WO2026007294A1 - 电力系统安全稳定快速评估方法、装置及终端 - Google Patents

电力系统安全稳定快速评估方法、装置及终端

Info

Publication number
WO2026007294A1
WO2026007294A1 PCT/CN2024/129722 CN2024129722W WO2026007294A1 WO 2026007294 A1 WO2026007294 A1 WO 2026007294A1 CN 2024129722 W CN2024129722 W CN 2024129722W WO 2026007294 A1 WO2026007294 A1 WO 2026007294A1
Authority
WO
WIPO (PCT)
Prior art keywords
power
pixelated
stability
power system
image
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2024/129722
Other languages
English (en)
French (fr)
Inventor
范士雄
赵泽宁
马士聪
王铁柱
郭剑波
罗魁
荆逸然
侯玮琳
徐浩田
郭雅蓉
单熙雯
李苏宁
李东琦
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Electric Power Research Institute Co Ltd CEPRI
Original Assignee
China Electric Power Research Institute Co Ltd CEPRI
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by China Electric Power Research Institute Co Ltd CEPRI filed Critical China Electric Power Research Institute Co Ltd CEPRI
Publication of WO2026007294A1 publication Critical patent/WO2026007294A1/zh
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—ELECTRIC POWER NETWORKS; CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for AC mains or AC distribution networks
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—ELECTRIC POWER NETWORKS; CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J2103/00—Details of circuit arrangements for mains or AC distribution networks
    • H02J2103/30—Simulating, planning, modelling, reliability check or computer assisted design [CAD] of electric power networks

