Nonlinear mapping method and system for stability domain of isolated network based on real-time power of source, load and storage

Abstract

This invention provides a nonlinear mapping method and system for the stability domain of an isolated grid based on real-time power from source, load, and storage. By collecting real-time power data from source, load, and storage and constructing a multi-dimensional time-series power feature map, the method utilizes an encoder-decoder network incorporating a multi-head attention mechanism to perform differentiated weighted fusion of power features. Regression mapping yields key state indicators describing the stability domain boundary. A dynamic stability domain is then generated using a convex hull construction algorithm, and a real-time stability margin index is calculated based on the shortest distance from the current operating point to the stability domain boundary. When this index falls below a warning threshold, an emergency control signal is triggered. This application achieves end-to-end fast nonlinear mapping from high-dimensional time-series power data to the stability domain boundary, significantly improving the real-time performance and accuracy of online assessment of the isolated grid stability domain, and providing reliable support for the safe and stable operation of isolated grids.

Description

Technical Field

[0001] This application relates to the field of power system stability analysis and control technology, specifically to a nonlinear mapping method and system for the stability domain of an isolated grid based on real-time power from source, load, and storage. Background Technology

[0002] An isolated grid system, also known as a microgrid operating in island mode, refers to a local power network that operates independently after being disconnected from the main power grid. Due to the lack of rotational inertia support and power regulation / peak-shaving capabilities of the main grid, isolated grid systems have low equivalent inertia and weak damping. Their frequency and voltage are extremely sensitive to power disturbances such as power output at the source, load fluctuations at the load end, and charging / discharging switching at the energy storage end. When the real-time power of the source, load, and energy storage fluctuates significantly or rapidly, the system is highly susceptible to exceeding the transient stability boundary, leading to frequency collapse or voltage instability. Therefore, accurately and quickly characterizing the stable operating boundary (i.e., the stability domain) of the isolated grid system in the current and short-term future is a crucial prerequisite for ensuring the safe and stable operation of the isolated grid.

[0003] Currently, traditional methods for assessing the stability domain of power systems primarily rely on offline electromechanical transient time-domain simulation. These methods, based on detailed component models and network equations, calculate a series of stability indicators such as critical clearing time and frequency / voltage extremes by numerically integrating anticipated fault sets and operating modes, thereby artificially defining stability boundaries. However, time-domain simulation methods are computationally intensive, time-consuming per simulation, and their results are highly dependent on pre-set specific operating conditions, making it difficult to cover the massive operating scenarios resulting from the continuous and rapid changes in source-load-storage power in high-dimensional space. If directly applied to online assessment, it cannot keep up with the real-time changes in the isolated grid's operating state, resulting in significant lag in the assessment results.

[0004] To improve computational speed, some studies have proposed data-driven methods based on simplified models or machine learning. Simplified model methods, such as single-machine equivalent methods and energy function methods in direct methods, can provide fast criteria, but they essentially make strong approximations of complex systems, making it difficult to retain all the dynamic characteristics of the system, resulting in insufficient accuracy in describing the stability domain boundaries. On the other hand, methods that directly fit the relationship between operating parameters and stability margin using conventional fully connected or convolutional neural networks often treat source-load-storage power data of different times and types with equal weight, failing to highlight the differentiated impact of system inertial response and continuous temporal dependencies on stability, thus creating bottlenecks in scenario generalization ability and mapping accuracy.

[0005] In summary, existing technologies struggle to achieve high-precision online mapping of the stability domain of isolated networks while maintaining assessment speed. The root cause of this predicament lies in the lack of a technical solution capable of effectively mining the high-dimensional temporal characteristics of real-time power from sources, loads, and storage, and accurately and explicitly constructing the complex nonlinear mapping relationship between its time-varying fluctuations and the system stability boundary. Therefore, how to achieve a rapid and accurate mapping from high-dimensional time-series power data to the stability domain boundary is a pressing technical problem that needs to be solved. Summary of the Invention

[0006] This application provides a nonlinear mapping method and system for the stable domain of an isolated grid based on real-time power of source, load, and storage, which at least addresses the problems existing in the prior art.

[0007] A first aspect of this application provides a method for nonlinear mapping of the stability domain of an isolated grid based on real-time power of source, load, and storage, comprising the following steps: S1: By installing measurement units at each node and branch of the isolated network, the power generation at the source end, the load power at the load end, and the charging and discharging power at the storage end within a preset time window are collected to form a real-time power data sequence characterizing the system's operating status. S2: Use a sliding time window to segment and stack the real-time power data sequence to construct a multi-dimensional time-series power feature map containing time-series dependencies, which serves as the input to the trained nonlinear mapping network; S3: Input the multi-dimensional temporal power feature map into the nonlinear mapping network. The encoder in the nonlinear mapping network uses a multi-head attention mechanism to assign differentiated weights to power features of different time periods and different types in the feature map, and calculates a weighted fused power state feature vector. S4: The decoder in the nonlinear mapping network performs regression mapping on the power state feature vector and outputs N key state indicators used to describe the boundary of the stable domain under the current operating state, where N is an integer greater than 1; S5: Based on the values ​​of N key state indicators, in the N-dimensional state space with the N key state indicators as coordinate axes, use the convex hull construction algorithm to generate a minimum convex polyhedron with the values ​​as vertices, which serves as the dynamic stable domain at the current moment. S6: Map the weighted fused power state feature vector to an N-dimensional state space to obtain a current operating point; based on the preset stability margin evaluation function, calculate the shortest distance from the current operating point to each boundary surface of the dynamic stability domain, and normalize the shortest distance to obtain the real-time stability margin index. S7: When the real-time stability margin index is determined to be lower than the preset warning threshold, an alarm control trigger signal is sent to the isolated grid energy management system.

[0008] This application embodiment collects real-time power data from source, load, and storage and constructs a multi-dimensional time-series power feature map. It utilizes a nonlinear mapping network incorporating a multi-head attention mechanism to perform differentiated weighted fusion of the time-series dependencies of power features. Then, it obtains key state indicators and generates a dynamic stability domain through decoder regression mapping. Finally, it makes early warning judgments based on the real-time stability margin index. This achieves end-to-end fast nonlinear mapping from high-dimensional time-series power data to the stability domain boundary, significantly improving the real-time performance and accuracy of online evaluation of isolated network stability domain.

[0009] In this embodiment of the application, the training process of the nonlinear mapping network specifically includes: S31: Using a preset electromechanical transient simulation model of the power system, a large number of operating scenarios are generated by random combination within the preset fluctuation range of power generation at the source end, load power at the load end, and charging and discharging power at the storage end. S32: For each operating scenario, perform time-domain simulation, input the simulation trajectory data into the stability criterion analysis module, and automatically mark a set of key state indicators of the stability domain boundary corresponding to the scenario to form a labeled sample. S33: Using a multi-dimensional temporal power feature map as input and labeled samples as the desired output, the encoder-decoder network containing a multi-head attention mechanism is trained offline using the gradient descent method until the preset loss function converges, thus obtaining a nonlinear mapping network.

[0010] This application embodiment generates a massive number of operating scenarios by randomly combining source-load-storage power within a wide range of fluctuations. It also automatically labels key state indicators of the stability domain boundary using electromechanical transient simulation models and stability criteria, constructing high-quality labeled samples for offline training. This enables the nonlinear mapping network to fully learn the complex mapping relationship between power fluctuations and the stability domain boundary under different operating scenarios, thereby improving the network's generalization ability and mapping accuracy.

