, ΔQ) of the current iteration step and uses the Jacobian matrix to make corrections, gradually approximating the accurate solutions of node voltage and branch power until the preset convergence accuracy requirement is reached; that is, when the power deviation of all nodes meets the condition that the maximum error is less than a set threshold (e.g., 0.5%), or the node voltage calculation error is less than or equal to 0.5%, the iteration process terminates and the steady-state calculation result is output.

[0138] In this embodiment, when max(|ΔP) is satisfied i |,|ΔQ i |)≤∈(∈ is the convergence accuracy threshold, taken as 10 -4 The iteration terminates when the node voltage calculation error is ≤0.5% (MW / MVar) or ≤0.5%, and the steady-state calculation result V is output. cal,i ,P cal,i V cal,i ,P cal,iThe values ​​are the node voltage (in kV) and branch power (in MW) of the i-th group calculated by the physical mechanism steady-state model, respectively.

[0139] In this embodiment, historical data is used to evaluate the accuracy of the steady-state calculation model of the physical mechanism during parameter calibration. By adjusting parameters such as transformer impedance and line resistance, the deviation between the calculated values ​​of the physical mechanism model and the actual operating data is minimized, ensuring that the sum of squared errors is minimized.

[0140] In this embodiment, the calibration of the model parameters includes "least squares method and historical data iterative calibration". Historical power grid operation data from the past 6 months (including more than 200 sets of normal operating conditions and more than 50 sets of extreme operating conditions) is selected. The objective function is "minimizing the sum of squared errors between the calculated values ​​of the physical mechanism steady-state calculation model and the actual operating data". Parameters such as transformer impedance and line resistance are adjusted accordingly. After calibration, the error thresholds are set as follows: node voltage calculation error ≤ 0.8%, branch power calculation error ≤ 1.0%. If these requirements are not met, the historical data sample size is increased (supplemented to 300 sets of normal operating condition data) for recalibration until the error requirements are met, ultimately forming a calibrated model parameter library.

[0141] In this embodiment, the formula with the objective function of "minimizing the sum of squared errors between the calculated values ​​of the physical mechanism steady-state calculation model and the actual operating data" is as follows:

[0142]

[0143] Where J is the sum of squared errors; n is the number of historical data samples (n≥200, including normal and extreme operating conditions); V cal,i ,P cal,i V represents the node voltage (kV) and branch power (MW) of the i-th group calculated by the steady-state calculation model based on the physical mechanism. act,i ,P act,i These represent the historical actual node voltage and branch power of the i-th group, respectively; the calibration constraints are as follows:

[0144] |V cal,i -V act,i | / V act,i ≤0.8% (16)

[0145] |P cal,i -P act,i | / P act,i ≤1.0% (17)

[0146] In this embodiment, iterative calibration primarily involves comparing historical data with the calculation results of the physical mechanism steady-state calculation model to further adjust the model's parameters. First, historical power grid operation data from the past six months is selected, including at least 200 sets of normal operating conditions and 50 sets of extreme operating conditions. This data will be used to evaluate the accuracy of the physical mechanism steady-state calculation model. The calibration objective is to minimize the sum of squared errors between the calculated results and the actual historical data, thereby precisely adjusting various parameters in the power grid model, such as transformer impedance and line resistance. Iterative calibration is performed using the least squares method until the node voltage calculation error is ≤0.8% and the branch power calculation error is ≤1.0%. If the error does not reach the set threshold, the sample size of historical data is increased until the accuracy requirements are met, ultimately forming an accurate calibration parameter library.

[0147] In this embodiment, historical measurement data is used as the basis for calibration, which can align the model parameters with the actual operating characteristics and reduce deviations caused by equipment aging, parameter labeling errors, or model simplification. Thus, parameter adjustment based on the least squares method provides a quantifiable, stable, and easy-to-implement parameter optimization method, thereby improving the numerical accuracy of the mechanistic model and the consistency of steady-state calculations. This provides a more reliable baseline reference for the data-driven model and improves the accuracy of the final fusion prediction.