Definitions

  • This application relates to the field of power grid operation and control technology, specifically to methods, devices and terminals for rapid assessment of power system safety and stability.
  • the power system is one of the most important, largest, and most complex man-made systems in modern society. With the increasing uncertainty of power sources and loads, the vulnerability and openness of the power grid, the continuous expansion of the power grid scale, the more complex and variable the power grid operation mode, the constant changes in the transient stability boundary of the power system, the increasingly complex transient stability mechanism of the power system, and the continuous aggravation of the risk of power grid security and stability mismatch.
  • Transient stability issues in power systems are a key factor threatening the safe operation of large-scale power grids, and in severe cases, can lead to power system collapse.
  • time-domain simulation is highly adaptable but has low timeliness; direct methods, due to their simplified physical models, tend to yield conservative calculation results.
  • this application proposes a method, device and terminal for rapid assessment of power system safety and stability, aiming to solve the problems of insufficient timeliness and accuracy of existing power system transient stability assessment methods.
  • this application provides a method for rapid assessment of the safety and stability of a power system, including:
  • the measured or simulated steady-state power characteristic variables are pixelated to generate offline three-channel pixelated images, and a sample set is constructed;
  • the electrical quantity types of the steady-state power characteristic variables include: voltage, active power, and reactive power;
  • a fusion model based on deep residual network and Transformer is trained.
  • the trained fusion model is used for rapid assessment of power system safety and stability.
  • the real-time measured steady-state characteristic variables of the power system are pixelated to generate an online three-channel pixelated image; the generated online three-channel pixelated image is processed by the trained fusion model to generate a transient stability assessment conclusion of the power system.
  • step of pixelating the measured or simulated steady-state power characteristic variables includes:
  • the power grid is pixelated using an escaping mechanism, and each power grid component is mapped to a planar graphic within the base image; the base image includes non-bus base images and bus base images.
  • the step of pixelating the measured or simulated steady-state power characteristic variables to generate an offline three-channel pixelated image includes:
  • the power steady-state characteristic variables are mapped to the non-bus foundation image to generate a pixelated image of active power and a pixelated image of reactive power.
  • the power steady-state characteristic variables are mapped to the bus foundation image to generate a pixelated image of voltage.
  • the active power pixelated image, the reactive power pixelated image, and the voltage pixelated image are structured into the offline three-channel pixelated image.
  • the construction of the sample set includes:
  • faults that may cause system instability are set for various typical power grid operation modes, and the power grid is simulated offline to obtain the state variables of the power grid during stable operation before the fault is injected.
  • the selected steady-state power characteristic variables are pixelated to generate multiple sets of offline three-channel pixelated images
  • Each group of three-channel pixelated images is combined with its corresponding transient stable label to form a sample set.
  • the step of training a fusion model based on deep residual networks and Transformers using the constructed sample set includes:
  • the three-channel pixelated image corresponding to each sample in the sample set is used as the input of the deep residual network.
  • the matrix feature map generated by the deep residual network is used as the input of the Transformer based on the multi-head attention mechanism.
  • the transient stable label corresponding to each sample in the sample set is used as the output of the Transformer based on the multi-head attention mechanism.
  • the deep residual network includes n residual modules, where n is a natural number not less than 3;
  • Deep residual networks process three-channel pixelated images, including:
  • the acquired three-channel pixelated image is used as input for standard convolution processing to obtain the standard convolution result;
  • Max pooling is performed on the standard convolution result to obtain the max pooling result
  • the max pooling result is used as input and sequentially passed through N 1 cascaded first residual modules to generate the processing result of the first residual module group;
  • the processing result of the i-th residual module group is used as input and passed through N (i+1) cascaded (i+1)-th residual modules to generate the processing result of the (i+1)-th residual module group, where i is greater than or equal to 1 and less than n;
  • the processing result of the generated nth residual module group is used as the matrix feature map output after processing by the deep residual network.
  • the Transformer based on the multi-head attention mechanism includes multiple Transformer encoder modules, a position encoding module, and a classification information module;
  • the Transformer encoder module includes a multi-head attention layer, a feedforward network layer, a residual connection unit, and a layer normalization unit.
  • the training is based on a fusion model of deep residual networks and Transformers, including:
  • Evaluation metrics based on the confusion matrix including accuracy, precision, recall, and F1 score, were used to assess the performance of the fusion model.
  • this application provides a power system safety and stability rapid assessment device, comprising:
  • a sample set construction unit is used to pixelate the measured or simulated steady-state characteristic variables of the power system, generate offline three-channel pixelated images, and construct a sample set;
  • the electrical quantity types of the steady-state characteristic variables of the power system include: voltage, active power, and reactive power;
  • the fusion model training unit is used to train a fusion model based on deep residual network and Transformer using the constructed sample set.
  • the trained fusion model is used for rapid assessment of power system safety and stability.
  • the transient stability assessment unit is used to pixelate the real-time measured power steady-state characteristic variables to generate an online three-channel pixelated image; the generated online three-channel pixelated image is processed by the trained fusion model to generate a power system transient stability assessment conclusion.
  • this application provides a terminal, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect.
  • Figure 1 is a flowchart illustrating the rapid assessment method for power system security and stability according to an embodiment of this application
  • Figure 2 is a flowchart illustrating a rapid assessment method for power system security and stability according to another embodiment of this application;
  • Figure 3A is a schematic diagram of the power grid topology of the IEEE 39-node standard system in an embodiment of this application;
  • Figure 3B is a visual representation of the pixelated three-channel superimposed IEEE 39-node standard system shown in Figure 3A;
  • Figure 3C is a visual representation of the pixelated voltage image of the IEEE 39-node standard system shown in Figure 3A.
  • Figure 3D is a visual representation of the pixelated active power image of the IEEE 39-node standard system shown in Figure 3A.
  • Figure 3E is a visual representation of the pixelated reactive power image of the IEEE 39-node standard system shown in Figure 3A.
  • Figure 4A is a visual representation and a partial magnified view of the voltage pixelation image in the embodiment of this application;
  • Figure 4B is a visual representation and a partial magnified view of the active power pixelated image in the active power pixelation of the embodiment of this application.
  • Figure 4C is a visual representation and a partial magnified view of the reactive power pixelated image in the embodiment of this application.
  • Figure 5 is an overall structural diagram of the fusion model used for transient stability assessment in the power system security and stability rapid assessment method of this application embodiment
  • Figure 6 is a schematic diagram of the ResNet network structure of the fusion model used for transient stability assessment in the power system security and stability rapid assessment method of this application embodiment;
  • Figure 7 is a schematic diagram of the Transformer of the fusion model used for transient stability assessment in the power system security and stability rapid assessment method of this application embodiment
  • Figure 8 is a schematic diagram of the composition of the power system safety and stability rapid assessment device according to an embodiment of this application.
  • Figure 9 is a schematic diagram of the terminal composition of the power system safety and stability rapid assessment method according to an embodiment of this application.
  • time-domain simulation is highly adaptable but lacks timeliness when performing transient stability assessments; direct methods tend to yield conservative results due to the simplification of the physical model.
  • Traditional "pre-plan" safety strategies are ill-suited to complex and ever-changing operating conditions, and data-driven artificial intelligence technology holds promise for providing new technical means for assessing the safety and stability of power systems.
  • this application provides a rapid assessment technology solution for power system safety and stability, which is an engineering application of deep learning under the power grid situational awareness technology. It is conducive to deepening and improving the transient stability assessment technology of power system and has broad application prospects. By upgrading the smart grid control system of power management departments at all levels or within each administrative region, it can bring huge economic and social benefits.
  • WAMS wide-area measurement systems
  • PMUs phasor measurement units
  • the rapid power system safety and stability assessment technology solution provided in this application includes offline and online stages.
  • structured power data is acquired, an escaping mechanism is constructed, and a deep residual network and a Transformer network are fused to establish a computer image-based fusion model for power system safety and stability assessment.
  • real-time power grid operating parameters are acquired from the power grid control system. These parameters are then escaped to generate multi-channel pixelated images, including voltage, active power, and reactive power pixelated images.
  • the trained computer image-based fusion model processes these multi-channel pixelated images to generate a power system safety and stability assessment conclusion, such as system stability or system instability.
  • the above-mentioned AI-based rapid assessment method for safety and stability deeply mines the inherent mapping patterns in massive, diverse, and low-value-density power grid operation data to characterize the transient stability boundary of the power system, thereby assessing the transient stability state of the power grid.
  • This method is not limited by the accuracy of physical models, and the perception of the power grid situation can be achieved within milliseconds. It can meet the balance between timeliness and accuracy requirements of online safety assessment, and can be effectively integrated with the power grid control system, providing an effective technical means to improve the intelligence level of power grid situation awareness and reduce the workload of control personnel.
  • This method combines artificial intelligence technology with the need for transient stability assessment, and solves key problems in applying artificial intelligence technology in power grid situational awareness tasks, including: how to convert power system operation data into structured data suitable for artificial intelligence technology processing, and what kind of artificial intelligence model and its network structure to establish to solve the problems of uncertainty, vulnerability and openness in actual power grid application scenarios.
  • the power system safety and stability rapid assessment method of this application includes the following steps:
  • S100 Pixelate the measured or simulated steady-state characteristic variables of the power system to generate an offline three-channel pixelated image and construct a sample set;
  • the electrical quantity types of the steady-state characteristic variables of the power system include: voltage, active power, and reactive power;
  • S200 Using the constructed sample set, train a fusion model based on deep residual network and Transformer.
  • the trained fusion model is used for rapid assessment of power system safety and stability.
  • S300 Pixelate the real-time measured steady-state characteristic variables of the power system to generate an online three-channel pixelated image; process the generated online three-channel pixelated image using the trained fusion model to generate a power system transient stability assessment conclusion.
  • step S200 deep residual networks (ResNet) and networks based on multi-head attention mechanisms are used.
  • the Transformers are fused to obtain a fusion model for rapid assessment of power system security and stability (hereinafter referred to as an AI-based fusion model).
  • the fusion model can extract features from multi-channel pixelated images and establish a correlation mapping between multi-channel pixelated images and the transient stability state of the power grid, thereby enabling rapid assessment of the transient stability of the power system.
  • the overall technical framework of the power system safety and stability rapid assessment method is shown in Figure 2.
  • step S100 involves pixelating the measured or simulated steady-state characteristic variables of the power system to generate an offline three-channel pixelated image and constructing a sample set, including the following steps S110A, S110B, S120, and S130.
  • S110A Determine the topology and operation mode of the power grid, set typical faults, perform offline time-domain simulation of the power grid, and analyze its power system transient stability.
  • the transient stability of a power system is related to various factors such as the power grid's operating mode and the type of fault.
  • Clearly defining the power grid's operating mode is fundamental to power system transient stability analysis.
  • the actual operating range of a power grid is generally within the neighborhood of a typical power grid operating mode, and cannot fully cover all future power grid operating scenarios.
  • the probability of system instability after a power grid fault is low, resulting in very few system instability samples obtained through wide-area measurement systems, leading to a lack of sufficient sample size. Therefore, constructing a data-driven fusion model (the aforementioned AI-based fusion model) requires building a sample set that balances the amount of data on system stability samples and system instability samples under diverse power grid operating scenarios as input to train the model.
  • a pre-built time-domain simulation model of the power grid can be used.
  • faults that may lead to system instability such as single-phase ground faults, two-phase short circuits, and three-phase short circuits; typically, the fault type or severity with the greatest destructive impact under the current typical operating mode can be selected
  • This generates and acquires the state variables of the power grid during stable operation before the fault injection, and determines and records whether the power system is in an unstable or stable state after the fault injection as the conclusion of the power system's transient stability assessment. Furthermore, whether the power system is in an unstable or stable state after the fault injection is used as the transient stability label for the sample.
  • the electrical quantity types of steady-state characteristic variables of power include voltage, active power, and reactive power
  • the node types or component types for setting steady-state characteristic variables of power include: bus, line, generator, and load
  • the steady-state characteristic variables of power include: voltage of each bus; active power and reactive power of each line; active power and reactive power of each generator; active power and reactive power of each load.
  • data preprocessing includes converting data to per-unit values and supplementing missing data through interpolation. Data preprocessing is performed separately for each set of steady-state characteristic variables obtained under various typical power grid operating modes, and will not be elaborated further.
  • the steady-state characteristic variables of typical power grid operation modes and the transient stability labels of system instability or system stability after the injection of faults are combined into a sample and added to the sample set.
  • the ratio of the number of samples labeled as system-stable to the number of samples labeled as system-instable is adjusted based on engineering experience. This balances the amount of data corresponding to system-stable labels with the amount of data corresponding to system-instable labels, which helps to ensure the accuracy and generalization ability of the data-driven fusion model.
  • the key state variables that affect the transient stability of the power system are selected as power steady-state characteristic variables and data preprocessed to obtain each sample, thus constructing a sample set in which the amount of data corresponding to the system stability label is balanced with the amount of data corresponding to the system instability label.
  • a planar image refers to a planar graphic with M ⁇ N pixels at a specified resolution; it can be a grayscale image or a color image.
  • a grayscale image is a single-channel image, with the pixel value ranging from 0 to 255.
  • a color image is a three-channel image.
  • a color image can use the RGB color model, CMY and CMYK color models, or the HSI color model. When using the RGB color model, the pixel value in the R, G, and B channels ranges from 0 to 255.
  • a color image using any color model when displayed on a screen, can possess color in the human visual sense.
  • the steady-state power characteristic variables obtained under various typical power grid operation modes in the sample set are pixelated and then used as input to the fusion model.
  • power steady-state characteristic variables related to the transient stability of the power system are selected, and after pixelation, they are input into the fusion model to train the fusion model.
  • the power grid is pixelated using an escaping mechanism, and each power grid component is mapped to a planar graphic within the base image.
  • the base image includes non-bus base images and bus base images.
  • an escaping mechanism is used to generate non-bus foundation images (as shown in Figure 3D or Figure 3E) and bus foundation images (as shown in Figure 3C) corresponding to the power grid topology diagram.
  • the power grid topology diagram is a planar image with M ⁇ N pixels at a specified resolution.
  • the non-busbar foundation image and the busbar foundation image correspond to the power grid topology diagram, including: the resolution of the non-busbar foundation image and the busbar foundation image is directly proportional to the resolution of the power grid topology diagram; the number of pixels in the width and height directions of the non-busbar foundation image and the busbar foundation image are directly proportional to the number of pixels in the width and height directions of the power grid topology diagram, respectively.
  • the non-busbar foundation image and the busbar foundation image can be grayscale images or color images.
  • the power grid is pixelated using an escaping mechanism according to the power grid topology, including:
  • each generator is mapped to a first shape, such as a rectangle with m1 pixels.
  • the pixel value of each pixel in the rectangle is a pre-set first value or a random value, specifically, it can be zero.
  • each load is mapped to a second graphic in the non-busbar foundation image, such as a rectangle with m2 pixels.
  • the pixel value of each pixel in the rectangle is a pre-set second value or a random value, specifically, it can be zero.
  • each line is mapped to a third shape in the non-busbar base image, such as a polygon with m3 pixels.
  • the pixel value of each pixel in the polygon is a pre-set third value or a random value, specifically, it can be zero.
  • each bus is mapped to a fourth graphic in the bus base image, such as a rectangle with m4 pixels.
  • the pixel value of each pixel in the rectangle is a pre-set fourth value or a random value, specifically, it can be zero.