[0011] In this embodiment of the application, during the offline training process of S33, the encoder calculates the weighted fused power state feature vector through a multi-head attention mechanism, specifically including: S331: Perform a linear transformation on the input multidimensional time-series power feature map to generate a query matrix, key matrix, and value matrix corresponding to multiple attention heads; S332: For each attention head, calculate the dot product of the query matrix and the key matrix and scale it to obtain an attention score matrix that reflects the degree of correlation between power features at different time steps; S333: Use the Softmax function to convert the attention score matrix into an attention weight matrix, and then perform a weighted summation of the attention weight matrix and the value matrix to obtain the output features of the attention head; S334: Concatenate the output features of all attention heads and fuse them through a fully connected layer to generate a weighted fused power state feature vector.

[0012] This application embodiment utilizes a multi-head attention mechanism to perform linear transformations of the query matrix, key matrix, and value matrix; dot product scaling to calculate attention scores; softmax normalization to obtain attention weights; and multi-head feature splicing and fusion. This enables the network to simultaneously capture fine-grained correlations between power features at different time steps from multiple representation subspaces, thereby enhancing the ability of the fused power state feature vector to express key temporal information.

[0013] In this embodiment of the application, the N key state indicators output in step S4 to describe the boundary of the stable domain are specifically as follows: The decoder of the nonlinear mapping network outputs the coordinates of M vertices. The coordinates of each vertex are composed of the values ​​of N key state indicators, including the critical cut-off time, the lowest frequency, the lowest voltage, and the damping ratio of the dominant oscillation mode. The M vertices are used to jointly describe the boundary of the stability region. Here, N=4, and M is an integer greater than N.

[0014] This application's embodiments select the critical cut-off time, the lowest frequency point, the lowest voltage point, and the damping ratio of the dominant oscillation mode as key state indicators to describe the boundary of the stability domain. It comprehensively characterizes the boundary of the stability domain from multiple dimensions such as power angle stability, frequency stability, voltage stability, and small disturbance stability, so that the generated dynamic stability domain can fully reflect the transient stability characteristics of the isolated network system.

[0015] In this embodiment of the application, the dynamic stable region is generated in step S5 using a convex hull construction algorithm, specifically as follows: In an N-dimensional state space with N key state indices as coordinate axes, the coordinates of the M vertices output by the decoder are used as a vertex set. All vertices in the vertex set are connected by the fast convex hull algorithm to form a minimal convex polyhedron. The internal region of this convex polyhedron is the dynamic stability domain that ensures the transient stability of the system.

[0016] This application, based on M boundary vertices, explicitly defines a minimum convex polyhedron generated from this vertex set in N-dimensional space using a fast convex hull algorithm as the dynamic stability domain. This transforms discrete boundary vertex information into a continuous, closed geometric operating region. The internal region of this polyhedron strictly guarantees the transient stability of the system, providing a clear and computable geometric object basis for subsequently quantifying the stability margin by calculating the shortest distance from the current operating point to the boundary surface of the polyhedron. This achieves a precise mapping of the stability domain from an abstract index to a specific geometric space.

[0017] In this embodiment of the application, S6 calculates the shortest distance from the current running point to each boundary surface of the dynamic stable domain, including: calculating the shortest Euclidean distance from the current running point to each boundary surface of the dynamic stable domain using a point-to-polyhedron distance algorithm, and taking the minimum value among them as the shortest distance.

[0018] This application embodiment maps the power state feature vector to an N-dimensional state space to determine the current operating point, and calculates the shortest Euclidean distance from it to each face of the dynamic stable domain boundary using a point-to-polyhedron distance algorithm, taking the minimum value as the shortest distance, thereby achieving an accurate quantitative assessment of the danger level of the current operating state.

[0019] In this embodiment of the application, the method further includes: preprocessing the real-time power data sequence, the preprocessing including filtering out measurement noise using a Kalman filter and repairing missing data using linear interpolation to obtain a cleaned power data stream; and using a sliding time window to segment and stack the cleaned power data stream to construct a multi-dimensional time-series power feature map.

[0020] The embodiments of this application filter out measurement noise using a Kalman filter and repair missing data using linear interpolation before constructing feature maps. This effectively suppresses the adverse effects of measurement errors and missing data on the accuracy of subsequent network mapping, and improves the quality and stability of real-time power data sequences.

[0021] In this embodiment of the application, after S7, step S8 is further included, which includes: S81: Encapsulate the real-time stability margin index and its corresponding source-end power generation, load-end power, energy storage charging and discharging power, and dynamic stability domain boundary information into a historical operating state data packet; S82: Periodically feeds back historical running data packets to the training database, and when the preset model update triggering conditions are met, uses an incremental learning mechanism to fine-tune the parameters of the nonlinear mapping network online.

[0022] This application embodiment encapsulates and feeds back real-time stability margin index, source-load-storage power data, and stability domain boundary information to the training database, and uses an incremental learning mechanism to fine-tune the network parameters online when the update trigger condition is met. This enables the nonlinear mapping network to continuously adapt to the slow changes in the operating characteristics of the isolated network system, maintaining the mapping accuracy and adaptability for long-term operation.

[0023] This application embodiment also includes: S51: Obtain the future scheduling plan of the isolated grid system, which includes the predicted power generation at the source end, the predicted power load at the load end, and the planned power charging and discharging at the storage end within a preset future time period; S52: Input the power prediction curve corresponding to the scheduling plan into the nonlinear mapping network to predict the predicted stability region corresponding to each time section within the future preset period. S53: Based on the predicted stability domain, assess whether there will be a risk period in the future where the real-time stability margin index is lower than the warning threshold; S54: If there is a risk period, calculate the minimum energy storage charging and discharging power adjustment required to make the system's stability margin index reach above the warning threshold during the risk period, and generate a forward-looking energy storage pre-scheduling instruction accordingly, which is then fed back to the isolated grid energy management system.

[0024] This application embodiment obtains the power prediction curve in the future scheduling plan and inputs it into a nonlinear mapping network to predict the stability domain of future periods. Then, it assesses risk periods and calculates the minimum energy storage charging and discharging power adjustment amount to generate forward-looking pre-scheduling instructions. This extends the stability domain assessment from real-time monitoring to predictive control, providing technical support for the forward-looking safety management of isolated grids.

[0025] In this embodiment of the application, step S6 calculates the shortest distance from the current running point to each boundary surface of the dynamic stable domain and normalizes it to obtain the real-time stability margin index, specifically including: S61: In the N-dimensional state space, determine the coordinates of the current running point and identify the set of vertices of the boundary surface closest to the current running point; S62: Construct the spatial plane equation of the boundary surface based on the vertex set, and obtain the shortest distance by calculating the perpendicular distance from the current running point to the spatial plane; S63: Calculate the ratio of the shortest distance to the preset baseline distance value to obtain the normalized stability margin index; S64: Based on the stability margin index, generate the stability margin gradient direction vector of the current running point in the N-dimensional state space. The gradient direction vector points to the direction of the fastest growth of the stability margin, which is used to provide adjustment direction guidance for subsequent prevention and control decisions.

[0026] This application embodiment normalizes the shortest vertical distance from the current running point to the boundary surface by identifying the nearest boundary surface and constructing a spatial plane equation. At the same time, it generates a gradient direction vector pointing to the direction of the fastest growth of stability margin. This not only quantitatively assesses the degree of danger, but also provides clear adjustment direction guidance for subsequent prevention and control, thereby improving the pertinence and efficiency of control decisions.