[0148] Step 103: Extract features from the standardized dataset based on the preset power system operation semantics to obtain a first feature set, and construct a data-driven model based on the first feature set, the target mechanism steady-state model, and the preset machine learning model;

[0149] In this embodiment, the step of extracting features from the standardized dataset based on preset power system operation semantics to obtain a first feature set, and constructing a data-driven model based on the first feature set, the target mechanism steady-state model, and a preset machine learning model, includes:

[0150] Based on the preset power system operation semantics, feature extraction is performed on the standardized dataset to obtain time series features, static parameter features and original field features of topology structure, and to generate a candidate feature set;

[0151] The candidate feature set is normalized based on preset standardization rules to obtain the first feature set;

[0152] A data-driven model is constructed based on the first feature set, the target mechanism steady-state model, and the preset machine learning model.

[0153] In this embodiment, key features affecting steady-state operation (such as peak output of new energy sources, load factor, and number of topological connections) are extracted from the standardized dataset, and feature normalization and dimensionality reduction are performed to obtain a feature set.

[0154] In this embodiment, a feature catalog is first defined based on the semantics of power system operation: time-series features include instantaneous values ​​of renewable energy output sequences, rolling window statistics, peak-to-valley ratio, ramp rate / ramp rate, seasonal decomposition components, and frequency domain features; static parameter features include bus voltage levels, transformer rated capacity and impedance, branch length and rated current, and device type; topology features are derived from the power grid topology model, such as nodal degree, nodal voltage level grouping, branch connectivity component identification, electrical distance based on conductance matrix, betweenness centrality of branches and nodes, and shortest path length to the nearest dispatch node or grid connection point. This candidate feature set also includes derived fields output from the physical mechanism steady-state model, such as node voltage predictions, active and reactive power predictions, and branch power flow direction markings calculated by the mechanism model. These fields are used to introduce physical information into the subsequent data-driven model.

[0155] In this embodiment, a missing value handling strategy is first applied to each candidate feature, and then Z-Score standardization is used to unify the scale, ensuring consistency between features of different dimensions. Subsequently, redundant information is compressed through feature selection and dimensionality reduction processes to obtain a robust and informative first feature set.

[0156] In this embodiment, Z-Score normalization is used for feature normalization:

[0157]

[0158] Among them, z j x is the normalized value of the j-th feature; j μ is the original value of the j-th feature; j The sample mean of the j-th feature ( m is the total number of samples); σ j The sample standard deviation of the j-th feature

[0159]

[0160] In this embodiment, feature extraction is performed based on the operational semantics of the power system to ensure that the selected features have physical meaning and engineering interpretability, thereby improving the model's ability to understand the power grid operation mechanism. Then, the candidate features are normalized according to a unified rule to obtain the first feature set, ensuring that the numerical scale is consistent in training and online inference, which is conducive to accelerating convergence and improving the model's generalization. The data-driven model is constructed together with this feature set, the target mechanism model, and the preset machine learning model, which enables the data-driven model to absorb both topological and temporal information, and to use the mechanism output as a reference, thereby enhancing its physical perception ability and prediction stability.

[0161] In this embodiment, the step of constructing a data-driven model based on the first feature set, the target mechanism steady-state model, and a preset machine learning model includes:

[0162] The first feature set is input into the target mechanism steady-state model to obtain the prediction result corresponding to the first feature set, and training samples are generated based on the first feature set and the prediction result.

[0163] An initial data-driven model is constructed based on a preset learning model, and the initial data-driven model is trained based on the training samples to obtain a data-driven model.

[0164] In this embodiment, the first feature set is first fed into the calibrated target mechanism steady-state model according to a preset time window, and the output vector y of the target mechanism steady-state model is calculated and recorded. mech This includes target quantities such as node voltage and branch power; subsequently, a training sample vector S = {X} is constructed for each time sample. first ,y mech}, where X first Let y be the first feature set vector after normalization. mech This is the mechanism prediction vector.