  • variables m1, m2, m3, and m4 which represent the shape or size of planar graphics of different elements, can be adaptively adjusted according to the size of the pixelated image and customization requirements to achieve a better visual display effect.
  • the aforementioned pre-set first, second, third, and fourth values are all different.
  • the generator, load, and line each have different pixel values, resulting in better visual differentiation and display effects.
  • busbar generators, loads, and lines each have different pixel values, providing good visual differentiation and display effects.
  • comparing the power grid topology diagram with the comprehensive foundation image helps verify the accuracy of the escaping mechanism.
  • selecting m2 of the second graphic to be equal to the value of m1 of the first graphic simplifies the operation and improves the processing speed without affecting the display effect and visual effect.
  • the positions of buses, lines, generators, and loads in the power grid topology diagram can be expressed using coordinates. For example, with the top left corner of the power grid topology diagram as the origin, the left and right directions as the width direction, and the up and down directions as the height direction, the absolute positions of buses, lines, generators, and loads in the power grid topology diagram can be expressed using height and width coordinates. Alternatively, a certain element in the power grid topology diagram can be used as a reference point, and the relative positions of buses, lines, generators, and loads in the power grid topology diagram can be expressed based on the relative positions of other elements with respect to the reference point.
  • the location of each component within the power grid topology and the connection relationships between components are extracted; or a power grid topology diagram is obtained, and the location of each component within the power grid topology diagram and the relative positions between components are extracted;
  • the component type is any of the following: bus, line, generator, load; a power grid escape coordinate system is established, and the escape coordinates of each component are generated, in a flat... The components are distributed within the surface graphic.
  • the power grid coverage area and various components within the power grid are displayed using a pre-defined planar graphic (with a background of pre-defined pixel values, such as the blue color in the RGB color model in Figure 3C, Figure 3D, or Figure 3E).
  • a pre-defined planar graphic such as a rectangle, rhombus, trapezoid, or parallelogram.
  • Figure 3C shows the planar graphics corresponding to each of the M1 buses.
  • the pixel value of each planar graphic corresponds to the voltage value of the bus.
  • Each bus is displayed as a horizontal line segment with a line width of A.
  • the length of the line segment corresponds to the length of the bus in the IEEE 39-node standard diagram.
  • Figure 3D shows the planar graphics corresponding to each of the M2 lines.
  • the pixel value of each planar graphic corresponds to the active power value of the corresponding line.
  • Each line is displayed as a combination of vertical line segments and/or horizontal line segments with a line width of B.
  • the length of the line segment corresponds to the line length.
  • Figure 3E shows the planar graphics corresponding to each of the M2 lines.
  • the pixel value of each planar graphic corresponds to the reactive power value of the corresponding line.
  • Figure 3D shows the planar graphics corresponding to the M3 generators respectively.
  • the pixel value of each planar graphic corresponds to the active power value of the generator (generated).
  • the generator is represented by a square with a side length of C1.
  • the coverage area of the planar graphics corresponding to the generator is determined according to the rated power of the generator or the side length C1.
  • Figure 3E shows the planar graphics corresponding to the M3 generators respectively.
  • the pixel value of each planar graphic corresponds to the reactive power value of the generator (generated).
  • Figure 3D shows the planar graphs corresponding to M4 loads.
  • the pixel values of each planar graph correspond to the active power value of the corresponding load (consumption).
  • the load is represented by a square with a side length of C2.
  • the coverage area of the planar graph corresponding to the load is determined according to the rated power of the load or the side length C2.
  • Figure 3E shows the planar graphs corresponding to M4 loads.
  • the pixel values of each planar graph correspond to the reactive power value of the corresponding load (consumption).
  • each generator and each load is connected to one end of a line.
  • S130 Map the power steady-state characteristic variables to the non-bus foundation image to generate a pixelated image of active power and a pixelated image of reactive power; map the power steady-state characteristic variables to the bus foundation image to generate a pixelated image of voltage.
  • mapping the power steady-state characteristic variables to the non-bus base image to generate a pixelated active power image includes:
  • the active power values of each line, each generator, and each load in the obtained power steady-state characteristic variables are mapped to the pixel values of the corresponding planar graphics of each element.
  • the pixel values of each pixel in the planar graphics corresponding to each line, each generator, and each load are updated respectively to obtain the active power pixelated image.
  • the pixel value of each pixel point of each graphic corresponding to the generator or load corresponds to the active power value of each line, the active power value of each generator, and the active power value of each load in the steady-state characteristic variables of the power.
  • mapping the power steady-state characteristic variables to the non-bus foundation image to generate a pixelated reactive power image includes:
  • the reactive power values of each line, each generator, and each load in the obtained power steady-state characteristic variables are mapped to the pixel values of the corresponding planar graphics of each element.
  • the pixel values of each pixel in the planar graphics corresponding to each line, each generator, and each load are updated respectively to obtain a pixelated image of reactive power.
  • the pixel value of each pixel point of each graphic corresponding to the generator or load corresponds to the reactive power value of each line, the reactive power value of each generator, and the reactive power value of each load in the steady-state characteristic variables of the power.
  • mapping the power steady-state characteristic variables to the bus base image to generate a voltage pixelated image includes:
  • the voltage values of each bus in the acquired steady-state power characteristic variables are mapped to the pixel values of the corresponding planar graphics of each bus.
  • the pixel values of each pixel in the corresponding planar graphic of each bus are updated to obtain the voltage pixelated image.
  • the pixel value of each pixel point of each pattern corresponding to each bus corresponds to the voltage value of each bus in the steady-state characteristic variables of the power.
  • the pixel values of each pixel in the voltage pixelated image are converted into corresponding visual colors, resulting in the visual display image shown in Figure 3C.
  • the pixel values of each pixel in the active power pixelated image are converted into corresponding visual colors, resulting in the visual display image shown in Figure 3D.
  • the pixel values of each pixel in the reactive power pixelated image are converted into corresponding visual colors, resulting in the visual display image shown in Figure 3E.
  • each power grid component is mapped to a planar graphic within the base image.
  • Each electrical quantity within the steady-state characteristic variables is mapped to the pixel value of the corresponding planar graphic according to component type, and the resulting pixelated images of reactive power, active power, and voltage are updated.
  • each steady-state characteristic variable is pixelated into three different channels of computer images: a pixelated image of reactive power, a pixelated image of active power, and a pixelated image of voltage.
  • the AI-based fusion model can process computer images quickly and efficiently and extract their features. Therefore, pixelating the steady-state power characteristic variables is beneficial for maximizing the performance of the fusion model.
  • steady-state feature variable pixelation includes: mapping the voltage value in the steady-state feature variable corresponding to each element to the pixel value corresponding to the planar image, and updating the voltage pixelated image according to each pixel value;
  • the active power values in the steady-state characteristic variables corresponding to each component are mapped to the pixel values corresponding to the planar image, and the active power pixelated image is updated according to each pixel value.
  • the reactive power values in the steady-state characteristic variables corresponding to each component are mapped to the pixel values corresponding to the planar image, and the reactive power pixelated image is updated according to each pixel value.
  • the voltage pixelated image records the planar patterns corresponding to M1 buses.
  • Each planar pattern corresponding to a bus has a center coordinate and a coverage area.
  • the pixel value of each planar pattern corresponding to a bus is obtained by mapping the voltage value of each bus.
  • the active power pixelated image records the planar graphics corresponding to M2 lines, each planar graphic corresponding to a line having a central coordinate and a coverage area, and the pixel value of each planar graphic corresponding to a line is mapped from the active power value of each line;
  • the active power pixelated image records the planar graphics corresponding to M3 generators, each planar graphic corresponding to a generator having a central coordinate and a coverage area, and the pixel value of each planar graphic corresponding to a generator is mapped from the active power value of each generator;
  • the active power pixelated image records the planar graphics corresponding to M4 loads, each planar graphic corresponding to a load having a central coordinate and a coverage area, and the pixel value of each planar graphic corresponding to a load is mapped from the active power value of each load.
  • the reactive power pixelated image records the planar graphics corresponding to M2 lines, each planar graphic corresponding to a line having a central coordinate and a coverage area, and the pixel value of each planar graphic corresponding to a line is mapped from the reactive power value of each line;
  • the reactive power pixelated image records the planar graphics corresponding to M3 generators, each planar graphic corresponding to a generator having a central coordinate and a coverage area, and the pixel value of each planar graphic corresponding to a generator is mapped from the reactive power value of each generator;
  • the reactive power pixelated image records the planar graphics corresponding to M4 loads, each planar graphic corresponding to a load having a central coordinate and a coverage area, and the pixel value of each planar graphic corresponding to a load is mapped from the reactive power value of each load.
  • the pixelated visualization images of the IEEE 39-node standard system are shown in Figures 3C, 3D, and 3E.
  • the voltage pixelated image shown in Figure 3C multiple bus nodes or bus elements are displayed, and the pixel value of each pixel in each bus node or bus element corresponds to the voltage value carried by each bus.
  • the active power pixelated image shown in Figure 3D multiple line, generator, and load nodes are displayed, and the pixel value of each pixel in each node corresponds to the active power value carried by each line, generator, or load bus.
  • the reactive power pixelated image shown in Figure 3E multiple line, generator, and load nodes are displayed, and the pixel value of each pixel in each node corresponds to the reactive power value carried by each line, generator, or load bus.
  • FIG 4A the portion within the white dashed rectangle in the middle right area of the left-hand voltage pixelated image (containing four pixel values) is magnified and displayed as a 17 ⁇ 17 matrix on the right.
  • This matrix corresponds to 17 ⁇ 17 pixels, and the pixel value of each pixel corresponds to the voltage value carried by the bus node or bus element it belongs to.
  • Pixel value 0 corresponds to the background color;
  • pixel values 0.8, 0.6, and 0.7 correspond to the other three colors, respectively.
  • each set of three pixelated images including the active power pixelated image, the reactive power pixelated image, and the voltage pixelated image, contains information on the magnitude of steady-state characteristic variables and power grid topology information, and can reflect the spatial distribution of system voltage, active power, and reactive power.
  • each component is arranged in the bus base image and non-bus base image according to its relative position in the power grid topology.
  • the steady-state characteristic variables (voltage, active power, and reactive power) are converted into pixel values for each pixel.
  • the power grid topology information and the magnitude of the steady-state characteristic variables can be recorded in three pixelated images within a set, representing the shape and pixel values of each component.
  • the magnitude of the steady-state characteristic variables changes, the corresponding pixelated image will also change accordingly.
  • each group of three pixelated images can be represented by a high-dimensional array (h, w, c, v).
  • Each pixel can be represented by an element of this high-dimensional array, and each element has the form (h, w, c, v).
  • w represents the horizontal coordinate of the pixel (as mentioned above in the width direction)
  • h represents the vertical coordinate of the pixel (as mentioned above in the height direction)
  • c represents the channel where the pixel is located (e.g., c is 0, 1, or 2, as mentioned above).
  • the voltage pixelated image corresponds to 0, the active power pixelated image corresponds to 1, and the reactive power pixelated image corresponds to 2);
  • v represents the pixel value of the pixel.
  • the steady-state characteristic variables obtained under various typical power grid operating modes within the sample set are pixelated to generate three pixelated images within a set: active power pixelated image, reactive power pixelated image, and voltage pixelated image.
  • Each set of three pixelated images is used as a three-channel pixelated image, and a high-dimensional array is used for data structuring.
  • the resulting high-dimensional array can be used for subsequent construction of fusion models, training of fusion models, and prediction of power system stability using the trained fusion model.
  • the pixelation of the power grid topology and steady-state power characteristic variables to generate the input for the fusion model is one of the foundations of the overall technical framework.
  • the purpose of pixelation is to convert the power grid topology or steady-state power characteristic variables into graphical data. This is because graphical data is beneficial for leveraging the excellent image processing performance of the fusion model and its learning algorithm.
  • the high-dimensional array or three-channel pixelated image corresponding to each sample in the sample set is used as the input of the fusion model, and the transient stable label corresponding to each sample in the sample set is used as the output of the fusion model, so as to train the fusion model with supervised learning.
  • Rectified Linear Unit is a commonly used activation function in neural networks. In general, it refers to the ramp function in mathematics.
  • step S200 uses the constructed sample set to train a fusion model based on deep residual network and Transformer, including the following steps S210 and S220.
  • step S210 ResNet and Transformer are fused to obtain the framework of the fused model, including:
  • the sample set is divided into a training set and a test set, such as the training set accounting for 80% of the sample set and the test set accounting for 20% of the sample set.
  • the high-dimensional array or three-channel pixelated image corresponding to each sample in the sample set is used as the input of the deep residual network.
  • the matrix feature map generated by the deep residual network is used as the input of the Transformer based on the multi-head attention mechanism.
  • the transient stable label corresponding to each sample in the sample set is used as the output of the Transformer, that is, the output of the fusion model.
  • step S220 the fusion model is trained using the sample set constructed in the aforementioned steps to establish a correlation mapping between multi-channel pixelated images and the transient stability state of the power grid.
  • the constructed fusion model establishes a mapping relationship between power steady-state feature variables and power grid transient stability.
  • the output process matrix feature map is obtained.
  • the process matrix feature map has a reduced image resolution and an increased number of convolution kernel channels.
  • the generated matrix feature map or the output process matrix feature map is 8 ⁇ 8 ⁇ 1024.
  • the process matrix feature map output by ResNet is input into the Transformer encoder to calculate the final classification result, which includes transient stability labels such as system stability and system instability.
  • the identity mapping structure in each residual module of ResNet includes an identity mapping across short connections. These are illustrated in the solid or dashed boxes in Figure 6, showing the identity mapping path from the input matrix feature map to the matrix feature map generated by the residual path.
  • This identity mapping structure outputs the input matrix feature map with an identity and adds it to the matrix feature map from the residual path. After processing using the ReLU activation function, the processed matrix feature map is generated.
  • the deep residual network ResNet50 by adding identity mappings across short connections, can transform the complex function fitting problem into a problem of solving the residual function, ensuring that the gradient always exists during backpropagation. This effectively avoids the training difficulties or gradient vanishing problems caused by excessively deep network layers.
  • the three dashed boxes respectively illustrate the first residual module, the downsampling second residual module (e.g., a convolution with a stride of 2), and the downsampling third residual module (e.g., a convolution with a stride of 2).
  • the identity mapping path is implemented through a convolutional layer.
  • the input matrix feature map is output through the identity mapping path with the convolutional layer and added to the matrix feature map output by the residual path, thereby halving the number of pixels in the matrix feature map and achieving downsampling.
  • x is the input parameter or the input matrix feature map
  • y is the parameter or matrix feature map output by the residual structure
  • W, W1 and W2 are the network parameters whose values need to be determined by learning when training a neural network in a supervised manner.
  • the three-channel pixelated images are voltage pixelated images, active power pixelated images, and reactive power pixelated images, respectively.
  • the pixelated images shown in Figures 3C, 3D, and 3E are illustrated as having 128 ⁇ 128 pixels.
  • the power grid topology diagram shown in Figure 3A displays 10 generators, 46 lines, and 39 buses.
  • the matrix feature map is processed by a deep residual network, including:
  • the input matrix feature map is processed by standard convolution (Stdconv,7 ⁇ 7,64), with a kernel size of 7 ⁇ 7, the output channels are adjusted to 64, and the stride is 2 to obtain the standard convolution result;
  • Max pooling (3 ⁇ 3) is performed on the standard convolution result with a stride of 2 to obtain the max pooling result;
  • the max pooling result is used as input and passes through N 1 cascaded first residual modules in sequence to generate the processing result of the first residual module group;
  • the processing result of the first residual module group is taken as input and sequentially passed through N 2 cascaded second residual modules to generate the processing result of the second residual module group;
  • the processing result of the second residual module group is used as input, and then passed through N 3 cascaded third residual modules to generate the processing result of the third residual module group, which is used as the matrix feature map output after processing by the deep residual network.
  • the input matrix feature map is processed through the first residual module, including:
  • the ReLU activation function is used to process the first convolutional layer result to obtain the first activation result