[0027] In this embodiment of the application, the calculation in S54 of the minimum energy storage charging and discharging power adjustment required to keep the system stable during risky periods specifically includes: S541: Obtain the predicted stability domain corresponding to the risk period, and determine the most dangerous critical state indicator corresponding to the boundary closest to the predicted operating point in the predicted stability domain; S542: Using the sensitivity matrix of the most dangerous critical state index to the energy storage charging and discharging power as a constraint, and with the goal of minimizing the adjustment of the energy storage charging and discharging power, a linear programming model is constructed. S543: Solve the linear programming model to obtain the minimum energy storage charging and discharging power adjustment required to bring the most dangerous critical state index back to the predicted stability domain. S544: Based on the minimum energy storage charging and discharging power adjustment amount and the rated power and capacity constraints of the energy storage system, the feasibility of the forward-looking energy storage pre-dispatch command is verified. The pre-dispatch command is only fed back to the isolated grid energy management system when the adjustment amount is within the adjustable range of the energy storage system.

[0028] This application embodiment determines the most dangerous critical state index and uses its sensitivity matrix to energy storage power as a constraint to construct a linear programming model to solve for the minimum adjustment amount. Finally, it combines the rated power and capacity constraints of the energy storage system to perform feasibility verification, ensuring that the pre-scheduling command is technically feasible and the adjustment amount is optimal, thus realizing the refinement and reliability of prevention and control.

[0029] In this embodiment of the application, the calculation process of the multi-head attention mechanism further includes: S335: Perform sparsity constraint processing on the attention weight matrix corresponding to each attention head. The sparsity constraint processing includes: setting an attention weight threshold and setting the weight values ​​below the attention weight threshold to zero in order to suppress interference from irrelevant time steps. S336: Based on the attention weight matrix after sparsification constraint processing, extract the source-load-storage power features corresponding to time steps with non-zero weight values, and construct a sparse temporal feature subset that is strongly correlated with the dominant dynamic characteristics of the isolated network. S337: Perform a residual connection between the sparse temporal feature subset and the attention head output features obtained in S333 to obtain enhanced attention head output features. Use the enhanced attention head output features as input for concatenation in S334 to enhance the nonlinear mapping network's ability to identify key time segments in long temporal dependencies.

[0030] This application embodiment suppresses interference from irrelevant time steps by sparsely constraining the attention weight matrix, extracts strongly correlated sparse temporal feature subsets, and enhances the output features of the attention head through residual connections. This strengthens the network's ability to identify key time segments in long temporal dependencies, suppresses the negative impact of redundant temporal information on the network's generalization performance, and improves the robustness and mapping accuracy of the nonlinear mapping network in complex noise environments.

[0031] A second aspect of this application provides an islanded grid stability domain nonlinear mapping system based on source-load-storage real-time power, used to perform the above method. The system includes: The data acquisition module is configured to execute step S1. The feature map construction module is configured to execute step S2. The encoder module is configured to execute step S3; The decoder module is configured to perform step S4. The stability region generation module is configured to execute step S5. The margin assessment module is configured to execute step S6; The early warning feedback module is configured to execute step S7.

[0032] This application embodiment constructs a system architecture that fully corresponds to the steps of the mapping method by setting up the coordinated cooperation of various functional modules in the system. It realizes a complete closed-loop function from power data acquisition to stable domain generation and early warning feedback, ensuring the coordination and efficiency of system operation. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating a nonlinear mapping method for the stable domain of an isolated grid based on real-time power of source-load-storage, as provided in an embodiment of this application. Detailed Implementation

[0034] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0035] Islanded grid systems, operating independently from the main power grid, lack the rotational inertia and power support provided by the main grid. Their frequency and voltage are extremely sensitive to power disturbances such as power output at the source, load fluctuations at the load end, and charging / discharging switching at the energy storage end. In actual islanded grid operation, the real-time power of the source, load, and energy storage systems is constantly and rapidly changing, causing the system operating point to continuously migrate in the state space. Once the operating point crosses the transient stability boundary, it may trigger serious accidents such as frequency collapse or voltage instability. Therefore, how to quickly and accurately characterize the system's stable operating boundary, i.e., the stability domain, under the current operating state, and how to assess the system's safety margin in real time based on this stability domain, has become a crucial technical problem that needs to be solved to ensure the safe operation of islanded grids.

[0036] In existing technologies, the main method for obtaining the stability domain of a power system is offline electromechanical transient time-domain simulation. This method, based on detailed component models and network equations, calculates stability indicators such as critical clearing time and frequency extrema through numerical integration for preset faults and specific operating conditions. However, time-domain simulation is computationally intensive and cannot cover the massive operating scenarios formed by the continuous changes of source-load-storage power in high-dimensional space. If it is directly used for online assessment, the assessment results are significantly lagging and cannot provide timely stability situation awareness for the rapidly changing operating state of isolated grids. Another type of method based on simplified models, such as the single-machine equivalent method or energy function method, although computationally fast, suffers from insufficient accuracy in describing the stability domain boundary due to extensive approximations of the system, making it difficult to meet the requirements of refined control of isolated grids.

[0037] Recent data-driven stability assessment methods attempt to directly establish a mapping relationship between operating parameters and stability margins using neural networks. However, the inventors of this application discovered during their research that existing conventional fully connected networks or convolutional networks typically treat source-load-storage power data of different times and types with equal weights, failing to fully explore and utilize the continuous temporal dependencies brought about by the system's inertial response, and also failing to distinguish the differentiated impact of power fluctuations at different times on system stability. This results in significant bottlenecks in the network model's scenario generalization ability and mapping accuracy, especially when operating conditions fluctuate drastically, leading to a significant decrease in the reliability of the assessment results.

[0038] To address the aforementioned problems, the inventors, through in-depth research and repeated experiments, conceived and proposed the technical solution of this application. The inventors recognized that in continuous time-series power data operating in isolated networks, not all power fluctuations at all times have an equally significant impact on the current stability domain boundary. Large power disturbances at certain critical moments often determine the system's stability margin more decisively than long-term stable fluctuations. Based on this understanding, the inventors innovatively introduced a multi-head attention mechanism into the construction of the stability domain mapping network, enabling the network model to learn autonomously and assign differentiated weights to power features at different time steps, thereby focusing on the critical moments and variables affecting the stability boundary. Furthermore, the inventors designed a regression mapping path from power feature vectors to multi-dimensional key state indicators, and used a convex hull construction algorithm to generate dynamic stability domains in the state space with these indicators as vertices, thus establishing a complete mapping link from high-dimensional time-series power data to the geometrically defined stability domain boundary.

[0039] Building upon this foundation, the inventors further conceived a scheme for calculating the real-time stability margin index based on the shortest distance from the current operating point to the boundary of the dynamic stability domain, as well as a technical path for stability domain prediction and forward-looking energy storage pre-scheduling based on future scheduling plans. This forms a complete technical solution covering real-time assessment, early warning triggering, and predictive control. This solution effectively solves the technical problems of slow online stability domain assessment, low accuracy, and inability to distinguish the impact of time-series information differences in existing technologies. It achieves rapid, accurate, and real-time mapping of the stability domain of isolated networks, providing reliable technical support for the safe and stable operation of isolated networks. To make the purpose, technical solution, and advantages of this application clearer, the following will be discussed in conjunction with the appendix. Figure 1 The following is an explanation using specific examples.