[0165] In this embodiment, an initial data-driven model is constructed based on random forest, gradient boosting tree, or deep learning model.

[0166] In this embodiment, the initial data-driven model is trained based on the training sample vector. After training, the data-driven model is evaluated on the retained validation set, and indicators such as MAPE, RMSE, and constraint violation rate are calculated. Post-processing corrections are then performed to strictly meet key physical constraints, such as adjusting branch power proportionally to restore power balance, or ensuring voltage range constraints through a second minimum adjustment.

[0167] In this embodiment, the technical features of using the target mechanism model to predict the first feature set and generate training results provide rich and physically-background labels for the data-driven model, which helps to alleviate the problem of insufficient measurement data or the influence of noise; and use this to construct training samples and train the initial data-driven model, so that the data model can learn the system nonlinearity and empirical correction terms that are difficult for the mechanism model to characterize, thereby improving the adaptability to extreme working conditions and unseen situations.

[0168] Step 104: Based on the preset fusion algorithm and the preset residual compensation mechanism, fuse the target mechanism steady-state model and the data-driven model to obtain a steady-state operation prediction model, and perform steady-state operation analysis on the power system based on the steady-state operation prediction model.

[0169] In this embodiment, the process of fusing the target mechanism steady-state model and the data-driven model based on a preset fusion algorithm and a preset residual compensation mechanism to obtain a steady-state operation prediction model, and performing steady-state operation analysis of the power system based on the steady-state operation prediction model, includes:

[0170] The first prediction result of the steady-state model of the target mechanism is obtained based on a preset time window, and the first error is calculated based on the first prediction result;

[0171] The second prediction result of the data-driven model is obtained based on the time window, and the second error is calculated based on the second prediction result;

[0172] The residual values ​​of the target mechanism steady-state model and the data-driven model are determined based on the first prediction result and the second prediction result;

[0173] The fusion weights are determined based on the first error and the second error combined with a preset weight allocation function;

[0174] Based on the residual value, the preset residual compensation mechanism, and the fusion weight, the target mechanism steady-state model and the data-driven model are fused to obtain a steady-state operation prediction model. Based on the steady-state operation prediction model, the power system is subjected to steady-state operation analysis.

[0175] In this embodiment, by using mechanisms such as weighted fusion and residual compensation, the hard constraints of the physical mechanism steady-state calculation model (such as power balance) are combined with the predictive ability of the data-driven model to optimize the calculation accuracy and form a hybrid-driven power system steady-state operation calculation method (including feature set, fusion mechanism, and trained model).

[0176] In this embodiment, the first prediction result y of the steady-state model of the target mechanism is first acquired in parallel according to a preset time window (e.g., T = 5 minutes or T = 15 minutes). mech,i The second prediction result y of the data-driven model data,i For the sample set within this time window, calculate the real-time error indices for the two types of models: mechanistic model error e1 and data model error e2.

[0177] In this embodiment, dynamic weight allocation is adopted, and the weights are determined based on the real-time error of the mechanistic model and the data-driven model:

[0178]

[0179]

[0180] Where w1 represents the weights of the physical mechanism steady-state calculation model; w2 represents the weights of the data-driven model; e1 represents the calculation error of the physical mechanism steady-state calculation model (using the mean absolute percentage error); e2 represents the prediction error of the data-driven model; y mech,i y data,i y act,i These represent the calculated value from the steady-state calculation model of the i-th physical mechanism, the predicted value from the data-driven model, and the historical actual value, respectively.

[0181] In this embodiment, when calculating the initial residual value, a residual vector is calculated for each target component to be estimated (such as node voltage and branch power):

[0182] Δy=y data,i -y mech,i (twenty three)

[0183] In this embodiment, weights are assigned based on the error magnitude of each model, with the weights of the mechanistic model and the data-driven model dynamically adjusted according to their current performance. Typically, when the data-driven model performs better, its weight increases relatively, and vice versa, when the mechanistic model performs better, its weight increases relatively. This weighted fusion mechanism not only flexibly addresses prediction needs in different scenarios but also effectively reduces the risks associated with the bias of a single model.