  • a second convolution (e.g., Conv2d, 3 ⁇ 3, 64) on the first activation result with a stride of 1.
  • the kernel size is 3 ⁇ 3 and the number of channels is 64 to generate the second convolution result.
  • the ReLU activation function is used to process the results of the second convolution layer to obtain the second activation result
  • a third convolution (e.g., Conv2d, 1 ⁇ 1, 256) is performed on the second activation result with a stride of 1.
  • the kernel size is 1 ⁇ 1 and the number of channels is 256, generating the third convolution result.
  • an identity mapping convolution (e.g., Conv2d, 1 ⁇ 1, 256) on the input matrix feature map with a stride of 1, the convolution kernel size is 1 ⁇ 1 and the number of channels is 256, and generate the identity mapping convolution result;
  • the ReLU activation function is applied to generate the processing result of the first residual module group, which serves as the matrix feature map corresponding to the input matrix feature map.
  • the image resolution is reduced while the number of convolution kernel channels is increased.
  • the 1 ⁇ 1 convolutional kernel is used to conveniently adjust the number of channels in the intermediate layers.
  • the number of channels is reduced (256->64), and after passing through the 3 ⁇ 3 convolutional layer, the number of channels is restored (64->256), which can significantly reduce the number of network parameters.
  • the input matrix feature map is processed through the second residual module, including:
  • the ReLU activation function is used to process the first convolutional layer result to obtain the first activation result
  • a second convolution (e.g., Conv2d, 3 ⁇ 3, 128) is performed on the first activation result with a stride of 2.
  • the kernel size is 3 ⁇ 3 and the number of channels is 128, generating the second convolution result.
  • the ReLU activation function is used to process the results of the second convolution layer to obtain the second activation result
  • a third convolution layer (e.g., Conv2d, 1 ⁇ 1, 512) is performed on the second activation result with a stride of 1.
  • the kernel size is 1 ⁇ 1 and the number of channels is 512, generating the third convolution result.
  • an identity mapping convolution (e.g., Conv2d, 1 ⁇ 1, 512) on the input matrix feature map with a stride of 2, with a kernel size of 1 ⁇ 1 and 512 channels, to generate the identity mapping convolution result;
  • the ReLU activation function is applied to generate the processing result of the second residual module group, which serves as the matrix feature map corresponding to the input matrix feature map.
  • the data dimensionality is reduced through the processing of the second residual module.
  • the 1 ⁇ 1 convolutional kernel is used to conveniently adjust the number of channels in the intermediate layers.
  • the number of channels is reduced (512->128), and after passing through the 3 ⁇ 3 convolutional layer, the number of channels is restored (128->512), which can significantly reduce the number of network parameters.
  • the third residual module includes:
  • the ReLU activation function is used to process the first convolutional layer result to obtain the first activation result
  • a second convolution (e.g., Conv2d, 3 ⁇ 3, 256) is performed on the first activation result with a stride of 2.
  • the kernel size is 3 ⁇ 3 and the number of channels is 256, generating the second convolution result.
  • the ReLU activation function is used to process the results of the second convolution layer to obtain the second activation result
  • a third convolution layer (e.g., Conv2d, 1 ⁇ 1, 1024) is performed on the second activation result with a stride of 1.
  • the kernel size is 1 ⁇ 1 and the number of channels is 1024, generating the third convolution result.
  • an identity mapping convolution (e.g., Conv2d, 1 ⁇ 1, 1024) is performed on the input matrix feature map with a stride of 2.
  • the convolution kernel size is 1 ⁇ 1 and the number of channels is 1024, generating the identity mapping convolution result.
  • the ReLU activation function is applied to generate the third residual module group processing result, which serves as the matrix feature map corresponding to the input matrix feature map.
  • the image resolution is reduced while the number of convolution kernel channels is increased.
  • the 1 ⁇ 1 convolutional kernel is used to conveniently adjust the number of channels in the intermediate layers.
  • the number of channels is reduced (1024->256), and after passing through the 3 ⁇ 3 convolutional layer, the number of channels is restored (256->1024), which can significantly reduce the number of network parameters.
  • the number of channels in each convolutional layer in the second residual module is twice the number of channels in each convolutional layer in the first residual module; the number of channels in each convolutional layer in the third residual module is twice the number of channels in each convolutional layer in the second residual module.
  • a deep residual network ResNet can include more than three residual modules.
  • the residual paths and identity mapping paths of other residual modules can be implemented with reference to the first residual module, the second residual module, or the third residual module shown in Figure 6, and will not be described in detail here.
  • the max pooling result is used as input and sequentially passed through N+ 1 cascaded first residual modules to generate the processing result of the first residual module group, including:
  • the max pooling result or the processing result of the previous first residual module group is used as input, and it passes through N 1 cascaded first residual modules in sequence.
  • the processing result of the N 1 cascaded first residual module is used as the processing result of the first residual module group.
  • the max pooling result is taken as input and passed through (1+2) first residual modules in sequence to generate the processing result of the first residual module group, which will not be described in detail here.
  • the processing result of the first residual module or the processing result of the previous cascaded second residual module is used as input, and it passes through N2 cascaded second residual modules in sequence, and the processing result of the N2nd cascaded second residual module is used as the processing result of the second residual module group.
  • the processing result of the first residual module group is taken as input and passed through (1+3) cascaded second residual modules to generate the processing result of the second residual module group, which will not be described in detail here.
  • the processing result of the second residual module group is used as input, and sequentially passed through N+ 3 third residual modules to generate the processing result of the third residual module group, including:
  • the processing result of the second residual module group or the processing result of the previous third residual module is used as input, and it passes through N 3 third residual modules in sequence.
  • the processing result of the N 3 third residual module is used as the processing result of the third residual module group.
  • the processing result of the second residual module group is taken as input and passed through (1+8) third residual modules in sequence to generate the processing result of the third residual module group, which will not be described in detail here.
  • ResNet performs feature selection on the input pixelated image, selecting feature variables that are highly correlated with the transient stability of the power grid to form the optimal feature set, and outputs it in the form of a multi-dimensional feature map (e.g., 8 ⁇ 8 ⁇ 1024).
  • the Transformer includes multiple Transformer encoder modules, a position embedding module, and a classification information module.
  • the position embedding module helps the model identify positional relationships in the feature vectors; the classification information module provides 1 ⁇ 192 classification relationships within the feature vectors.
  • the classification information is a learnable parameter used to store data related to classification.
  • the Transformer encoder module includes a multi-head attention layer, a feedforward network layer, residual connection units, and layer normalization units, with the multi-head attention mechanism at its core.
  • This mechanism utilizes adaptive attention weight distribution to divide the model into multiple subspaces, thereby focusing on different aspects of power system transient stability features (or transient stability information).
  • transient stability information integrates the transient stability information from various aspects, giving it a global receptive field. This allows for efficient computation of attention among various transient stability features, assigning high weights to high-value transient stability information, and achieving efficient allocation of information processing resources.
  • the adaptive attention weight distribution Attention(Q,K,V) reflects the correlation between the Transformer encoder output data and the input matrix feature map (as shown in Figure 5 or Figure 7, the matrix feature map has 8 ⁇ 8 ⁇ 1024 dimensions after processing by ResNet50) and the value of the features.
  • the query matrix Q, the key matrix K, and the value matrix V are generated by transforming the feature vector matrix X, respectively.
  • WQ , WK , and WV are all linear transformation matrices; d is a scaling factor and is the length of the row vectors in the key matrix K.
  • Total number (the total number of subspaces is a pre-defined model parameter);
  • H ⁇ sub>i ⁇ /sub> represents the single-head attention value of the i-th subspace
  • MultiHead(Q,K,V) integrates the metastable information from various aspects
  • W ⁇ sub>O ⁇ /sub> is the transformation matrix used to connect the attention values of all subspaces.
  • each layer of neurons shares the mean and variance, normalizing the layer input to a standard normal distribution, which helps to enhance the flow of information during backpropagation.
  • each computer image is processed using residual connection units and layer normalization units after passing through the multi-head attention layer and feedforward network layer.
  • the residual connection unit is an identity mapping type, which can make gradient propagation more efficient and avoid model degradation.
  • training a fusion model based on ResNet and Transformer includes:
  • the sample set is input into the constructed ResNet and Transformer-based fusion model for training until the model converges.
  • the cross-entropy loss function used for classification is adopted, and the gradient descent method is used for optimization.
  • the optimization method performs a global minimization of the loss function until the model converges.
  • evaluation metrics are used to test and evaluate the effectiveness of the fusion model for power system transient stability assessment, which can guide the online application of the model or model optimization and updates.
  • Accuracy refers to the proportion of correctly predicted unstable and stable samples out of the total number of samples in the training or test set, used to indicate overall accuracy.
  • Precision refers to the proportion of correctly predicted unstable samples out of the total number of predicted unstable samples.
  • Recall refers to the proportion of correctly predicted unstable samples out of the total number of unstable samples.
  • the F1 score is the harmonic mean of precision and recall for unstable samples, used to focus on scenarios involving unstable samples.
  • TP represents a correctly judged stable sample (predicting a stable sample as a stable system);
  • TN represents a correctly judged unstable sample (predicting an unstable sample as an unstable system);
  • FP represents an incorrectly predicted or misjudged unstable sample;
  • FN represents an incorrectly predicted or misjudged stable sample.
  • the evaluation of the fusion model based on ResNet and Transformer includes:
  • the offline-trained fusion model is tested on the test set. Once the evaluation metric value of the model on the test set (generated using the aforementioned evaluation metrics) meets the preset threshold, the trained fusion model is put into online practical application.
  • step S300 the real-time measured power steady-state characteristic variables are pixelated to generate an online three-channel pixelated image; the generated online three-channel pixelated image is processed using a trained fusion model to generate a power system transient stability assessment conclusion.
  • the steady-state characteristic variables of the power system obtained from the wide-area measurement system are converted into pixelated images and input into the trained fusion model to evaluate the transient stability of the power system, generating a transient stability evaluation conclusion for the power system.
  • the power system transient stability assessment conclusions generated by the fusion model are sent to the power grid control system.
  • this invention combines artificial intelligence technology with the needs of power system transient stability assessment, proposes a power system transient stability assessment method based on pixelation and fusion model of power steady-state feature variables, clarifies the escaping mechanism of power steady-state feature variables to computer images, establishes a fusion model, and proposes an overall technical framework of data-escaping-image-model.
  • This transient stability assessment method for power systems provides practical guidance and assistance on how artificial intelligence technology can be integrated with power system situational awareness technology.
  • the proposed pixelation method for power steady-state characteristic variables is conducive to giving full play to the advantages of artificial intelligence technology in processing image data. Combined with the powerful data mining capabilities of the ResNet and Transformer fusion model, it meets the timeliness and accuracy requirements of online security assessment, giving full play to the role of machine intelligence. It provides an effective means to improve the intelligence level of power grid regulation and reduce the workload of regulation personnel, and has high engineering applicability.
  • the power system safety and stability rapid assessment device includes:
  • the sample set construction unit 10 is used to pixelate the measured or simulated steady-state characteristic variables of the power, generate offline three-channel pixelated images, and construct a sample set;
  • the electrical quantity types of the steady-state characteristic variables of the power include: voltage, active power, and reactive power;
  • the fusion model training unit 20 is used to train a fusion model based on a deep residual network and a Transformer using the constructed sample set.
  • the fusion model once trained, is used for rapid assessment of the safety and stability of power systems.
  • the transient stability assessment unit 30 is used to pixelate the real-time measured power steady-state characteristic variables to generate an online three-channel pixelated image; the generated online three-channel pixelated image is processed by the trained fusion model to generate a power system transient stability assessment conclusion.
  • the power system safety and stability rapid assessment device and the power system safety and stability rapid assessment method of the present application embodiments are based on the same inventive concept, can be implemented with reference, and have the same beneficial effects as the methods they adopt, operate or implement.
  • FIG9 shows a schematic diagram of a terminal provided by some embodiments of this application.
  • the terminal 8 includes: a processor 800, a memory 801, a bus 802, and a communication interface 803.
  • the processor 800, the communication interface 803, and the memory 801 are connected via the bus 802.
  • the memory 801 stores a computer program that can run on the processor 800.
  • the processor 800 runs the computer program, it executes the method provided by any embodiment of this application.
  • the memory 801 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this device network element and at least one other network element is achieved through at least one communication interface 803 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.
  • RAM random access memory
  • non-volatile memory such as at least one disk storage device.
  • Communication between this device network element and at least one other network element is achieved through at least one communication interface 803 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.
  • Bus 802 can be an ISA bus, PCI bus, or EISA bus, etc.
  • the bus can be divided into an address bus, a data bus, a control bus, etc.
  • the memory 801 is used to store programs. After receiving an execution instruction, the processor 800 executes the program. The method disclosed in any embodiment of this application can be applied to the processor 800, or implemented by the processor 800.
  • the processor 800 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the method can be completed by the integrated logic circuitry in the hardware of the processor 800 or by instructions in software form.
  • the processor 800 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application.
  • the general-purpose processor can be a microprocessor or any conventional processor.
  • the steps of the method disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor.
  • the software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art.
  • the storage medium is located in memory 801.
  • Processor 800 reads the information in memory 801 and, in conjunction with its hardware, completes the steps of the method.
  • the terminal provided in this application embodiment and the method in this application embodiment are based on the same inventive concept and have the same beneficial effects as the methods they adopt, operate or implement.
  • the computer-readable storage medium is an optical disc, on which a computer program (i.e., a program product) is stored.
  • a computer program i.e., a program product
  • the computer program When the computer program is run by a processor, it executes the method provided in any of the embodiments.
  • examples of the computer-readable storage medium may also include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical and magnetic storage media, which will not be elaborated here.
  • PRAM phase-change memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • RAM random access memory
  • ROM read-only memory
  • EEPROM electrically erasable programmable read-only memory
  • flash memory or other optical and magnetic storage media, which will not be elaborated here.
  • the computer-readable storage medium provided in the embodiments of this application is based on the same inventive concept as the method in the embodiments of this application, and has the same beneficial effects as the method adopted, run or implemented by the application stored therein.
  • the rapid power system safety and stability assessment technology includes offline and online stages.
  • the offline stage structured power data is acquired, an escaping mechanism is constructed, and a deep residual network and a Transformer network are fused to establish a computer image-based fusion model, thereby achieving power system safety and stability assessment.
  • real-time power grid operating parameters are acquired from the power grid control system. These parameters are then escaped to generate multi-channel pixelated images, including voltage, active power, and reactive power pixelated images.
  • the trained computer image-based fusion model is used to process these multi-channel pixelated images, generating a power system safety and stability assessment conclusion, such as system stability or system instability.
  • this AI-based rapid safety and stability assessment method by deeply mining the inherent mapping patterns in massive, diverse, and low-value-density power grid operating data, characterizes the transient stability boundary of the power system, achieving assessment of power grid transient stability (transient state). Assessment of stable states.
  • This method is not limited by the accuracy of physical models, and the perception of the power grid situation can be achieved within milliseconds. It can meet the balance between timeliness and accuracy requirements of online security assessment, and can be effectively integrated with the power grid control system, providing an effective technical means to improve the intelligence level of power grid situation perception and reduce the workload of control personnel.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Power Engineering (AREA)
  • Supply And Distribution Of Alternating Current (AREA)
  • Testing, Inspecting, Measuring Of Stereoscopic Televisions And Televisions (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本申请公开电力系统安全稳定快速评估方法、装置及终端。该方法包括：将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，并构建样本集；利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，训练完成的所述融合模型用于电力系统安全稳定快速评估；将实时测量的电力稳态特征变量像素化，生成在线的三通道像素化图像；利用训练完成的所述融合模型处理生成的所述在线的三通道像素化图像，生成电力系统暂态稳定状态评估结论。如此，结合ResNet和Transformer的融合模型能够满足电力系统在线安全稳定评估的时效性和准确性需求。