[0040] Please refer to Figure 1 , Figure 1 The first aspect of this application provides a method for nonlinear mapping of the stability domain of an isolated grid based on real-time power of source, load, and storage, comprising the following steps: S1: By installing measurement units at each node and branch of the isolated network, the power generation at the source end, the load power at the load end, and the charging and discharging power at the storage end within a preset time window are collected to form a real-time power data sequence characterizing the system's operating status. S2: Use a sliding time window to segment and stack the real-time power data sequence to construct a multi-dimensional time-series power feature map containing time-series dependencies, which serves as the input to the trained nonlinear mapping network; S3: Input the multi-dimensional temporal power feature map into the nonlinear mapping network. The encoder in the nonlinear mapping network uses a multi-head attention mechanism to assign differentiated weights to power features of different time periods and different types in the feature map, and calculates a weighted fused power state feature vector. S4: The decoder in the nonlinear mapping network performs regression mapping on the power state feature vector and outputs N key state indicators used to describe the boundary of the stable domain under the current operating state, where N is an integer greater than 1; S5: Based on the values ​​of N key state indicators, in the N-dimensional state space with the N key state indicators as coordinate axes, use the convex hull construction algorithm to generate a minimum convex polyhedron with the values ​​as vertices, which serves as the dynamic stable domain at the current moment. S6: Map the weighted fused power state feature vector to an N-dimensional state space to obtain a current operating point; based on the preset stability margin evaluation function, calculate the shortest distance from the current operating point to each boundary surface of the dynamic stability domain, and normalize the shortest distance to obtain the real-time stability margin index. S7: When the real-time stability margin index is determined to be lower than the preset warning threshold, an alarm control trigger signal is sent to the isolated grid energy management system.

[0041] In this application, in S1, the measurement unit refers to the power acquisition device installed at each power generation node, load node, energy storage grid connection point, and key interconnection branch of the islanded grid system. It can be understood as a phasor measurement unit or a module with power sampling function in an intelligent electronic device, and its sampling frequency can be configured according to the dynamic response characteristics of the islanded grid. The power generation at the source includes the active and reactive power output of intermittent power sources such as photovoltaic and wind power. The load power at the load end refers to the active and reactive power consumed by various electrical loads. The charging and discharging power at the storage end is represented by positive values ​​for discharging and negative values ​​for charging. The three together constitute a real-time power data sequence. The preset time window refers to the length of the historical period covered by each acquisition. For example, it can be set to a time range of several hundred milliseconds to several seconds before the fault or the current moment to fully capture the dynamic response process of the system after the disturbance.

[0042] In S2, the sliding time window refers to the operation of progressively shifting and extracting multiple equal-length data subsequences from a continuously arriving real-time power data sequence at fixed time steps. By stacking the data subsequences within multiple adjacent windows along the depth dimension, a three-dimensional data structure with a time axis, a power type axis, and a window batch axis is formed, namely a multi-dimensional time-series power feature map. This feature map, while preserving the original sampling time sequence, expresses the temporal dependencies between various power types at different times in an explicit tensor form, providing a structured input for the subsequent network model to simultaneously extract dynamic features at short and long time scales.

[0043] In S3, the nonlinear mapping network refers to a pre-trained deep neural network model whose internal parameters are determined through an offline learning process. It approximates the complex nonlinear function between the power input and the stability domain description index. The encoder is the front-end of this network, responsible for feature extraction and fusion. Its core computational unit is the multi-head attention mechanism. The multi-head attention mechanism simultaneously projects the input features into multiple different representation subspaces, independently calculating the correlation between power features at different time steps in each subspace. This allows it to capture the temporal impact patterns of power fluctuations from multiple perspectives, assigning differentiated weights reflecting the importance of the originally equally weighted time-series data to the stability boundary. The weighted fusion power state feature vector refers to a compact one-dimensional feature representation output by the encoder. This vector extracts and integrates the parts of the original high-dimensional time-series power information that are strongly correlated with the stability domain determination, while suppressing redundant and irrelevant components.

[0044] Imagine an isolated grid system comprising photovoltaic power generation, conventional loads, and battery energy storage. A sliding time window of five consecutive sampling times is used, recording three types of power at each time point: source-side power generation, load-side power, and storage-side charging / discharging power (discharging is a negative value). After step S2, this data is constructed into a 5x3 multidimensional time-series power characteristic map.

[0045] To illustrate the allocation and weighted fusion process of differentiated weights, assume the following events occur within the window: the system is stable for the first two time points; at time 3, the load power suddenly surges, and the energy storage system instantaneously increases its discharge power to support the frequency; at time 4, the load slightly decreases but remains high; at time 5, the system tends towards a new equilibrium. The source-side power generation response is relatively slow due to inertia, and its changes are relatively delayed.

[0046] The power value at this time can be represented as follows: Time 1: Source 100, Load 95, Storage 0 Time 2: Source 100, Load 100, Storage -10 Time 3: Source 95, Load 140, Storage -40 Time 4: Source 90, Load 120, Storage -25 Time 5: Source 100, Load 105, Storage 0 When processing this feature map, the multi-head attention mechanism inside the encoder calculates an attention weight for each small square representing a specific time and type of power feature. This weight reflects the importance of the information contained in that square for determining the current stable domain boundary.

[0047] Suppose the encoder is configured with two attention heads that observe the same set of data from different representation subspaces and give highly differentiated weight assignments.

[0048] Attention Head 1 may evolve into a load mutation sensor. From its perspective, the load power characteristics at time 3 are strongly correlated with the stability region boundary, and it will be assigned a very high weight, such as 0.48; the load power at time 4, as a continuation state after a surge, will have a slightly lower weight of 0.18; while the load power at time 2, as a precursor to a surge, will also receive a weight of 0.08. For source and storage power at the same time interval, the attention head pays very little attention, with weights generally close to 0. The weights for all other times (such as times 1 and 5) and irrelevant power types are also extremely small.

[0049] Attention head 2 may evolve into an energy storage support behavior sensor. It keenly detects the sudden increase in energy storage discharge power at time 3, thus assigning a weight as high as 0.55 to the energy storage power characteristics at time 3; at time 4, energy storage continues to discharge, with a weight of 0.20; and the initial energy storage discharge action at time 2 also receives a weight of 0.10. This head is not particularly concerned with the absolute value of the load power itself, only assigning a small weight when the load is related to the change in energy storage, while the weight of the source power is extremely low.

[0050] The so-called weighted fusion uses these differentiated weights to condense the originally scattered feature map containing 15 basic units into two compact intermediate vectors rich in the information of interest. For attention head 1, all power features of all times and types are weighted and summed according to the weights it generates, resulting in a vector. Since this head allows the load power features at time 3 to contribute the vast majority of information, this vector highly condenses the event of "the load experiencing a violent positive impact at time 3". Similarly, the vector obtained after weighted summation by attention head 2 mainly encodes the dynamic of "the energy storage system providing strong discharge support during times 3 to 4".