[0184] In this embodiment, predictions from two types of models are acquired and short-term errors are calculated according to a time window, enabling the fusion decision to have "short-term performance awareness" and thus dynamically respond to the time-varying nature of model performance. Furthermore, the fusion weights are determined based on the first and second errors combined with a preset weight allocation function, which helps to allocate confidence based on real-time performance. This avoids the bias caused by long-term fixed weights and makes full use of the model that performs better in the short term. Therefore, fusion is performed based on residuals, residual compensation mechanisms, and fusion weights, which can explicitly process the difference information between the two models during fusion. This preserves the physical consistency of the mechanistic model and introduces the correction information of the data model, thereby improving the overall prediction accuracy and reducing the frequency of extreme biases.

[0185] In this embodiment, the step of fusing the target mechanism steady-state model and the data-driven model based on the residual value, a preset residual compensation mechanism, and fusion weights to obtain a steady-state operation prediction model, and performing steady-state operation analysis of the power system based on the steady-state operation prediction model, includes:

[0186] The residual value is compared with a preset residual threshold.

[0187] When the residual value is less than or equal to the preset residual threshold, the target mechanism steady-state model and the data-driven model are fused based on the fusion weight to form an initial fusion model, and the initial fusion model is used as the steady-state operation prediction model.

[0188] When the residual value is greater than the residual threshold, the target mechanism steady-state model and the data-driven model are fused based on the fusion weight to obtain an initial fusion model, and the initial fusion model is corrected based on the residual value and the preset residual compensation coefficient to obtain a steady-state operation prediction model.

[0189] The steady-state operation prediction model is used to perform steady-state operation analysis on the power system.

[0190] In this embodiment, the mechanistic model and the data-driven model may produce significant deviations in certain situations, especially under extreme conditions. Therefore, the system automatically calculates the residuals of each model after the fusion process. When these residuals exceed a preset threshold, the system triggers a residual compensation mechanism. The purpose of this mechanism is to ensure the accuracy and physical plausibility of the predictions by adjusting the deviations in the model prediction results. By compensating for the residuals, the system can effectively correct prediction errors and improve the stability and reliability of predictions.

[0191] In this embodiment, after each model prediction result is obtained, it is first determined whether the residual exceeds a preset threshold. If the residual is below the threshold, the system directly uses the weighted fusion prediction result to output the final steady-state operation analysis result. If the residual exceeds the threshold, it indicates that the difference between the models is large, and the system enters the residual compensation stage. The core of the residual compensation mechanism is that the system corrects the prediction result according to the magnitude and distribution of the residual. Specifically, the system adjusts the model with large deviations by introducing a compensation coefficient to correct the fusion result of the two models. These adjustments are not arbitrary but take into account the physical constraints of the power grid, such as power balance, voltage amplitude limits, and line heat flow limits. In this way, the residual compensation mechanism can not only improve the accuracy of the prediction result but also ensure that the calculation result conforms to the physical and operational constraints of the power grid.

[0192] In this embodiment, the residual compensation mechanism is triggered when the residual value exceeds the true value by 0.5%, i.e., |Δy| / y act,i >0.5%.

[0193] In this embodiment, when the residual value exceeds 0.5% of the true value, the steady-state operation prediction model is as follows:

[0194] y final,i =w1y mech,i +w2y data,i +k×Δy (24)

[0195] Among them, y final,i This represents the final calculation result for the i-th group of hybrid drives; k is the residual compensation coefficient (initial value 0.8, adjusted to 0.5-0.7 if the power balance constraint is violated); power balance constraint: |∑P in,i -∑P out,i |≤0.1MW(∑P in,i For the total input power of the region, ∑P out,i (Total output power of the region).