Description

电力系统安全稳定快速评估方法、装置及终端
相关申请的交叉引用
本申请要求在2024年07月02日提交中国专利局、申请号为202410873998X、发明名称为“电力系统安全稳定快速评估方法、装置及终端”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及电网运行与控制技术领域，具体涉及电力系统安全稳定快速评估方法、装置及终端。
背景技术
电力系统是现代社会最重要、最庞大、最复杂的人造系统之一。伴随着源荷不确定性、电网脆弱性和开放性进一步加强，电网规模不断扩大，电网运行方式更加复杂多变，电力系统暂态稳定边界不断发生变化，电力系统暂态稳定机理愈发复杂，电网安全稳定失配风险持续加剧。
电力系统的暂态稳定问题是威胁大规模电网安全运行的关键因素，严重时会导致电力系统崩溃。目前，暂态稳定评估的方法中，时域仿真法适应性强，但时效性低；直接法由于简化物理模型而导致计算结果偏保守。
因此亟需研究能够快速进行电力系统暂态稳定评估的技术方案。
发明内容
鉴于此，本申请提出电力系统安全稳定快速评估方法、装置及终端，旨在解决现有电力系统暂态稳定评估方法的时效性和准确性不足的问题。
第一方面，本申请提供一种电力系统安全稳定快速评估方法，包括：
将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，并构建样本集；所述电力稳态特征变量的电气量类型包括：电压、有功功率、无功功率；
利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，训练完成的所述融合模型用于电力系统安全稳定快速评估；
将实时测量的电力稳态特征变量像素化，生成在线的三通道像素化图像；利用训练完成的所述融合模型处理生成的所述在线的三通道像素化图像，生成电力系统暂态稳定状态评估结论。
进一步地，所述将测量获取的或仿真生成的电力稳态特征变量像素化，包括：
根据电网拓扑结构，利用转义机制，将电网进行像素化处理，将电网各元件映射为基础图像内的各平面图形；所述基础图像包括非母线基础图像及母线基础图像。
进一步地，所述将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，包括：
将所述电力稳态特征变量映射到所述非母线基础图像，生成有功功率像素化图像及无功功率像素化图像，将所述电力稳态特征变量映射到所述母线基础图像，生成电压像素化图像；
将所述有功功率像素化图像、所述无功功率像素化图像及所述电压像素化图像结构化为所述离线的三通道像素化图像。
进一步地，所述构建样本集，包括：
利用预先构建的电网时域仿真模型，对多种典型电网运行方式设置可能导致系统失稳的故障，对电网进行离线仿真，获取注入故障之前电网稳定运行时的状态量；
从多种典型电网运行方式下获取的注入故障之前的电网稳定运行时的状态量中选取多个用于表征系统暂态稳定性的电气量作为电力稳态特征变量；
将选取的所述电力稳态特征变量像素化，生成多组离线的三通道像素化图像；
将各组三通道像素化图像与其对应的暂态稳定标签组合为各样本，形成样本集。
进一步地，所述利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，包括：
将样本集内各样本对应的三通道像素化图像作为深度残差网络的输入，将深度残差网络生成的矩阵特征图作为基于多头注意力机制的Transformer的输入，将样本集内各样本对应的暂态稳定标签作为基于多头注意力机制的Transformer的输出。
进一步地，所述深度残差网络包括n个残差模块，其中，n为不小于3的自然数；
深度残差网络处理三通道像素化图像，包括：
将获取的三通道像素化图像作为输入做标准卷积处理，得到标准卷积结果；
对所述标准卷积结果做最大池化，得到最大池化结果；
在残差路径上，将所述最大池化结果作为输入，依次经过N1个串联的第一残差模块，生成第一残差模块组处理结果；
将第i残差模块组处理结果作为输入，依次经过N(i+1)个串联的第(i+1)残差模块，生成第(i+1)残差模块组处理结果，其中，i大于或等于1且小于n；
将生成的第n残差模块组处理结果，作为经深度残差网络处理后输出的矩阵特征图。
进一步地，所述基于多头注意力机制的Transformer包括多个Transformer编码器模块、位置编码模块、分类信息模块；
所述Transformer编码器模块包括多头注意力层、前馈网络层、残差连接单元和层归一化单元。
进一步地，所述训练基于深度残差网络与Transformer的融合模型，包括：
选用基于混淆矩阵的准确率、精确率、召回率以及F1值构建评估指标，以评估融合模型的性能。
第二方面，本申请提供一种电力系统安全稳定快速评估装置，包括：
构建样本集单元，用于将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，并构建样本集；所述电力稳态特征变量的电气量类型包括：电压、有功功率、无功功率；
融合模型训练单元，用于利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，训练完成的所述融合模型用于电力系统安全稳定快速评估；
暂态稳定评估单元，用于将实时测量的电力稳态特征变量像素化，生成在线的三通道像素化图像；利用训练完成的所述融合模型处理生成的所述在线的三通道像素化图像，生成电力系统暂态稳定状态评估结论。
第三方面，本申请提供一种终端，包括：存储器、处理器及存储在该存储器上并可在该处理器上运行的计算机程序，所述处理器执行所述计算机程序以实现第一方面说明的所述方法。
本申请的附加方面和优点将在下面的描述中部分给出，部分将从下面的描述中变得明显，或通过本申请的实践了解到。
附图说明
通过阅读下文优选实施方式的详细描述，各种其他的优点和益处对于本领域普通技术人员将变得清楚明了。附图仅用于示出优选实施方式的目的，而并不认为是对本申请的限值。而且在整个附图中，用相同的参考符号表示相同的部件。在附图中：
图1为本申请实施例的电力系统安全稳定快速评估方法的流程示意图；
图2为本申请另一实施例的电力系统安全稳定快速评估方法的流程示意图；
图3A为本申请实施例中IEEE 39节点标准系统的电网拓扑结构示意图；
图3B为图3A所示的IEEE 39节点标准系统像素化后三通道叠加后的视觉展示图；
图3C为图3A所示的IEEE 39节点标准系统像素化后的电压像素化图像的视觉展示图；
图3D为图3A所示的IEEE 39节点标准系统像素化后的有功功率像素化图像的视觉展示图；
图3E为图3A所示的IEEE 39节点标准系统像素化后的无功功率像素化图像的视觉展示图；
图4A为本申请实施例中电压像素化中电压像素化图像的视觉展示图及局部放大图；
图4B为本申请实施例中有功功率像素化中有功功率像素化图像的视觉展示图及局部放大图；
图4C为本申请实施例中无功功率像素化中无功功率像素化图像的视觉展示图及局部放大图；
图5为本申请实施例的电力系统安全稳定快速评估方法中用于暂态稳定评估的融合模型的整体结构图；
图6为本申请实施例的电力系统安全稳定快速评估方法中用于暂态稳定评估的融合模型的ResNet网络结构示意图；
图7为本申请实施例的电力系统安全稳定快速评估方法中用于暂态稳定评估的融合模型的Transformer示意图；
图8为本申请实施例的电力系统安全稳定快速评估装置的组成示意图；
图9为应用本申请实施例的电力系统安全稳定快速评估方法的终端的组成示意图。
具体实施方式
下面将参照附图更详细地描述本公开的示例性实施例。虽然附图中显示了本公开的示例性实施例，然而应当理解，可以以各种形式实现本公开而不应被这里阐述的实施例所限值。相反，提供这些实施例是为了能够更透彻地理解本公开，并且能够将本公开的范围完整的传达给本领域的技术人员。需要说明的是，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。
新能源大规模并网和高比例电力电子设备的投运导致电力系统惯性降低，运行状态不确定性增加、可预估性降低、暂态稳定的内在机理发生深刻变化。
受人力限制和时间成本的影响，传统“离线仿真、人工决策”的预案式分析无法适应复杂多样的电网运行环境，在面对复杂多样的大电网时越来越力不从心。相应地，传统的“预案式”安全策略也难以适应电力系统复杂多变的运行方式，在解决暂态稳定评估时具有计算效率低和时效性差等缺陷。
目前，人工智能技术得益于庞大的数据量、成熟的学习算法以及强大的算力支撑，在视觉、语言和交通等领域取得技术突破，有效提高了不同工业领域的智能化水平。
参考前述说明，完成暂态稳定评估任务时，时域仿真法适应性强，但时效性低；直接法由于简化物理模型导致计算结果偏保守。传统的“预案式”安全策略难以适应复杂多变的运行方式，数据驱动的人工智能技术有望为电力系统的安全稳定评估提供新的技术手段。
未来电网发展的主要特征为数字化、信息化和智能化，亟需先进的人工智能技术来增强电力系统态势感知任务的时效性。基于数据驱动的人工智能技术具有较强的感知预测和决策能力，将人工智能技术与电网调控系统进行有效的结合，可以为提高电网调控智能化水平和减轻调控人员工作强度提供有效手段。
为此，本申请提供电力系统安全稳定快速评估技术方案，是深度学习在电网态势感知技术下的工程性应用，有利于深化与提升电力系统暂态稳定评估技术，应用前景广阔，通过升级各级电力管理部门或各本行政区域内的智能电网调控系统，可带来巨大的经济效益和社会效益。
目前，基于相量测量单元(Phasor Measurement Unit，PMU)的广域量测系统(Wide-Area Measurement System，WAMS)日趋完善，高速实时的通信网络使得电力系统的可观性大为提高，为电网的数智化建设提供了充分的数据支撑。
本申请提供的电力系统安全稳定快速评估技术方案，包括离线环节及在线环节。在离线环节，获取电力结构化数据，构建转义机制，融合深度残差网络与Transformer网络，建立基于计算机图像的融合模型，实现电力系统安全稳定评估。在线环节实时从电网调控系统获取电网运行参数，将获取的电网运行参数转义，生成电压像素化图像、有功功率像素化图像及无功功率像素化图像等多通道像素化图像；利用训练完成的基于计算机图像的融合模型处理生成的多通道像素化图像，生成电力系统安全稳定评估结论，如系统稳定或系统失稳。如此，在暂态稳定评估任务中应用人工智能技术具有快速准确的优点。
具体地，以上基于人工智能的安全稳定快速评估方法，通过在海量、多型和低价值密度的电网运行数据中深度挖掘内在映射规律，刻画电力系统的暂稳边界，实现对电网暂稳(暂态稳定)状态的评估。此方法不受物理模型精度限制，对电网态势的感知可以在毫秒内实现，能够满足在线安全评估在时效性和准确性上的需求平衡，能够与电网调控系统进行有效的结合，为提高电网态势感知智能化水平和减轻调控人员工作强度提供有效技术手段。
在各级电力管理部门运行的电网调控系统中实施本申请提供的电力系统安全稳定快速评估的技术方案，能够提高电网态势感知智能化水平，减轻调控人员工作强度。
此方法结合人工智能技术和暂态稳定评估的需求，解决了在电网态势感知任务中应用人工智能技术面临的关键问题，包括：如何将电力系统运行数据转换为适合人工智能技术处理的结构化数据，以及建立何种人工智能模型及其网络结构来解决电网实际应用场景中存在的不确定性、脆弱性和开放性等难题。
如图1所示，本申请实施例的电力系统安全稳定快速评估方法，包括以下步骤：
S100：将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，并构建样本集；所述电力稳态特征变量的电气量类型包括：电压、有功功率、无功功率；
S200：利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，训练完成的所述融合模型用于电力系统安全稳定快速评估；
S300：将实时测量的电力稳态特征变量像素化，生成在线的三通道像素化图像；利用训练完成的融合模型处理生成的所述在线的三通道像素化图像，生成电力系统暂态稳定状态评估结论。
以上步骤S200中，对深度残差网络(Residual Network，ResNet)及基于多头注意力机制的 Transformer进行融合，得到用于电力系统安全稳定快速评估的融合模型(以下又称基于人工智能的融合模型)。融合模型能够对多通道像素化图像进行特征提取，建立起多通道像素化图像与电网暂态稳定状态的关联映射，进而可以快速对电力系统的暂态稳定性进行评估。
在一些实施例中，电力系统安全稳定快速评估方法的总体技术框架如图2所示。
在一些实施例中，步骤S100中将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，并构建样本集，包括以下步骤S110A、S110B、S120及S130。
S110A：确定电网的拓扑模型、运行方式，设置典型故障，针对电网进行离线时域仿真，分析其电力系统暂态稳定性。
电力系统暂态稳定性与电网运行方式以及故障类型等多种因素有关，明确电网运行方式是电力系统暂态稳定性分析的基础。电网实际运行区间一般处于典型电网运行方式的邻域内，无法完整覆盖未来电网的所有运行场景。另一方面，在实际运行中，电网遭受故障后发生系统失稳的概率较低，通过广域量测系统测量获取的系统失稳样本非常少，样本缺乏，样本量不足。而构建基于数据驱动的融合模型(前述基于人工智能的融合模型)，则需要构建在多样化电网运行场景下系统稳定样本的数据量与系统失稳样本的数据量相平衡的样本集作为模型的输入来训练模型。
为此，除外测量获取电力稳态特征变量，还可以利用预先构建的电网时域仿真模型，对任一典型电网运行方式，设置可能导致系统失稳的故障(如单相接地故障、两相短路、三相短路，通常可以设置当前典型运行方式下破坏程度最大的故障类型或故障程度)，对电网进行离线仿真，生成并获取注入故障之前电网稳定运行时的各状态量，判断并记录注入故障之后电力系统为失稳状态或稳定状态作为电力系统暂态稳定状态评估结论。进一步地，将注入故障之后电力系统为失稳状态或稳定状态作为样本的暂态稳定标签。
如此，在多样化的典型电网运行场景中，分别设置故障，针对电网进行离线时域仿真，在电网实际运行区间内分析电力系统的暂态稳定性。
通常，通过调整电网时域仿真模型的系统参数，实现以下任一项的多样化的典型电网运行场景：系统内发电机的出力发生变化；系统内负荷水平发生变化。
S110B：根据在多样化的典型电网运行方式下分析电力系统暂态稳定性的过程及结论，从测量获取的或仿真生成的状态量中选取多个用于表征系统暂态稳定性的电气量作为电力稳态特征变量。
表1电力稳态特征变量
如表1所示，电力稳态特征变量的电气量类型包括电压、有功功率及无功功率；设置电力稳态特征变量的节点类型或元件类型包括：母线、线路、发电机及负荷；电力稳态特征变量包括：各母线电压；各线路有功功率及线路无功功率；各发电机有功功率及各发电机无功功率；各负荷有功功率及各负荷无功功率。
根据在多样化的典型电网运行方式下分析电力系统暂态稳定性获取的各组稳态特征变量，进行数据预处理，以后续对其进行像素化处理。
具体地，数据预处理包括转换为标幺值、针对缺失的数据通过插值进行补充。数据预处理针对各典型电网运行方式下获取的各组稳态特征变量分别进行，不再赘述。
进一步地，数据预处理之后，将典型电网运行方式时的各稳态特征变量、注入故障之后系统失稳或系统稳定的暂态稳定标签组合为一个样本补充到样本集中。
进一步地，在构建样本集时，根据工程经验调整样本集中样本标签为系统稳定的样本量与样本标签为系统失稳的样本量的比例关系，使得系统稳定标签对应的数据量与系统失稳标签对应的数据量相平衡，有利于保证基于数据驱动的融合模型的准确性和泛化能力。
以上，针对仿真生成的典型运行场景下的各状态量，从中选择影响电力系统暂态稳定的关键的状态量作为电力稳态特征变量并进行数据预处理，得到各样本，构建起系统稳定标签对应的数据量与系统失稳标签对应的数据量相平衡的样本集。
平面图像是指在指定分辨率下具有M×N个像素点的平面图形；可以是灰度图像或彩色图像。灰度图像时，为单通道图像，图像内各像素点的像素值的取值范围为0至255。彩色图像时，图像为三通道图像。彩色图像可以是RGB彩色模型、CMY和CMYK彩色模型或HSI彩色模型。采用RGB彩色模型时，在R通道、G通道及B通道，各像素点的像素值的取值范围为0至255。任一彩色模型的彩色图像展示在显示屏幕上时，可以具有人类视觉意义上的彩色。
如图2、图5所示，将样本集内各典型电网运行方式下获取的电力稳态特征变量进行像素化处理后，作为融合模型的输入。
如此，选择与电力系统暂态稳定有关的电力稳态特征变量，并对其像素化处理后输入到融合模型中，对融合模型进行训练。
应该理解为，基于数据驱动的融合模型的输入的通道数或维度数或数据量，在模型训练、模型性能评估及在线暂态稳定评估(也即在线应用)这3个阶段中保持一致。
S120：根据电网拓扑结构，利用转义机制，将电网进行像素化处理，将电网各元件映射为基础图像内的各平面图形，所述基础图像包括非母线基础图像及母线基础图像。
具体地，基于图3A所示的电网拓扑结构图，利用转义机制，生成与电网拓扑结构图对应的非母线基础图像(如图3D或图3E)及母线基础图像(如图3C)。
如图3A所示，电网拓扑结构图为在指定分辨率下具有M×N个像素点的平面图像。具体地，非母线基础图像及母线基础图像与电网拓扑结构图对应，包括：非母线基础图像及母线基础图像的分辨率与电网拓扑结构图的分辨率成正比例，非母线基础图像及母线基础图像在宽度方向及高度方向上分别具有的像素点数与电网拓扑结构图在宽度方向及高度方向上分别具有的像素点数成正比例。自然地，非母线基础图像及母线基础图像可以是灰度图像或彩色图像。
在一些实施例中，根据电网拓扑结构，利用转义机制，将电网进行像素化处理，包括：
基于图3A所示的电网拓扑结构图，确定母线、线路、发电机和负荷分别在电网拓扑结构图中的位置；
根据确定的发电机在电网拓扑结构图中的位置，在非母线基础图像中，将各发电机映射为第一图形，如具有m1个像素点的矩形，矩形内各像素点的像素值为预先设定的第一数值或随机值，具体地，可以取零值；
根据确定的负荷在电网拓扑结构图中的位置，在非母线基础图像中，将各负荷映射为第二图形，如具有m2个像素点的矩形，矩形内各像素点的像素值为预先设定的第二数值或随机值，具体地，可以取零值；
根据确定的线路在电网拓扑结构图中的位置，在非母线基础图像中，将各线路映射为第三图形，如具有m3个像素点的多边形，多边形内各像素点的像素值为预先设定的第三数值或随机值，具体地，可以取零值；
根据确定的母线在电网拓扑结构图中的位置，在母线基础图像中，将各母线映射为第四图形，如具有m4个像素点的矩形，矩形内各像素点的像素值为预先设定的第四数值或随机值，具体地，可以取零值。
应该理解为，所述转义机制中，代表不同元件的平面图形的形状或尺寸的变量m1、m2、m3、m4的数值可以根据像素化图像的尺寸以及定制化需求适应性地调整，以具有更好的视觉展示效果。
优选地，前述预先设定的第一数值、第二数值、第三数值、第四数值各不相同。如此，在非母线基础图像中，发电机、负荷及线路分别具有不同的像素值，具有更好的视觉区分度和视觉展示效果。
进一步地，将非母线基础图像与母线基础图像叠加后，得到整体基础图像。整体基础图像中，母线发电机、负荷及线路分别具有不同的像素值，具有良好的视觉区分度和视觉展示效果。以及，将电网拓扑结构图与整体基础图像对比，有利于核对转义机制的准确性。
优选地，选择所述第二图形的m2与所述第一图形m1的数值相等，则能够在不影响展示效果及视觉效果的前提下，简化操作，提高处理速度。
在一些实施例中，母线、线路、发电机和负荷分别在电网拓扑结构图中的位置可以用坐标来表达，如，以电网拓扑结构图的左上角为坐标原点，左右方向为宽度方向，上下方向为高度方向，以高度坐标、宽度坐标来表达母线、线路、发电机和负荷分别在电网拓扑结构图中的绝对位置。又或者，以电网拓扑结构图的某个元件为参照基准，根据其他元件与参照基准的相对位置来表达母线、线路、发电机和负荷分别在电网拓扑结构图中的相对位置。
在一些实施例中，提取电网拓扑结构内各元件在电网内的位置、各元件之间的连接关系；或获取电网拓扑结构图，提取各元件在电网拓扑结构图内的位置、各元件之间的相对位置；参考前述说明，元件类型为以下任一种：母线、线路、发电机、负荷；建立电网转义坐标系，生成各元件的转义坐标，在平 面图形内分布各元件。
具体地，以预设尺寸的平面图形(背景为预设像素值，如图3C、图3D或图3E中RGB彩色模型中的蓝色)展示电网覆盖范围及电网内各元件，其中，各元件或节点以预设的平面图形(如矩形、菱形、梯形或平行四边形等多边形)展示。
与图3A相对应，图3C中示出了M1个母线分别对应的平面图形，各平面图形的像素值对应该母线的电压值，各母线以线宽为A的横向线段展示，线段的长度与IEEE39节点标准图中母线的长度对应。
与图3A相对应，图3D中示出了M2个线路分别对应的平面图形，各平面图形的像素值对应该线路上的有功功率值，各线路以线宽为B的竖向线段和/或横向线段的组合展示，线段的长度与线路长度相对应；与图3A相对应，图3E中示出了M2个线路分别对应的平面图形，各平面图形的像素值对应该线路上的无功功率值。