[0051] Subsequently, these two vectors, representing different perspectives, are concatenated and then processed through a fully connected network layer for information integration and dimensionality reduction, ultimately outputting a fixed-length power state feature vector. Understandably, this final vector contains neither isolated "time-3 load values" nor the original "energy storage power amplitude." Instead, it integrates information such as the magnitude and rate of load surges, as well as the depth and timing of the instantaneous energy storage response, into a set of abstract stable-state representations. For example, certain dimensions in this vector may exhibit high positive activation values, collectively encoding the pattern of the system tending towards critical oscillation due to a sharp imbalance between supply and demand; conversely, if the system remains stable throughout the window, the activation values ​​of these dimensions will be significantly different.

[0052] Differentiated weights enable the nonlinear mapping network to automatically ignore static, redundant periods, focusing computational resources on key power segments and types that affect the stability domain boundary. The resulting weighted fused power state feature vector then becomes a reliable basis for the subsequent decoder to accurately regress the dynamic stability domain boundary.

[0053] In S4, the decoder is the back-end of the nonlinear mapping network, receiving the power state feature vector output by the encoder. It typically consists of a multi-layer fully connected network and is used to regress this compact feature representation to the stability domain description space. Key state indicators refer to the set of physical quantities that can quantitatively characterize the boundary of the system's transient stability domain. N represents the number of indicators, taking a value greater than 1, to collectively define the boundary shape of the stability domain from multiple dimensions. It can be understood that these indicators correspond to different aspects of the power system's transient stability. Describing the boundary through a set of indicators, rather than a single one, allows for a more complete capture of the critical conditions of the isolated grid under different instability modes.

[0054] In S5, the N-dimensional state space is an abstract mathematical space spanned by N key state indicators as coordinate axes. Each point in this space corresponds to a stability domain boundary state of the system. The convex hull construction algorithm refers to a method for calculating the smallest convex polyhedron containing all points given a set of points. Examples include the fast convex hull algorithm or the incremental convex hull algorithm. The dynamic stability domain is the smallest convex polyhedron generated by the convex hull construction algorithm in this N-dimensional state space. Its interior region represents the operating range within which the system can maintain transient stability, while the boundary surface characterizes the boundary state where the system is critically stable. This stability domain is dynamically updated according to changes in real-time power data.

[0055] In S6, the current operating point is the coordinate position obtained by transforming the weighted fused power state feature vector to an N-dimensional state space through a mapping relationship before the decoder output. This point represents the stability margin state of the system at the current moment. The stability margin evaluation function is a pre-defined mathematical calculation process that calculates the shortest spatial distance between the current operating point and each boundary surface of the dynamic stability domain, and normalizes it by comparing this distance with a reference value to obtain a dimensionless real-time stability margin index. The magnitude of this index intuitively reflects the relative distance between the current operating point of the system and the critical boundary.

[0056] In S7, the warning threshold is a lower limit of the stability margin index pre-set according to the islanded grid safe operation procedures or offline analysis. When the real-time stability margin index is lower than this threshold, it indicates that the system operating point has become too close to or has entered the boundary region of the dynamic stability domain, posing a risk of instability. The alarm control trigger signal is an alarm and command message sent to the islanded grid energy management system. This signal can trigger preset stability control strategies such as generator tripping, load shedding, or emergency power support from energy storage to pull the system operating point back into the stability domain.

[0057] The above steps form a progressive and interconnected processing chain, with the output of each step serving as the input for the next step, collaboratively completing the entire process from raw power data to stability domain assessment and early warning.

[0058] In the data acquisition and processing phase, S1 acquires real-time power data from each port of the isolated network system through the measurement unit, providing fundamental information reflecting the actual operating status of the system for the entire evaluation process. Based on this, S2 utilizes a sliding time window to perform temporal structured organization, converting the discrete power samples arriving in chronological order into a multi-dimensional temporal power feature map that preserves temporal dependencies. This conversion is not a simple data format adjustment, but rather explicitly presents the dynamic evolution patterns implicit in the continuous sampling sequence in a structured form. This allows the network model to perceive the sequential correlation between power at different times without additional time encoding, laying a structural foundation for the subsequent effective weight allocation by the attention mechanism. In the feature extraction and mapping phase, S3 and S4 collaboratively constitute the core inference link of the nonlinear mapping. The encoder in S3 receives the multi-dimensional temporal power feature map and uses a multi-head attention mechanism to calculate the correlation strength between power features at different time steps in parallel across multiple representation subspaces. Since the impact of power fluctuations at different time steps on the stability domain boundary varies, the attention mechanism can autonomously learn, through data-driven methods, which power changes at which times are more decisive for the stability margin and assign them higher weights accordingly, while irrelevant or disruptive times are assigned lower weights. The weighted and fused power state feature vector thus becomes a highly refined information carrier, concentrating and retaining the components strongly correlated with stability from the original time-series features. The decoder of S4 takes this feature vector and further maps it into N key state indicators with dimensions far lower than the original input. Understandably, these key state indicators directly correspond to the physical meaning of the stability domain boundary. This mapping process achieves dimensionality reduction and abstraction from a high-dimensional time-series measurement space to a low-dimensional stability description space, allowing the stability domain boundary to be explicitly expressed with a small number of geometric elements, providing the basic coordinates for subsequent geometric construction. In the stability domain construction and evaluation stage, S5 uses the key state indicator values ​​output by S4 to generate the dynamic stability domain. By using these indicators as vertices in the N-dimensional state space and constructing a minimum convex polyhedron using the convex hull algorithm, the dynamic stability domain geometrically and intuitively presents the current stable operating range of the system, with its boundary surface representing the location of the critical instability state in the state space. S6 then associates the geometric domain with the system's current operating point. By calculating and normalizing the shortest distances from the current operating point to each boundary surface, the resulting real-time stability margin index can quantitatively describe the relative safety level of the current operating state. Compared to directly using a single index threshold for judgment, the evaluation method based on point-to-polyhedron distance can comprehensively consider boundary constraints in various dimensions. When the system approaches the critical point first in a certain dimension, the distance calculation can promptly capture the dangerous trend in that dimension. In terms of feedback loop, S7 compares the real-time stability margin index output by S6 with the warning threshold. When the margin is insufficient, an emergency control signal is triggered, converting the evaluation result into executable control commands.This step establishes a closed-loop connection between assessment and control, enabling assessment results to directly drive protection actions and shortening the time delay from state perception to control response.

[0059] The above steps, with power data flow as the main thread, sequentially complete the progressive processing of time-series structuring, attention-weighted fusion, dimensionality reduction mapping, geometric domain construction, margin quantification and evaluation, and early warning feedback. Time-series structuring makes implicit dynamic dependencies explicit; the attention mechanism distinguishes the importance of information at different times; dimensionality reduction mapping allows the stability boundary to be compactly expressed by a few indicators; convex hull construction provides an intuitive geometric reference for the boundary; distance calculation quantifies the level of security; and early warning feedback enables the evaluation results to be practically applied. Each step is interconnected and interdependent, decomposing the technical challenge of mapping high-dimensional time-series power data to the stability domain boundary into multiple cascaded, processable sub-problems. This overall synergy enables rapid end-to-end processing capabilities from raw measurements to control signals.

[0060] In some embodiments of this application, the training process of the nonlinear mapping network specifically includes: S31: Using a preset electromechanical transient simulation model of the power system, a large number of operating scenarios are generated by random combination within the preset fluctuation range of power generation at the source end, load power at the load end, and charging and discharging power at the storage end. S32: For each operating scenario, perform time-domain simulation, input the simulation trajectory data into the stability criterion analysis module, and automatically mark a set of key state indicators of the stability domain boundary corresponding to the scenario to form a labeled sample. S33: Using a multi-dimensional temporal power feature map as input and labeled samples as the desired output, the encoder-decoder network containing a multi-head attention mechanism is trained offline using the gradient descent method until the preset loss function converges, thus obtaining a nonlinear mapping network.