[0196] In this embodiment, by comparing the residuals with a preset threshold and processing according to the threshold branch, a more cautious correction process can be triggered when the residuals are abnormal to prevent error propagation. When the residuals are small, the initial fusion model is obtained directly based on the fusion weights, which simplifies the calculation and improves real-time performance when the models are consistent. When the residuals are large, the initial fusion model is corrected by applying the residual compensation coefficient. The useful correction driven by data is incorporated into the final result through a controllable correction amount. At the same time, the power balance or voltage constraint is avoided by adjusting the compensation intensity. Therefore, the reliability of the fused model to sudden changes and abnormal situations is significantly improved while ensuring that the physical constraints are met.

[0197] In this embodiment, to ensure the accuracy and robustness of the algorithm, the system comprehensively evaluates its performance. First, the accuracy evaluation uses the Mean Absolute Percentage Error (MAPE) as the primary indicator. This indicator measures the deviation between the model's predicted results and actual observations. The calculated results are compared with actual data to derive an overall error level. During the evaluation process, the MAPE value is required to not exceed 2.0% to ensure the algorithm's accuracy in practical applications. This accuracy requirement guarantees that the system can provide sufficiently accurate prediction results when processing steady-state operation analysis of power systems, thereby meeting the high accuracy requirements of power grid operation.

[0198] In this embodiment, the accuracy index (MAPE) is:

[0199]

[0200] In this embodiment, the acceptable threshold is: MAPE ≤ 2.0%.

[0201] In this embodiment, to evaluate the robustness of the model, the MAPE change rate (CR) after disturbance is also introduced as a robustness indicator. During this evaluation process, a 5% random disturbance is added to the input data to simulate different possible data deviations or anomalies. The change rate is calculated by comparing the MAPE after disturbance with the MAPE under normal conditions. If the MAPE value after disturbance remains within a reasonable range (the change rate does not exceed 15%), the system can still maintain good robustness even with incomplete data or noise. Specifically, when the MAPE value after disturbance does not exceed 2.3%, it indicates that the system has good robustness, can adapt to power system data under different operating conditions, and ensures that it can still provide stable and reliable prediction results when facing uncertainties.

[0202]

[0203] Where CR is the rate of change of MAPE; MAPE normal MAPE under normal input data; MAPE disturb Add 5% random perturbation (x) to the input data. disturb,i =x normal,i MAPE after ×(1+0.05×∈), ∈∈[-1,1]);

[0204] Acceptable threshold: CR ≤ 15% (i.e., MAPE) disturb ≤2.3%.

[0205] In this embodiment, through these two performance evaluations, the system not only verifies the accuracy of the algorithm under normal data, but also confirms its stability and adaptability under extreme or uncertain data conditions, thereby meeting the practical application needs in complex power grid environments.

[0206] In this embodiment, by collecting basic power system data and constructing standardized datasets and a structured power grid topology model, the consistency, traceability, and topological semantic integrity of the input data are guaranteed, thus providing a reliable data foundation for subsequent modeling. Furthermore, by constructing a target mechanism steady-state model based on the power grid topology and preset steady-state physical mechanism equations, and calibrating it with historical data, the model ensures that it conforms to physical laws and fits the actual operating state, improving the accuracy and interpretability of mechanism calculations. Moreover, feature extraction based on power system operating semantics and data-driven model construction enhances the model's comprehensive learning ability for temporal, static, and topological correlation information, thereby improving its ability to capture complex nonlinear relationships and implicit patterns. Finally, a preset fusion algorithm and residual compensation mechanism are used to fuse the mechanism model and the data model, effectively balancing the physical constraints of the mechanism model with the predictive capabilities of the data model, improving computational accuracy, robustness, and engineering applicability. Ultimately, the obtained steady-state operation prediction model achieves a balance between accuracy, reliability, and real-time performance, which is beneficial for actual scheduling and operational decision support.