与图3A相对应，图3D中示出了M3个发电机分别对应的平面图形，各平面图形的像素值对应该发电机(发出)的有功功率值，发电机以边长为C1的正方形展示，发电机分别对应的平面图形的覆盖范围根据发电机额定功率或边长C1确定；图3E中示出了M3个发电机分别对应的平面图形，各平面图形的像素值对应该发电机(发出)的无功功率值。
与如图3A相对应，如图3D中示出了M4个负荷对应的平面图形，各平面图形的像素值对应该负荷(消耗)的有功功率值，负荷以边长为C2的正方形展示，负荷对应的平面图形的覆盖范围根据负荷额定功率或边长C2确定；图3E中示出了M4个负荷对应的平面图形，各平面图形的像素值对应该负荷(消耗)的无功功率值。
如图3D及图3E所示，各发电机及各负荷分别连接在一线路的一端。
S130：将所述电力稳态特征变量映射到所述非母线基础图像，生成有功功率像素化图像及无功功率像素化图像，将所述电力稳态特征变量映射到所述母线基础图像，生成电压像素化图像。
具体地，将所述电力稳态特征变量映射到所述非母线基础图像，生成有功功率像素化图像，包括：
将获取的电力稳态特征变量中各线路的有功功率值、各发电机的有功功率值、各负荷的有功功率值分别映射为各元件对应的平面图形的像素值；
在所述非母线基础图像中，分别更新各线路、各发电机、各负荷对应的平面图形中各像素点的像素值，得到有功功率像素化图像。
如此，在有功功率像素化图像中，发电机或负荷对应的各图形的各像素点的像素值与电力稳态特征变量中各线路的有功功率值、各发电机的有功功率值、各负荷的有功功率值相对应。
具体地，将所述电力稳态特征变量映射到所述非母线基础图像，生成无功功率像素化图像，包括：
将获取的电力稳态特征变量中各线路的无功功率值、各发电机的无功功率值、各负荷的无功功率值分别映射为各元件对应的平面图形的像素值；
在所述非母线基础图像中，分别更新各线路、各发电机、各负荷对应的平面图形中各像素点的像素值，得到无功功率像素化图像。
如此，在无功功率像素化图像中，发电机或负荷对应的各图形的各像素点的像素值与电力稳态特征变量中各线路的无功功率值、各发电机的无功功率值、各负荷的无功功率值相对应。
具体地，将所述电力稳态特征变量映射到所述母线基础图像，生成电压像素化图像，包括：
将获取的电力稳态特征变量中各母线的电压值分别映射为各母线对应的平面图形的像素值；
在母线基础图像中，分别更新各母线对应的平面图形中各像素点的像素值，得到电压像素化图像。
如此，在电压像素化图像中，各母线对应的各图形的各像素点的像素值与电力稳态特征变量中各母线的电压值相对应。
进一步地，以预设的RGB彩色模型中的色彩方案，将电压像素化图像中各像素点的像素值转化为对应的视觉色彩，得到如图3C所示的视觉展示图。
进一步地，以预设的RGB彩色模型中的色彩方案，将有功功率像素化图像中各像素点的像素值转化为对应的视觉色彩，得到如图3D所示的视觉展示图。
进一步地，以预设的RGB彩色模型中的色彩方案，将无功功率像素化图像中各像素点的像素值转化为对应的视觉色彩，得到如图3E所示的视觉展示图。
以上，根据电网拓扑结构及转义机制，将电网各元件映射为基础图像内的各平面图形，将电力稳态特征变量内各电气量按照元件类型映射为对应平面图形的像素值，并更新得到无功功率像素化图像、有功功率像素化图像、电压像素化图像。如此，将各稳态特征变量像素化处理为无功功率像素化图像、有功功率像素化图像、电压像素化图像这三个不同通道的计算机图像。
参考前述说明，基于人工智能的融合模型能够快速、高效地处理计算机图像，提取计算机图像的特征，因此，将电力稳态特征变量像素化有利于发挥融合模型的性能。
在一些实施例中，稳态特征变量像素化，包括：将各元件对应的稳态特征变量中的电压值映射为平面图像对应的像素值，并根据各所述像素值更新电压像素化图像；
将各元件对应的稳态特征变量中的有功功率值映射为平面图像对应的像素值，并根据各所述像素值更新有功功率像素化图像；
将各元件对应的稳态特征变量中的无功功率值映射为平面图像对应的像素值，并根据各所述像素值更新无功功率像素化图像。
如此，电压像素化图像记载M1个母线对应的平面图形，各母线对应的平面图形具有一中心坐标及覆盖面积，各母线对应的平面图形的像素值根据各母线的电压值映射得到。
如此，有功功率像素化图像记载M2个线路分别对应的平面图形，各线路分别对应的平面图形具有一中心坐标及覆盖面积，各线路分别对应的平面图形的像素值根据各线路的有功功率值映射得到；有功功率像素化图像记载M3个发电机分别对应的平面图形，各发电机分别对应的平面图形具有一中心坐标及覆盖面积，各发电机分别对应的平面图形的像素值根据各发电机的有功功率值映射得到；有功功率像素化图像记载M4个负荷对应的平面图形，各负荷对应的平面图形具有一中心坐标及覆盖面积，各负荷对应的平面图形的像素值根据各负荷的有功功率值映射得到。
如此，无功功率像素化图像记载M2个线路分别对应的平面图形，各线路分别对应的平面图形具有一中心坐标及覆盖面积，各线路分别对应的平面图形的像素值根据各线路的无功功率值映射得到；无功功率像素化图像记载M3个发电机分别对应的平面图形，各发电机分别对应的平面图形具有一中心坐标及覆盖面积，各发电机分别对应的平面图形的像素值根据各发电机的无功功率值映射得到；无功功率像素化图像记载M4个负荷对应的平面图形，各负荷对应的平面图形具有一中心坐标及覆盖面积，各负荷对应的平面图形的像素值根据各负荷的无功功率值映射得到。
以图3A所示的IEEE 39节点标准系统为例，对IEEE 39节点标准系统像素化处理后的可视化图像如图3C、图3D及图3E所示。图3C中示出的电压像素化图像中，展示有多个母线节点或母线元件，各母线节点或母线元件中各像素点的像素值与各母线承载的电压值相对应。图3D中示出的有功功率像素化图像中，展示有多个线路、发电机和负荷节点，各节点中各像素点的像素值与各线路、发电机或负荷母线承载的有功功率值相对应。图3E中示出的无功功率像素化图像中，展示有多个线路、发电机和负荷节点，各节点中各像素点的像素值与各线路、发电机或负荷母线承载的无功功率值相对应。
以图4A、图4B、图4C所示的电压像素化图像、有功功率像素化图像、无功功率像素化图像为例，对转义机制及像素化处理进一步具体说明。
图4A中左侧电压像素化图像内右中区域的白色虚线矩形框内的部分(具有4种像素值)放大后，展示为右侧的17×17矩阵。该矩阵对应17×17个像素点，各像素点的像素值与所在母线节点或母线元件承载的电压值相对应。其中，像素值0与背景色相对应；像素值0.8、0.6及0.7分别对应其他3个颜色。
图4B中左侧电压像素化图像内右上区域的白色虚线矩形框内的部分(具有5种像素值)放大后，展示为右侧的17×17矩阵。该矩阵对应17×17个像素点，各像素点的像素值与所在线路、发电机或负荷节点承载的有功功率值相对应。其中，像素值0与背景色相对应；像素值0.9、0.7、0.6及0.4分别对应其他4个颜色。
图4C中左侧电压像素化图像内右上区域的白色虚线矩形框内的部分(具有5种像素值)放大后，展示为右侧的17×17矩阵。该矩阵对应17×17个像素点，各像素点的像素值与所在线路、发电机或负荷节点承载的无功功率值相对应。其中，像素值0与背景色相对应；像素值0.8、0.7、0.5及0.8分别对应其他4个颜色。
经过所述像素化处理，有功功率像素化图像、无功功率像素化图像及电压像素化图像等每一组三个像素化图像，包含稳态特征变量的大小信息以及电网拓扑结构信息，能够反映系统电压、有功和无功的空间分布。
如此，通过转义机制，各元件按照在电网拓扑结构中的相对位置分别排列在母线基础图像及非母线基础图像中；通过以上，电力稳态特征变量(电压、有功功率和无功功率)转换为各像素点的像素值，如此，能够将电网拓扑结构信息及电力稳态特征变量的大小信息以各像素化图像中各元件的形状及像素值记载在一组内的三个像素化图像中。自然地，电网稳态特征变量的大小发生变化后，其对应的像素化图像也会相应地发生变化。
在以结构化数据存储以上有功功率像素化图像、无功功率像素化图像及电压像素化图像时，每一组三个像素化图像可由一个高维数组(h，w，c，v)表示，任一个像素点可以由高维数组的一个元素来对应，各元素均具有(h，w，c，v)形式。其中，w代表着像素点的横坐标(如前述的宽度方向)，h代表着像素点的纵坐标(如前述的高度方向)，c代表着像素点所在的通道(如c为0、1或2，如前述 的电压像素化图像对应为0，有功功率像素化图像对应为1，无功功率像素化图像对应为2)；v代表着像素点的像素值。
如此，重复以上步骤，将样本集内各典型电网运行方式下获取的稳态特征变量进行像素化处理，分别生成有功功率像素化图像、无功功率像素化图像及电压像素化图像这样一组内的三个像素化图像。将每一组三个像素化图像作为三通道像素化图像，并采用高维数组进行数据结构化，得到的高维数组可以用于后续构建融合模型、训练融合模型及利用训练完成的融合模型预测电力系统的稳定性。
以上，将电网拓扑结构和电力稳态特征变量进行像素化处理，生成融合模型的输入，是整个总体技术框架的基础之一。像素化处理的目的是将电网拓扑结构或电力稳态特征变量转换为图像化数据。这是因为图像化数据有利于发挥融合模型及其学习算法优异的图像处理性能。
以下步骤中，将样本集内各样本对应的高维数组或三通道像素化图像作为融合模型的输入，将样本集内各样本对应的暂态稳定标签作为融合模型的输出，以有监督学习对融合模型进行训练。
线性整流函数，Rectified Linear Unit，简称ReLU，是神经网络中常用的激活函数。通常意义下，其指代数学中的斜坡函数。
在一些实施例中，步骤S200利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，包括以下步骤S210及S220。
具体地，步骤S210中，对ResNet和Transformer进行融合，得到融合模型的框架，包括：
将样本集划分为训练集和测试集，如训练集占样本集内样本数的80％，测试集占样本集内样本数的20％；
将样本集内各样本对应的高维数组或三通道像素化图像作为深度残差网络的输入，将深度残差网络生成的矩阵特征图作为基于多头注意力机制的Transformer的输入，将样本集内各样本对应的暂态稳定标签作为所述Transformer的输出，也即融合模型的输出。
具体地，步骤S220中，利用前述步骤构建的所述样本集，训练所述融合模型，建立起多通道像素化图像与电网暂态稳定状态的关联映射，由构建的融合模型建立电力稳态特征变量与电网暂态稳定性之间的映射关系。
如图5及图6所示，输入数据(如前述的三通道像素化图像或高维数组)经ResNet处理后，输出过程矩阵特征图，所述过程矩阵特征图与输入的三通道像素化图像或高维数组相比，图像的分辨率降低且卷积核通道数实现了升维。
以分别具有128×128×3像素点的三通道像素化图像为例，经ResNet50处理后，生成的矩阵特征图或输出的过程矩阵特征图为8×8×1024。
如图5及图7所示，将ResNet输出的过程矩阵特征图输入到Transformer编码器中计算得到最终的分类结果，所述分类结果包括系统稳定、系统失稳等暂态稳定标签。
如图6所示，ResNet中，各残差模块中的恒等映射结构包括跨层短连接的恒等映射，如图6中各实线框或虚线框内展示的自输入的矩阵特征图跨层短连接到残差路径生成的矩阵特征图的恒等映射路径。该跨层短连接的恒等映射结构将输入的矩阵特征图恒等输出，并与来自残差路径的矩阵特征图相加，利用ReLU激活函数处理后，生成残差模块处理后的矩阵特征图。
如此，深度残差网络ResNet50通过增加跨层短连接的恒等映射，能够将复杂的函数拟合问题转换为对残差函数的求解问题，使反向传播运算时梯度一直存在，能够有效避免因网络层数过深造成的训练困难或梯度消失问题。
如图6上部所示，自左向右，3个虚线框内分别展示了第一残差模块、降采样型的第二残差模块(如步幅为2的卷积)及降采样型的第三残差模块(如步幅为2的卷积)。降采样型的残差模块中，恒等映射路径通过一卷积层实现，输入的矩阵特征图经设置有卷积层的恒等映射路径输出，并与残差路径输出的矩阵特征图相加，使得矩阵特征图的像素点数量减半，实现降采样。
具体地，带有卷积层的恒等映射路径及残差连接通过以下公式计算：
y＝F(x,W)+x                              (1)
F(x,W)＝W2×ReLU(W1x)                           (2)
式中：x是输入参数或输入的矩阵特征图，y是残差结构输出的参数或矩阵特征图；W、W1和W2分别是通过有监督方式训练神经网络时，需要通过学习确定其数值的网络参数。
参考前述说明，三通道像素化图像分别为电压像素化图像、有功功率像素化图像和无功功率像素化图像。在一些实施例中，以图3C、图3D及图3E所示的各像素化图像具有128×128个像素点为示例进行说明。如图3A所示的电网拓扑结构图内，展示有10台发电机、46条线路及39条母线。
如图6所示，以三通道像素化图像作为数据或输入的矩阵特征图，经深度残差网络处理，包括：
对输入的矩阵特征图做标准卷积(Stdconv,7×7,64)处理，卷积核尺寸为7×7，调整输出通道为64，步幅为2，得到标准卷积结果；
对标准卷积结果做最大池化(maxpool 3×3)，步幅为2，得到最大池化结果；
在残差路径上，将最大池化结果作为输入，依次经过N1个串联的第一残差模块，生成第一残差模块组处理结果；
在残差路径上，将第一残差模块组处理结果作为输入，依次经过N2个串联的第二残差模块，生成第二残差模块组处理结果；
在残差路径上，将第二残差模块组处理结果作为输入，依次经过N3个串联的第三残差模块，生成第三残差模块组处理结果，作为经深度残差网络处理后输出的矩阵特征图。
具体地，如图6所示，将输入的矩阵特征图经过第一残差模块，包括：
以步幅(Stride)1对输入的矩阵特征图做第一层卷积(如Conv2d,1×1,64)，卷积核尺寸为1×1，通道数为64，生成第一层卷积结果；
利用ReLU激活函数处理第一层卷积结果，得到第一次激活结果；
以步幅(Stride)1对第一次激活结果做第二层卷积(如Conv2d,3×3,64)，卷积核尺寸为3×3，通道数为64，生成第二层卷积结果；
利用ReLU激活函数处理第二层卷积结果，得到第二次激活结果；
以步幅(Stride)1对第二次激活结果做第三层卷积(如Conv2d,1×1,256)，卷积核尺寸为1×1，通道数为256，生成第三层卷积结果；
以步幅(Stride)1对输入的矩阵特征图做恒等映射卷积(如Conv2d,1×1,256)，卷积核尺寸为1×1，通道数为256，生成恒等映射卷积结果；
将第三层卷积结果与恒等映射卷积结果合并后，利用ReLU激活函数处理，生成第一残差模块组处理结果，作为与所述输入的矩阵特征图对应的矩阵特征图。如此，经第一残差模块处理，图像的分辨率降低且卷积核通道数实现了升维。
以上第一残差模块中，1×1的卷积核用于方便地调整中间层的通道数。在进入3×3的卷积层之前，减少通道数(256->64)，经过该3×3卷积层后再恢复通道数(64->256)，可以显著减少网络的参数量。
具体地，如图6所示，将输入的矩阵特征图经过第二残差模块，包括：
以步幅(Stride)1对输入的矩阵特征图做第一层卷积(如Conv2d,1×1,128)，卷积核尺寸为1×1，通道数为128，生成第一层卷积结果；
利用ReLU激活函数处理第一层卷积结果，得到第一次激活结果；
以步幅(Stride)2对第一次激活结果做第二层卷积(如Conv2d,3×3,128)，卷积核尺寸为3×3，通道数为128，生成第二层卷积结果；
利用ReLU激活函数处理第二层卷积结果，得到第二次激活结果；
以步幅(Stride)1对第二次激活结果做第三层卷积(如Conv2d,1×1,512)，卷积核尺寸为1×1，通道数为512，生成第三层卷积结果；
以步幅(Stride)2对输入的矩阵特征图做恒等映射卷积(如Conv2d,1×1,512)，卷积核尺寸为1×1，通道数为512，生成恒等映射卷积结果；
将第三层卷积结果与恒等映射卷积结果合并后，利用ReLU激活函数处理，生成第二残差模块组处理结果，作为与所述输入的矩阵特征图对应的矩阵特征图。如此，经第二残差模块处理，进行了数据的降维。
以上第二残差模块中，1×1的卷积核用于方便地调整中间层的通道数。在进入3×3的卷积层之前，减少通道数(512->128)，经过该3×3卷积层后再恢复通道数(128->512)，可以显著减少网络的参数量。
具体地，如图6所示，第三残差模块包括：
以步幅(Stride)1对输入的矩阵特征图做第一层卷积(如Conv2d,1×1,256)，卷积核尺寸为1×1，通道数为256，生成第一层卷积结果；
利用ReLU激活函数处理第一层卷积结果，得到第一次激活结果；
以步幅(Stride)2对第一次激活结果做第二层卷积(如Conv2d,3×3,256)，卷积核尺寸为3×3，通道数为256，生成第二层卷积结果；
利用ReLU激活函数处理第二层卷积结果，得到第二次激活结果；
以步幅(Stride)1对第二次激活结果做第三层卷积(如Conv2d,1×1,1024)，卷积核尺寸为1×1，通道数为1024，生成第三层卷积结果；
在恒等映射路径上，以步幅(Stride)2对输入的矩阵特征图做恒等映射卷积(如Conv2d,1×1,1024)，卷积核尺寸为1×1，通道数为1024，生成恒等映射卷积结果；
将第三层卷积结果与恒等映射卷积结果合并后，利用ReLU激活函数处理，生成第三残差模块组处理结果，作为与所述输入的矩阵特征图对应的矩阵特征图。如此，所述输入的矩阵特征图经第三残差模块处理后，图像的分辨率降低且卷积核通道数实现了升维。
以上第三残差模块中，1×1的卷积核用于方便地调整中间层的通道数。在进入3×3的卷积层之前，减少通道数(1024->256)，经过该3×3卷积层后再恢复通道数(256->1024)，可以显著减少网络的参数量。
参考图6所示，第二残差模块内各卷积层的通道数为第一残差模块内各卷积层的通道数的2倍；第三残差模块内各卷积层的通道数为第二残差模块内各卷积层的通道数的2倍，有利于增强特征提取能力，提取和类别相关的更加高级的特征，增加模型的复杂度和表达能力。
自然地，深度残差网络ResNet可以包括大于3个的残差模块，其他残差模块的残差路径及恒等映射路径可以参考图6所示的第一残差模块、第二残差模块或第三残差模块实现，不再赘述。
进一步地，在残差路径上，将最大池化结果作为输入，依次经过N1个串联的第一残差模块，生成第一残差模块组处理结果，包括：
在残差路径上，将最大池化结果或前一个第一残差模块组处理结果作为输入，依次经过N1个串联的第一残差模块，并将第N1个串联的第一残差模块处理结果作为第一残差模块组处理结果。
如图6所示，将最大池化结果作为输入，依次经过(1+2)个第一残差模块，生成第一残差模块组处理结果，不再赘述。
进一步地，在残差路径上，将第一残差模块处理结果或前一个串联的第二残差模块处理结果作为输入，依次经过N2个串联的第二残差模块，并将第N2个串联的第二残差模块处理结果作为第二残差模块组处理结果。
如图6所示，将第一残差模块组处理结果作为输入，依次经过(1+3)个串联的第二残差模块，生成第二残差模块组处理结果，不再赘述。
进一步地，在残差路径上，将第二残差模块组处理结果作为输入，依次经过N3个第三残差模块，生成第三残差模块组处理结果，包括：
在残差路径上，将第二残差模块组处理结果或前一个第三残差模块处理结果作为输入，依次经过N3个第三残差模块，并将第N3个第三残差模块处理结果作为第三残差模块组处理结果。
如图6所示，将第二残差模块组处理结果作为输入，依次经过(1+8)个第三残差模块，生成第三残差模块组处理结果，不再赘述。
如此，ResNet对输入的像素化图像进行特征选择，能够选择与电网暂态稳定关联程度高的特征变量组成最优特征集，并以多维特征图(如8×8×1024)的形式输出。
针对电网暂态稳定评估的分类问题，如图7所示，Transformer包括多个Transformer编码器模块、位置编码模块(Position Embedding)、分类信息模块。其中，位置编码模块可以帮助模型识别特征向量中的位置关系；分类信息模块用于提供特征向量中1×192的分类关系。具体地，分类信息是一个可学习的参数，用于存放与分类相关的数据信息。
其中，如图5及图7所示，Transformer编码器模块包括多头(Multi-Head)注意力层、前馈网络层、残差连接单元和层归一化(Layer Norm)单元，其核心为多头注意力机制。多头注意力机制利用自适应注意力权重分布使模型分在多个子空间下，进而去关注不同方面的电力系统暂态稳定特征信息(或暂稳信息)，最后将各个方面暂稳信息综合起来，使其具备全局感受野，可以实现各个暂稳特征间注意力的有效计算，为暂态稳定高价值信息分配高权重，实现对信息处理资源的高效分配。
自适应注意力权重分布Attention(Q,K,V)体现了Transformer编码器输出数据与输入的矩阵特征图(如图5或图7中，经ResNet50处理后，矩阵特征图具有8×8×1024维度)间的相关性以及特征的价值度，计算公式如下：