[0061] In this application, S31, the preset power system electromechanical transient simulation model refers to a numerical simulation environment pre-constructed based on the actual topology of the isolated grid, the dynamic parameters of each component, and the logic of the control and protection system. It can be understood as a full-system mathematical model built in power system simulation software capable of simulating electromechanical transient processes. The preset fluctuation ranges for source-side power generation, load-side power, and energy storage charging / discharging power refer to the range of maximum and minimum values ​​that various power outputs may occur, determined based on historical operating data statistics or operational planning of the isolated grid. For example, photovoltaic output can be set to any value between zero and its rated installed capacity, load power can be set to fluctuate between peak and valley loads, and energy storage power can be set between maximum charging power and maximum discharging power. Random combination refers to independently extracting and combining the values ​​of each power variable within their respective preset fluctuation ranges using random sampling methods such as Latin hypercube sampling or Monte Carlo sampling to generate a set of operating scenarios covering a massive number of different operating conditions in a high-dimensional power space.

[0062] In S32, time-domain simulation refers to using the operating scenario generated in S31 as the initial operating condition, and under preset fault triggering conditions (such as a three-phase short circuit on a certain line), using an electromechanical transient simulation model to progressively solve the trajectory of each state variable of the system changing over time through numerical integration. The stability criterion analysis module is an automatic analysis program unit integrated into the simulation environment. Its function is to determine whether the system is unstable based on the dynamic response characteristics of each generator, such as the power angle deviation, system frequency shift, and node voltage drop, according to preset engineering practical stability criteria, and to record the corresponding key state index values ​​under the simulation conditions of critical stability. Automatic labeling refers to the stability criterion analysis module performing iterative simulations by repeatedly adjusting disturbance parameters such as fault clearing time until it finds the system exactly in a critical stable boundary state. It then extracts the corresponding critical clearing time, minimum frequency point, and other index values ​​under this boundary state as label values, thereby establishing a correspondence between input features and expected outputs for each operating scenario, forming labeled samples. It can be understood that an operating scenario may generate a set of key state indices for the stability domain boundary.

[0063] In S33, gradient descent refers to an iterative optimization algorithm that minimizes the loss function by progressively updating the parameter values ​​along the negative gradient direction of the loss function with respect to the network parameters. For example, it can be stochastic gradient descent or its variants such as the adaptive moment estimator optimizer. The preset loss function is a mathematical expression that measures the degree of difference between the network output and the labeled samples. It can be a mean squared error function or a smoothed average absolute error function. The training process improves the mapping accuracy of the network by continuously reducing this difference. Offline training refers to the process where, before the nonlinear mapping network is used online, the network parameters are trained all at once or in batches using all the sample data pre-generated and labeled in S31 and S32 until the loss function value converges to a preset small level or tends to stabilize. At this point, the network parameters are fixed, forming a nonlinear mapping network capable of mapping from multi-dimensional time-series power feature maps to key state indicators. An encoder-decoder network is a network structure where the encoder compresses the input into an intermediate representation, and the decoder restores this intermediate representation to the desired output.

[0064] The three steps of the above training process constitute a complete and tightly connected sample generation and network learning chain, which solves the key problems faced by isolated network stability domain mapping networks in the training stage: how to obtain a sufficient number of high-quality training samples with accurate labels, and how to use these samples to enable the network to effectively approximate the complex mapping relationship between source-load-storage power fluctuations and stability domain boundaries.

[0065] First, S31 and S32 work together to construct the training dataset from the problem domain. In isolated network operation, the actual samples that can be collected are often concentrated within the normal operating range, while samples of critical instability scenarios are extremely scarce. Directly using historical data for training makes it difficult for the network to learn the ability to accurately characterize the boundaries of the stability domain. S31 generates a massive number of operating scenarios by randomly combining them within preset fluctuation ranges of various power variables, covering a wide area of ​​the power space from the source. This includes not only the normal operating range but also conditions near or even beyond the critical boundary, ensuring the spatial distribution of training samples. S32 then performs time-domain simulation and automatic labeling on each operating scenario generated by S31. Through iterative search by the stability criterion analysis module, it obtains accurate key state indicators of the stability domain boundaries, thereby accurately labeling the expected output value for each scenario. The effect of the two-step collaboration is that S31 provides a sufficient number of scenario inputs with diverse operating conditions, while S32 provides accurate labels with physical basis for these inputs. Together, they form a labeled sample set that is large in scale, reliable in labeling, and comprehensive in coverage, effectively making up for the lack of critical samples in actual historical data.

[0066] Next, S33 utilizes the labeled sample set co-produced by S31 and S32 for supervised offline training. In each training iteration, the encoder-decoder network receives a multi-dimensional temporal power feature map as input. The encoder extracts and weights the temporal features through a multi-head attention mechanism, then the decoder regresses and maps them to predicted key state indicators. Gradient descent is used to continuously compare the predicted values ​​with the labeled values ​​in the labeled samples and update the network parameters accordingly. Since the labeled samples generated by S31 and S32 cover a wide range of operating conditions, from safe to critical to unstable, the network can repeatedly learn the correspondence between power fluctuation patterns and stability boundaries under various operating conditions during training, gradually establishing a generalized approximation ability for the mapping relationship between source-load-storage power changes and stability domain boundaries. Once the loss function converges, the parameters learned by the network implicitly encode the nonlinear mapping from power temporal features to stability domain boundaries. This allows the network, when deployed online, to quickly output the corresponding key state indicators with the accuracy validated during training, even when faced with new power fluctuation patterns that were not previously trained.

[0067] In some embodiments of this application, during the offline training process of S33, the encoder calculates a weighted fused power state feature vector through a multi-head attention mechanism, specifically including: S331: Perform a linear transformation on the input multidimensional time-series power feature map to generate a query matrix, key matrix, and value matrix corresponding to multiple attention heads; S332: For each attention head, calculate the dot product of the query matrix and the key matrix and scale it to obtain an attention score matrix that reflects the degree of correlation between power features at different time steps; S333: Use the Softmax function to convert the attention score matrix into an attention weight matrix, and then perform a weighted summation of the attention weight matrix and the value matrix to obtain the output features of the attention head; S334: Concatenate the output features of all attention heads and fuse them through a fully connected layer to generate a weighted fused power state feature vector.

[0068] In this application, in S331, the linear transformation refers to performing matrix multiplication on the input multidimensional temporal power feature map using a set of learnable weight matrices, projecting it from the original feature space to three different representation spaces used for querying, key-value matching, and feature extraction. Corresponding to multiple attention heads, the linear transformation does not generate a set of query, key, and value matrices, but rather multiple sets of parallel matrix triples, each corresponding to an independent attention operation channel, i.e., an attention head. The query matrix can be understood as the question representation actively proposed by the network in the current attention head, the key matrix as the index representation of the input features used for matching in the current attention head, and the value matrix as the representation of the actual information content contained in the input features in the current attention head. For example, if the number of attention heads is set to 8, then 8 sets of independent linear transformations are performed, each generating a query, key, and value matrix with appropriately reduced dimensionality. Different attention heads project the input features into different representation subspaces through their own independent weight parameters.