[0207] Please refer to Figure 2 , Figure 2 The present invention provides a schematic diagram of a power system steady-state operation analysis system, comprising: a data acquisition module 201, a mechanism model construction module 202, a driving model construction module 203, and a fusion module 204;

[0208] The data acquisition module is used to collect basic power system data and construct a standardized dataset and a grid topology structured model based on the basic power system data.

[0209] The mechanism model construction module is used to acquire historical operating data, construct a physical mechanism steady-state model based on the power grid topology structured model and the preset steady-state physical mechanism equation, and calibrate the physical mechanism steady-state model based on the historical operating data to obtain the target mechanism steady-state model.

[0210] The driving model construction module is used to extract features from the standardized dataset based on the preset power system operation semantics, obtain a first feature set, and construct a data-driven model based on the first feature set, the target mechanism steady-state model, and the preset machine learning model.

[0211] The fusion module is used to fuse the target mechanism steady-state model and the data-driven model based on a preset fusion algorithm and a preset residual compensation mechanism to obtain a steady-state operation prediction model, and to perform steady-state operation analysis on the power system based on the steady-state operation prediction model.

[0212] In this embodiment of the invention, a terminal device is also provided, 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 above-described power system steady-state operation analysis method.

[0213] In this embodiment of the invention, a computer-readable storage medium is also provided, which includes a stored computer program, wherein the computer program controls the device where the computer-readable storage medium is located to execute the above-described power system steady-state operation analysis method when it is running.

[0214] 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 descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that 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 for those skilled in the art.

Claims

1. A method of steady state operation analysis of an electric power system, characterized by, The method comprises the following steps: collecting basic data of a power system, and constructing a standardized data set and a power grid topology structured model based on the basic data of the power system; obtaining historical operation data, constructing a physical mechanism steady-state model based on the power grid topology structured model and a preset steady-state physical mechanism equation, and calibrating the physical mechanism steady-state model based on the historical operation data to obtain a target mechanism steady-state model; extracting features from the standardized data set based on a preset power system operation semantic to obtain a first feature set, and constructing a data-driven model based on the first feature set, the target mechanism steady-state model and a preset machine learning model; fusing the target mechanism steady-state model and the data-driven model based on a preset fusion algorithm and a preset residual error compensation mechanism to obtain a steady-state operation prediction model, and performing steady-state operation analysis on the power system based on the steady-state operation prediction model.

2. A method of steady state operation analysis of an electric power system as claimed in claim 1, characterized in that, The method comprises the following steps: collecting basic data of a power system, and constructing a standardized data set and a power grid topology structured model based on the basic data of the power system; collecting basic data of the power system, wherein the basic data comprises original data, power grid topology drawings and uncertainty statistical data; performing data cleaning on the original data and the uncertainty statistical data to obtain a standard data set; 3. A method of steady state operation analysis of an electric power system as claimed in claim 2, characterized in that, performing topology analysis on the power grid topology drawings in combination with the standard data set to obtain a power grid topology structured model. The method comprises the following steps: quantitatively processing the uncertainty statistical data to obtain statistical quantitative data; 4. A method of steady state operation analysis of an electric power system as claimed in claim 3, characterized in that, performing abnormal rejection processing and missing completion processing on the original data and the statistical quantitative data to obtain a standard data set. The method comprises the following steps: obtaining historical operation data; constructing a steady-state physical mechanism equation of the power system based on the power grid topology structured model, and constructing a physical mechanism steady-state model based on the steady-state physical mechanism equation; 5. A method of steady state operation analysis of an electric power system as claimed in claim 4, c h a r a c t e r i s e d b y calibrating the physical mechanism steady-state model based on the historical operation data to obtain a target mechanism steady-state model. The method comprises the following steps: obtaining historical operation data; constructing a steady-state physical mechanism equation of the power system based on the power grid topology structured model, and constructing a physical mechanism steady-state model based on the steady-state physical mechanism equation; 6. A method of steady state operation analysis of an electric power system as claimed in claim 5, characterized in that, calibrating the physical mechanism steady-state model based on the historical operation data to obtain a target mechanism steady-state model. The method comprises the following steps: obtaining historical operation data; constructing a steady-state physical mechanism equation of the power system based on the power grid topology structured model, and constructing a physical mechanism steady-state model based on the steady-state physical mechanism equation; calibrating the physical mechanism steady-state model based on the historical operation data to obtain a target mechanism steady-state model. The method comprises the following steps: obtaining historical operation data; constructing a steady-state physical mechanism equation of the power system based on the power grid topology structured model, and constructing a physical mechanism steady-state model based on the steady-state physical mechanism equation; calibrating the physical mechanism steady-state model based on the historical operation data to obtain a target mechanism steady-state model. Based on the historical node voltage data, historical branch power data, and the least squares method, the parameters of the power grid equipment in the physical mechanism steady-state model are calibrated to obtain the target mechanism steady-state model.