MultiHead(Q,K,V)＝[H1,…,Hh]WO                        (6)
其中，查询矩阵Q、键矩阵K、价值度矩阵V分别由特征向量矩阵X变换生成，WQ、WK和WV均为线性变换矩阵；d为缩放因子，为键矩阵K中行向量的长度；
WQ i、WK i和WV i分别是将Q、K、V各矩阵投影到第i个子空间的转换矩阵，i＝1,2...h，h为子空间 总数(子空间总数为预先设定的模型参数)；
Hi为第i个子空间的单头注意力值；MultiHead(Q,K,V)将各个方面暂稳信息综合起来，WO为转换矩阵，用于连接所有子空间的注意力值。
层归一化(Layer Norm)单元中，每层神经元共享均值和方差，使层输入归一为标准正态分布，有利于增强反向传播信息的流动性。
如图5及图7所示，各计算机图像经过多头注意力层和前馈网络层后均会采用残差连接单元和层归一化(Layer Norm)单元进行处理。
如图5及图7所示，残差连接单元为恒等映射型，能够使梯度传递更为高效，避免模型退化。
具体地，训练基于ResNet和Transformer的融合模型，包括：
将所述样本集输入到构建的基于ResNet和Transformer的融合模型中进行训练，直至模型收敛。具体地，损失函数采用用于分类的交叉熵损失函数，优化方法采用梯度下降法。随着模型多次训练，优化方法会对损失函数进行全局最小化寻优，直至模型收敛。
训练基于ResNet和Transformer的融合模型时，利用评估指标来检验和评估融合模型用于电力系统暂态稳定评估的效果，用于指导模型在线应用或者模型优化更新。
为了评估融合模型的性能与有效性，选用基于混淆矩阵的准确率(accuracy)、精确率(precision)、召回率(recall)以及F1值构建评估指标。
准确率accuracy指预测正确的失稳样本与稳定样本的数量和占训练集或测试集内样本总数的比例，用于指示整体准确率。精确率precision指预测正确的失稳样本占所有预测失稳样本总数的比例。召回率recall指失稳样本预测正确(将失稳样本预测为失稳结论)的数量占失稳样本总数的比例。F1值指针对失稳样本的精确率和召回率的调和均值，用于重点关注失稳样本场景。
所述各评估指标的计算公式如下所示：



式中：TP表示判断正确的稳定样本(将稳定样本预测为系统稳定)；TN表示判断正确的失稳样本(将失稳样本预测为系统失稳)；FP表示预测错误或错判的失稳样本；FN表示预测错误或错判的稳定样本。
具体地，评估基于ResNet和Transformer的融合模型，包括：
基于建立的评估指标，将离线训练完成的融合模型在测试集上进行测试；当模型在测试集中的评价指标值(利用前述评估指标生成)满足预先设定的阈值后，将训练完成的融合模型进行在线的实际应用。
在一些实施例中，步骤S300中，将实时测量的电力稳态特征变量像素化，生成在线的三通道像素化图像；利用训练完成的融合模型处理生成的所述在线的三通道像素化图像，生成电力系统暂态稳定状态评估结论。
具体地，融合模型在线应用过程中，将从广域量测系统测量获取的电力稳态特征变量，转义为像素化图像，并输入到训练完成的融合模型进行电力系统暂态稳定评估，生成的电力系统暂态稳定评估结论。
优选地，将融合模型生成的电力系统暂态稳定评估结论发送至电网调控系统。
综上，本发明结合人工智能技术与电力系统暂态稳定评估的需求，提出了基于电力稳态特征变量像素化和融合模型的电力系统暂态稳定评估方法，明确了电力稳态特征变量到计算机图像的转义机制，建立了融合模型，提出了数据-转义-图像-模型的总体技术框架。
该电力系统暂态稳定评估方法为人工智能技术如何与电力系统态势感知技术相融合提供了实际性的指导和帮助，所提出的电力稳态特征变量像素化方法有利于充分发挥人工智能技术对图像数据的处理优势；结合ResNet和Transformer的融合模型强大的数据挖掘能力，满足在线安全评估的时效性和准确性需求，充分发挥了机器智能的作用，为提高电网调控智能化水平和减轻调控人员工作强度提供有效手段，具有较高的工程适用性。
如图8所示，本申请实施例的电力系统安全稳定快速评估装置，包括：
构建样本集单元10，用于将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，并构建样本集；所述电力稳态特征变量的电气量类型包括：电压、有功功率、无功功率；
融合模型训练单元20，用于利用构建的所述样本集，训练基于深度残差网络与Transformer的融合 模型，训练完成的所述融合模型用于电力系统安全稳定快速评估；
暂态稳定评估单元30，用于将实时测量的电力稳态特征变量像素化，生成在线的三通道像素化图像；利用训练完成的融合模型处理生成的所述在线的三通道像素化图像，生成电力系统暂态稳定状态评估结论。
本申请实施例的电力系统安全稳定快速评估装置与本申请实施例的电力系统安全稳定快速评估方法出于相同的发明构思，可以参考实施，并具有与其采用、运行或实现的方法相同的有益效果。
本申请实施方式还提供一种终端，以执行所述方法。请参考图9其示出了本申请的一些实施方式所提供的一种终端的示意图。参考图9所示，终端8包括：处理器800，存储器801，总线802和通信接口803，所述处理器800、通信接口803和存储器801通过总线802连接；所述存储器801中存储有可在所述处理器800上运行的计算机程序，所述处理器800运行所述计算机程序时执行本申请所述任一实施方式所提供的方法。
其中，存储器801可能包含高速随机存取存储器(RAM：Random Access Memory)，也可能还包括非不稳定的存储器(non-volatile memory)，例如至少一个磁盘存储器。通过至少一个通信接口803(可以是有线或者无线)实现该装置网元与至少一个其他网元之间的通信连接，可以使用互联网、广域网、本地网、城域网等。
总线802可以是ISA总线、PCI总线或EISA总线等。所述总线可以分为地址总线、数据总线、控制总线等。其中，存储器801用于存储程序，所述处理器800在接收到执行指令后，执行所述程序，所述本申请实施例任一实施方式揭示的所述方法可以应用于处理器800中，或者由处理器800实现。
处理器800可能是一种集成电路芯片，具有信号的处理能力。在实现过程中，所述方法的各步骤可以通过处理器800中的硬件的集成逻辑电路或者软件形式的指令完成。所述的处理器800可以是通用处理器，包括中央处理器(Central Processing Unit，简称CPU)、网络处理器(Network Processor，简称NP)等；还可以是数字信号处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以实现或者执行本申请实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。结合本申请实施例所公开的方法的步骤可以直接体现为硬件译码处理器执行完成，或者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器801，处理器800读取存储器801中的信息，结合其硬件完成所述方法的步骤。
本申请实施例提供的终端与本申请实施例的方法出于相同的发明构思，具有与其采用、运行或实现的方法相同的有益效果。
本申请实施方式还提供一种与所述实施方式所提供的方法对应的计算机可读存储介质，所述计算机可读存储介质为光盘，其上存储有计算机程序(即程序产品)，所述计算机程序在被处理器运行时，会执行所述任意实施方式所提供的方法。
需要说明的是，所述计算机可读存储介质的例子还可以包括，但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他光学、磁性存储介质，在此不再一一赘述。
本申请的所述实施例提供的计算机可读存储介质与本申请实施例的方法出于相同的发明构思，具有与其存储的应用程序所采用、运行或实现的方法相同的有益效果。
显然，本领域的技术人员可以对本申请进行各种改动和变型而不脱离本申请的精神和范围。这样，倘若本申请的这些调整和变型属于本申请权利要求及其等同技术的范围之内，则本申请也意图包含这些改动和变型在内。
工业实用性
本申请实施例中，电力系统安全稳定快速评估技术方案包括离线环节及在线环节。在离线环节，获取电力结构化数据，构建转义机制，融合深度残差网络与Transformer网络，建立基于计算机图像的融合模型，实现电力系统安全稳定评估。在线环节实时从电网调控系统获取电网运行参数，将获取的电网运行参数转义，生成电压像素化图像、有功功率像素化图像及无功功率像素化图像等多通道像素化图像；利用训练完成的基于计算机图像的融合模型处理生成的多通道像素化图像，生成电力系统安全稳定评估结论，如系统稳定或系统失稳。如此，基于人工智能的安全稳定快速评估方法，通过在海量、多型和低价值密度的电网运行数据中深度挖掘内在映射规律，刻画电力系统的暂稳边界，实现对电网暂稳(暂态 稳定)状态的评估。此方法不受物理模型精度限制，对电网态势的感知可以在毫秒内实现，能够满足在线安全评估在时效性和准确性上的需求平衡，能够与电网调控系统进行有效的结合，为提高电网态势感知智能化水平和减轻调控人员工作强度提供有效技术手段。