[0069] In S332, the dot product of the query matrix and the key matrix involves performing an inner product operation between each row of the query matrix and the corresponding column of the key matrix. The result reflects the original correlation strength between the time step features represented by the query matrix and the time step features represented by the key matrix. Scaling involves dividing the dot product result by a scaling factor, typically the square root of the key matrix dimension or the attention head dimension. This adjusts the dot product value to a suitable range for subsequent Softmax operations, preventing gradient vanishing due to excessively large dot product values. The attention score matrix is ​​a scaled two-dimensional numerical matrix. Each row corresponds to a time step of interest, and each column corresponds to a reference time step. The magnitude of the matrix elements reflects the degree of correlation between the power features of the corresponding two time steps. This score matrix can capture dependencies across time steps, such as whether there is a causal or correlated effect between two power fluctuations separated by hundreds of milliseconds.

[0070] In S333, the Softmax function is an exponential normalization operation that transforms any real-valued vector into a probability distribution vector. It is performed along the rows of the attention score matrix, ensuring that the values ​​in each row are non-negative and sum to 1. The attention weight matrix is ​​the result after Softmax normalization; its elements can be understood as the probability of information dependence of a given time step on other time steps under the current attention head. Values ​​closer to 1 indicate a higher level of attention given to the power features of that time step during fusion. The value matrix carries the actual information content of the power features at each time step. Weighted summation uses the attention weight matrix as weighting coefficients to perform row-by-row weighted summation on the value matrix, giving a larger contribution share to the value features of time steps closely related to the current time step, while suppressing the value features of sparsely related time steps, ultimately yielding the output features of the attention head.

[0071] In S334, concatenation refers to linking the output features calculated independently by each attention head along the feature dimension to form a long feature vector containing temporal correlation information from different aspects captured by all attention heads. A fully connected layer is a feedforward network layer with one or more interconnected neurons. It mixes and reorganizes the concatenated input features using learned weights, achieving cross-attention head information interaction and integration. The weighted fusion power-state feature vector is a compact feature representation output after processing by the fully connected layer. It integrates the local temporal correlation patterns extracted from multiple representation subspaces by the multi-head attention mechanism into a unified feature representation for subsequent mapping by the decoder.

[0072] The four sub-steps described above collectively define the complete computational path of the multi-head attention mechanism in the encoder, from the original feature map to the fused feature vector. Each step exhibits a progressive and functionally differentiated collaborative relationship. Functionally, S331 undertakes the tasks of feature decoupling and multi-angle projection. Through multiple independent linear transformations, the same input feature map is simultaneously projected to multiple different representation subspaces, each corresponding to an attention head, enabling subsequent capture of temporal dependencies from different perspectives. With only a single attention head, the network might focus only on a dominant temporal pattern, ignoring other equally influential temporal associations with different performance patterns. The multi-head design allows the network to distribute risk in parallel channels, each specializing in learning association patterns between different time scales or different power levels. S332 and S333, within each attention head, inherit the output of S331, collaboratively completing the computation from the original association metric to probabilistic attention allocation. S332 generates the attention score matrix through dot product calculation, achieving a quantitative expression of association strength across time steps, while the scaling operation creates numerical conditions for the smooth operation of the Softmax function in S333. S333 utilizes the normalization property of Softmax to transform the score matrix into a probability distribution weight matrix. This transformation makes the weight allocation competitive and selective, allowing the model to automatically focus on time segments with real information value while suppressing interference from irrelevant or redundant time steps. There is a direct connection between the score matrix generated by S332 and the weight matrix generated by S333. The former provides a raw measure of the degree of correlation, while the latter, based on the former, imparts probabilistic semantics to attention allocation through normalization.

[0073] After single-head computation is complete, S334 integrates the outputs of all heads. Different attention heads may have focused on different temporal dependency patterns in their preceding steps. For example, some heads tend to capture short-range dependencies between adjacent time steps, while others may focus on long-range dependencies between distant but causally related time steps. The concatenation operation preserves the pattern information captured independently by each head, while the learning capability of the fully connected layer allows different patterns to interact and recombine, producing richer and more expressive fused features than single-head features. The final output power-state feature vector thus integrates all the temporal correlation information mined by the multi-head parallel processing, providing the decoder with a high-information-density and high-fidelity input representation.

[0074] Overall, S331 achieves diversified decomposition of the feature space through multi-head linear projection, S332 and S333 implement adaptive time-step attention focusing within each subspace, and S334 achieves the convergence and fusion of information from multiple subspaces. This progressive computational architecture enables the encoder to automatically identify and enhance time segments and power types that significantly influence the stable domain mapping in the multi-dimensional temporal power feature map in a data-driven manner, while weakening redundancy and interference components, providing a high-quality feature supply for the subsequent high-precision regression mapping of the decoder.

[0075] In some embodiments of this application, the N key state indicators output in step S4 to describe the boundary of the stable domain are specifically as follows: The decoder of the nonlinear mapping network outputs the coordinates of M vertices. The coordinates of each vertex are composed of the values ​​of N key state indicators, including the critical cut-off time, the lowest frequency, the lowest voltage, and the damping ratio of the dominant oscillation mode. The M vertices are used to jointly describe the boundary of the stability region. Here, N=4, and M is an integer greater than N.

[0076] In this application, the decoder is a back-end component of the nonlinear mapping network, receiving the power state feature vector output by the encoder. Internally, it performs regression mapping on this feature vector through a multi-layer fully connected neural network. In this step, the decoder output is configured as the coordinates of M vertices, which are used together to describe the position and shape of the stability domain boundary in the state space. M is an integer greater than N. When N is 4, M can be 8, 12, or 20, etc. Its specific value can be preset before network training according to the required level of detail in characterizing the stability domain.

[0077] The coordinates of each vertex are composed of the values ​​of N key state indicators arranged sequentially. N is 4, and these four indicators are the critical clearing time, the lowest frequency point, the lowest voltage point, and the damping ratio of the dominant oscillation mode. The critical clearing time refers to the maximum duration a power system can withstand a pre-set short-circuit fault without losing synchronization stability; it is usually measured in milliseconds, and a larger value indicates a more ample power angle stability margin. The lowest frequency point refers to the minimum deviation of the system's frequency from its rated value during dynamic processes after experiencing a high-power disturbance; it is measured in Hertz and is a typical indicator of frequency stability margin. The lowest voltage point refers to the lowest amplitude to which the voltage of the system's key bus drops during dynamic recovery after a disturbance; it is usually expressed in per-unit values ​​and reflects the system's dynamic voltage support capability. The damping ratio of the dominant oscillation mode is a dimensionless parameter, referring to the damping ratio corresponding to the mode that has the greatest impact on dynamic quality among multiple oscillation modes induced by a small disturbance. A positive and large damping ratio indicates rapid oscillation decay; when it approaches zero or is negative, the system may experience continuous oscillation or even instability. These four indicators define the transient and dynamic stability boundaries of an isolated grid system under current operating conditions from four dimensions: power angle stability, frequency stability, voltage stability, and small-disturbance dynamic stability. It is understandable that the combination of these four indicators is not the only option; depending on the characteristics of the isolated grid system, other physical quantities that can quantitatively characterize the stability boundary can be replaced or added, such as the acceptable duration of transient voltage drops or the extreme value of the transient rate of change.