7. A method of steady state operation analysis of an electric power system as claimed in claim 6, c h a r a c t e r i z e d b y The step of extracting features from the standardized dataset based on preset power system operation semantics to obtain a first feature set, and constructing a data-driven model based on the first feature set, the target mechanism steady-state model, and a preset machine learning model, includes: Based on the preset power system operation semantics, feature extraction is performed on the standardized dataset to obtain time series features, static parameter features and original field features of topology structure, and to generate a candidate feature set; The candidate feature set is normalized based on preset standardization rules to obtain the first feature set; A data-driven model is constructed based on the first feature set, the target mechanism steady-state model, and the preset machine learning model.

8. A method of steady state operation analysis of an electric power system as claimed in claim 7, c h a r a c t e r i z e d b y The step of constructing a data-driven model based on the first feature set, the target mechanism steady-state model, and a preset machine learning model includes: The first feature set is input into the target mechanism steady-state model to obtain the prediction result corresponding to the first feature set, and training samples are generated based on the first feature set and the prediction result. An initial data-driven model is constructed based on a preset learning model, and the initial data-driven model is trained based on the training samples to obtain a data-driven model.

9. A method of steady state operation analysis of an electric power system as claimed in claim 8, c h a r a c t e r i s e d b y The process involves fusing the target mechanism steady-state model and the data-driven model based on a preset fusion algorithm and a preset residual compensation mechanism to obtain a steady-state operation prediction model. Based on this steady-state operation prediction model, steady-state operation analysis of the power system is performed, including: The first prediction result of the steady-state model of the target mechanism is obtained based on a preset time window, and the first error is calculated based on the first prediction result; The second prediction result of the data-driven model is obtained based on the time window, and the second error is calculated based on the second prediction result; The residual values ​​of the target mechanism steady-state model and the data-driven model are determined based on the first prediction result and the second prediction result; The fusion weights are determined based on the first error and the second error combined with a preset weight allocation function; Based on the residual value, the preset residual compensation mechanism, and the fusion weight, the target mechanism steady-state model and the data-driven model are fused to obtain a steady-state operation prediction model. Based on the steady-state operation prediction model, the power system is subjected to steady-state operation analysis.

10. A method of steady state operation analysis of an electric power system as claimed in claim 9, characterized in that, The process of fusing the target mechanism steady-state model and the data-driven model based on the residual value, a preset residual compensation mechanism, and fusion weights to obtain a steady-state operation prediction model, and performing steady-state operation analysis of the power system based on the steady-state operation prediction model, includes: The residual value is compared with a preset residual threshold. When the residual value is less than or equal to the preset residual threshold, the target mechanism steady-state model and the data-driven model are fused based on the fusion weight to form an initial fusion model, and the initial fusion model is used as the steady-state operation prediction model. When the residual value is greater than the residual threshold value, fusing the target mechanism steady-state model and the data-driven model based on the fusion weight to obtain an initial fusion model, and correcting the initial fusion model based on the residual value and a preset residual compensation coefficient to obtain a steady-state operation prediction model; Performing steady-state operation analysis on the power system based on the steady-state operation prediction model.

Images