Claims (10)

  1. 一种电力系统安全稳定快速评估方法，包括：
    将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，并构建样本集；所述电力稳态特征变量的电气量类型包括：电压、有功功率、无功功率；
    利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，训练完成的所述融合模型用于电力系统安全稳定快速评估；
    将实时测量的电力稳态特征变量像素化，生成在线的三通道像素化图像；利用训练完成的所述融合模型处理生成的所述在线的三通道像素化图像，生成电力系统暂态稳定状态评估结论。
  2. 如权利要求1所述的电力系统安全稳定快速评估方法，其中，所述将测量获取的或仿真生成的电力稳态特征变量像素化，包括：
    根据电网拓扑结构，利用转义机制，将电网进行像素化处理，将电网各元件映射为基础图像内的各平面图形；所述基础图像包括非母线基础图像及母线基础图像。
  3. 如权利要求2所述的电力系统安全稳定快速评估方法，其中，所述将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，包括：
    将所述电力稳态特征变量映射到所述非母线基础图像，生成有功功率像素化图像及无功功率像素化图像，将所述电力稳态特征变量映射到所述母线基础图像，生成电压像素化图像；
    将所述有功功率像素化图像、所述无功功率像素化图像及所述电压像素化图像结构化为所述离线的三通道像素化图像。
  4. 如权利要求3所述的电力系统安全稳定快速评估方法，其中，所述构建样本集，包括：
    利用预先构建的电网时域仿真模型，对多种典型电网运行方式设置可能导致系统失稳的故障，对电网进行离线仿真，获取注入故障之前电网稳定运行时的状态量；
    从多种典型电网运行方式下获取的注入故障之前的电网稳定运行时的状态量中选取多个用于表征系统暂态稳定性的电气量作为电力稳态特征变量；
    将选取的所述电力稳态特征变量像素化，生成多组离线的三通道像素化图像；
    将各组三通道像素化图像与其对应的暂态稳定标签组合为各样本，形成样本集。
  5. 如权利要求3所述的电力系统安全稳定快速评估方法，其中，所述利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，包括：
    将样本集内各样本对应的三通道像素化图像作为深度残差网络的输入，将深度残差网络生成的矩阵特征图作为基于多头注意力机制的Transformer的输入，将样本集内各样本对应的暂态稳定标签作为基于多头注意力机制的Transformer的输出。
  6. 如权利要求5所述的电力系统安全稳定快速评估方法，其中，所述深度残差网络包括n个残差模块，其中，n为不小于3的自然数；
    深度残差网络处理三通道像素化图像，包括：
    将获取的三通道像素化图像作为输入做标准卷积处理，得到标准卷积结果；
    对所述标准卷积结果做最大池化，得到最大池化结果；
    在残差路径上，将所述最大池化结果作为输入，依次经过N1个串联的第一残差模块，生成第一残差模块组处理结果；
    将第i残差模块组处理结果作为输入，依次经过N(i+1)个串联的第(i+1)残差模块，生成第(i+1)残差模块组处理结果，其中，i大于或等于1且小于n；
    将生成的第n残差模块组处理结果，作为经深度残差网络处理后输出的矩阵特征图。
  7. 如权利要求5所述的电力系统安全稳定快速评估方法，其中，
    所述基于多头注意力机制的Transformer包括多个Transformer编码器模块、位置编码模块、分类信息模块；
    所述Transformer编码器模块包括多头注意力层、前馈网络层、残差连接单元和层归一化单元。
  8. 如权利要求5所述的电力系统安全稳定快速评估方法，其中，所述训练基于深度残差网络与Transformer的融合模型，包括：
    选用基于混淆矩阵的准确率、精确率、召回率以及F1值构建评估指标，以评估融合模型的性能。
  9. 一种电力系统安全稳定快速评估装置，包括：
    构建样本集单元，用于将测量获取的或仿真生成的电力稳态特征变量像素化，生成离线的三通道像素化图像，并构建样本集；所述电力稳态特征变量的电气量类型包括：电压、有功功率、无功功率；
    融合模型训练单元，用于利用构建的所述样本集，训练基于深度残差网络与Transformer的融合模型，训练完成的所述融合模型用于电力系统安全稳定快速评估；
    暂态稳定评估单元，用于将实时测量的电力稳态特征变量像素化，生成在线的三通道像素化图像；利用训练完成的所述融合模型处理生成的所述在线的三通道像素化图像，生成电力系统暂态稳定状态评估结论。
  10. 一种终端，包括：存储器、处理器及存储在该存储器上并可在该处理器上运行的计算机程序，其中，所述处理器执行所述计算机程序以实现如权利要求1至8中任一项所述方法。
PCT/CN2024/129722 2024-07-02 2024-11-04 电力系统安全稳定快速评估方法、装置及终端 Pending WO2026007294A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202410873998.X 2024-07-02
CN202410873998.XA CN118410305B (zh) 2024-07-02 2024-07-02 电力系统安全稳定快速评估方法、装置及终端

Publications (1)

Publication Number Publication Date
WO2026007294A1 true WO2026007294A1 (zh) 2026-01-08

Family

ID=91996337

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2024/129722 Pending WO2026007294A1 (zh) 2024-07-02 2024-11-04 电力系统安全稳定快速评估方法、装置及终端

Country Status (2)

Country Link
CN (1) CN118410305B (zh)
WO (1) WO2026007294A1 (zh)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118410305B (zh) * 2024-07-02 2024-09-27 中国电力科学研究院有限公司 电力系统安全稳定快速评估方法、装置及终端
CN118706192B (zh) * 2024-08-27 2025-01-21 青岛峻海物联科技有限公司 一种电气柜电路保护防火监测方法及系统
CN120049438A (zh) * 2025-04-25 2025-05-27 南京邮电大学 一种电力系统状态估计方法、装置、介质及设备
CN120782296A (zh) * 2025-07-07 2025-10-14 三峡大学 一种基于gaf-swin transformer的电力系统暂态电压稳定评估方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200184308A1 (en) * 2018-12-06 2020-06-11 University Of Tennessee Research Foundation Methods, systems, and computer readable mediums for determining a system state of a power system using a convolutional neural network
CN115311509A (zh) * 2022-08-29 2022-11-08 广东电网有限责任公司广州供电局 基于图像化数据驱动的电力系统暂态稳定评估方法及系统
CN117475462A (zh) * 2023-10-27 2024-01-30 内蒙古电力(集团)有限责任公司包头供电分公司 配网线路成图要素及特征自动识别匹配系统及方法
CN118410305A (zh) * 2024-07-02 2024-07-30 中国电力科学研究院有限公司 电力系统安全稳定快速评估方法、装置及终端

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018125220A1 (en) * 2016-12-30 2018-07-05 Intel Corporation Systems, methods, and apparatuses for implementing opc modeling via machine learning on simulated 2d optical images for sed and post sed processes

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200184308A1 (en) * 2018-12-06 2020-06-11 University Of Tennessee Research Foundation Methods, systems, and computer readable mediums for determining a system state of a power system using a convolutional neural network
CN115311509A (zh) * 2022-08-29 2022-11-08 广东电网有限责任公司广州供电局 基于图像化数据驱动的电力系统暂态稳定评估方法及系统
CN117475462A (zh) * 2023-10-27 2024-01-30 内蒙古电力(集团)有限责任公司包头供电分公司 配网线路成图要素及特征自动识别匹配系统及方法
CN118410305A (zh) * 2024-07-02 2024-07-30 中国电力科学研究院有限公司 电力系统安全稳定快速评估方法、装置及终端

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"Engineering Science and Technology II, China Master's Theses Full-Text Database", 1 June 2020, article HAO BOWEN: "Intelligent Fault Diagnosis Technology Based on Computer-visualized Power Flow", XP093385081, DOI: 10.27061/d.cnki.ghgdu.2020.003000 *
FAN SHIXIONG, LI LIXIN, WANG SONGYAN, LIU XINGWEI, YU YIJUN, HAO BOWEN: "Application Analysis and Exploration of Artificial Intelligence Technology in Power Grid Dispatch and Control", POWER SYSTEM TECHNOLOGY, SHUILI-DIANLIBU DIANLI KEXUE YANJIUYUAN, BEIJING, CN, vol. 44, no. 2, 2 December 2019 (2019-12-02), CN , pages 401 - 411, XP093385086, ISSN: 1000-3673, DOI: 10.13335/j.1000-3673.pst.2019.1842 *

Also Published As

Publication number Publication date
CN118410305B (zh) 2024-09-27
CN118410305A (zh) 2024-07-30

Similar Documents

Publication Publication Date Title
WO2026007294A1 (zh) 电力系统安全稳定快速评估方法、装置及终端
Hu et al. Real-time transient stability assessment in power system based on improved SVM
Nie et al. Optimizing the post-disaster control of islanded microgrid: A multi-agent deep reinforcement learning approach
Du et al. Fast cascading outage screening based on deep convolutional neural network and depth-first search
Nejati et al. A new stochastic search technique combined with scenario approach for dynamic state estimation of power systems
Wang et al. Actuator placement for enhanced grid dynamic performance: A machine learning approach
CN113935237A (zh) 一种基于胶囊网络的输电线路故障类型判别方法及系统
Valencia-Zuluaga et al. A fast decomposition method to solve a security-constrained optimal power flow (SCOPF) problem through constraint handling
CN119762290A (zh) 故障分析、电力设备故障分析及故障分析模型训练方法
Yin et al. CNN-based digital twin model for ultra-high voltage direct current system loss measurement
Thotakura et al. Time Series Prediction of Power System Events Using LSTM, GRU, and Transformer Networks to Increase Computation Efficiency
Zhang et al. Model and data driven machine learning approach for analyzing the vulnerability to cascading outages with random initial states in power systems
CN118349847A (zh) 一种提高预测效率的日用电负荷预测方法及系统
Wang et al. Image classification of missing insulators based on EfficientNet
CN118693833A (zh) 一种电压稳定紧急控制方法、系统、终端及介质
He et al. Contour–context joint blind image inpainting network for molecular sieve particle size measurement of SEM images
Moradzadeh et al. Image processing-based data integrity attack detection in dynamic line rating forecasting applications
Wen et al. Design and Implementation of a Power Grid Fault State Analysis System Based on Deep Learning
Wang et al. Interpretable Transient Frequency Stability Prediction Based on Graph Convolutional Learning
CN118232318B (zh) 台风场景下的短期风电功率预测方法、系统、介质及设备
Peng et al. Improved YOLOX Lightweight Algorithm for Power System Component and Defect
CN120671563B (zh) 一种电力混合模型构建方法、系统、电子设备及存储介质
CN121146491A (zh) 考虑运行方式变化的电力系统快速静态安全分析方法
Zhang et al. Data-driven frequency stability assessment of power systems considering wind power uncertainty
CN120234735A (zh) 一种电网局部放电缺陷识别方法、装置、终端设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 24946160

Country of ref document: EP

Kind code of ref document: A1