[0078] The M vertices collectively describe the boundary of the stability domain. This means that after locating these M vertices in the N-dimensional state space, they delineate the outline of the stability domain boundary from different orientations. Each vertex corresponds to a critical operating condition dominated by a different instability mode. For example, some vertices may be dominated by power angle stability constraints, with relatively small critical cutoff times and relatively lenient other indicators; other vertices may be dominated by voltage stability constraints, with relatively low voltage minimum values. By combining the M vertices distributed in different orientations of the boundary, the continuous hypersurface shape of the true stability domain boundary can be approximated as a discrete point set.

[0079] The design of the decoder outputting the coordinates of M vertices directly connects to the subsequent step of generating the dynamic stable region using a convex hull construction algorithm. When N is 4, if the decoder only outputs a single set of coordinates consisting of four index values, this point is merely an isolated location marker in four-dimensional space, unable to define a closed operating range with an internal region. By expanding the output to M vertices, where M is greater than 4, the decoder provides a sufficient number of boundary points, ensuring that at least a non-degenerate convex polyhedron can be formed in four-dimensional space. These M vertices each correspond to the critical states under different instability paths, defining the location of the stability boundary from different combinations of the four dimensions of power angle, frequency, voltage, and oscillation damping, providing the necessary geometric material for subsequent convex hull construction.

[0080] From the perspective of the connection between feature extraction, boundary description, and geometric domain construction, the decoder internally transforms the weighted fusion features extracted from the temporal power data by the encoder into boundary constraint points distributed in different stable dimensions through regression mapping. The distribution of these vertices in the state space implicitly reflects the constraint relationships between various instability modes under the current operating conditions. When these vertices are passed to the subsequent convex hull construction algorithm, the algorithm encloses all vertices within a minimal convex polyhedron formed by connecting the vertices, making the interior region of this polyhedron a computationally stable operating space. The boundaries of this space in each dimension are precisely defined by the corresponding critical vertices, thus realizing a complete mapping chain from multidimensional temporal power data weighted feature extraction via an attention mechanism to a closed stable domain in N-dimensional geometric space. Throughout the process, the decoder output acts as the interface between the abstract features of the network model and the concrete objects of computational geometry. The M vertices are both a complete description of the stable boundary states under the current operating conditions and a direct basis for constructing a dynamic stable domain with quantifiable safety margins.

[0081] In some embodiments of this application, the dynamic stable region is generated in S5 using a convex hull construction algorithm. Specifically, in an N-dimensional state space, the values ​​of N key state indicators are determined as vertex coordinates. All vertices are connected by a fast convex hull algorithm to form a minimum convex polyhedron. The internal region of this convex polyhedron is the dynamic stable region that ensures the transient stability of the system.

[0082] In this application, the N-dimensional state space is an abstract mathematical space spanned by N key state indicators as mutually orthogonal coordinate axes. In this step, N is set to 4, and the four coordinate axes correspond to the critical cut-off time, the lowest frequency point, the lowest voltage point, and the damping ratio of the dominant oscillation mode, respectively. Any point in this space uniquely corresponds to a specific transient stable state of the system. The distance and position in the space have clear physical meanings, that is, the displacement in different coordinate axis directions represents the state change of the system in different instability mode dimensions. Determining the values ​​of the key state indicators as vertex coordinates means that the values ​​of each indicator obtained in the previous steps are directly used as the coordinate values ​​of each coordinate axis in the state space without transformation or dimensionality reduction, thereby uniquely determining the position of a point in the space. For example, if the critical cut-off time obtained in a certain evaluation is 150 milliseconds, the lowest frequency point is 49.2 Hz, the lowest voltage point is 0.82 per unit, and the damping ratio of the dominant oscillation mode is 0.05, then these values ​​directly constitute a coordinate point in the four-dimensional space, which is used as a vertex of the convex hull construction.

[0083] The fast convex hull algorithm is a classic algorithm in computational geometry used to solve for the convex hull from a given set of points. This algorithm recursively approaches the convex hull boundary by finding the point farthest from the current hyperplane, and its time complexity is superior to traditional incremental algorithms in most application scenarios. Compared to envelope methods that require pre-defined shape parameters, the convex hull algorithm does not require prior assumptions about the geometry of the stable domain boundary; instead, it automatically generates the most compact envelope surface based on the positional distribution of the data itself. It is understood that the fast convex hull algorithm in this application is implemented using existing technologies. Besides the fast convex hull algorithm, other convex hull construction methods that can achieve the same function can also be used, such as incremental convex hull algorithms or Gift Wrapping algorithms.

[0084] A minimal convex polyhedron is a convex geometric solid that contains all vertices of its vertex set and has the smallest volume. In four-dimensional space, this polyhedron is enclosed by multiple three-dimensional hyperplane facets, each facet being a basic unit constituting the polyhedron's surface. The interior region of the convex polyhedron is considered a dynamic stability domain that ensures the transient stability of the system. This means that when a point representing the system's operating state falls inside the polyhedron, the system is within the critical value in all current stability dimensions; when the operating point falls on the polyhedron's surface or outside, it indicates that at least one stability dimension has reached or exceeded the stability boundary. The term "dynamic" in "dynamic stability domain" emphasizes that the polyhedron is not fixed but updates in real time with changes in the real-time power of the source load storage and the addition of new boundary vertex samples. Its shape and volume may change under different operating conditions.

[0085] The core function of step S5 is the construction of a closed geometric domain from discrete vertices. This transformation relies on the coordinated operation of vertex localization and the convex hull algorithm. The key state index values ​​output from the preceding steps are first identified as vertices with definite coordinates in the N-dimensional state space. Each of these vertices corresponds to a critical instability condition and is distributed at different locations on the boundary of the stable domain. After receiving these vertices, the fast convex hull algorithm calculates the spatial relationships between them, filters out vertices located on the outer envelope, and eliminates redundant points that might fall inside, thereby generating a minimal convex polyhedron enclosed by multiple hyperplane patches. Each vertex provides constraint information of the stable boundary in various directions in space, while the convex hull algorithm integrates these discrete directional constraints into a continuous closed geometric object. This division of labor between the two ensures that the transformation from discrete points to a closed domain retains the precise localization capability of vertices to the boundary while also providing a geometric basis for determining the attribution of any position within the domain.

[0086] Step S5 plays an intermediate role in the overall evaluation chain, transforming the index values ​​into a geometric domain. The key state index values ​​output by step S4 are abstract numbers and do not yet possess the functions of spatial enclosure and attribution judgment. Step S5 completes the transformation from numerical points to a continuous geometric domain by locating these values ​​as vertices in the state space and constructing a minimum convex polyhedron using the convex hull algorithm. This transformation provides a directly operable geometric object for calculating the shortest distance from the current running point to each boundary surface of the dynamic stable domain in subsequent step S6. Without this transformation step, step S6 would lack a clear and calculable boundary reference surface, and the quantitative evaluation of the real-time stability margin index would be impossible. At the same time, since the convex hull algorithm generates a minimum convex polyhedron without pre-setting the boundary shape, this polyhedron can closely approximate the complex hypersurface shape of the real stable domain boundary in the state space, providing a reasonable geometric reference for subsequent margin evaluation.

[0087] In some embodiments of this application, calculating the shortest distance from the current running point to each boundary surface of the dynamic stable domain in S6 includes: calculating the shortest Euclidean distance from the current running point to each boundary surface of the dynamic stable domain using a point-to-polyhedron distance algorithm, and taking the minimum value among them as the shortest distance.

[0088] In S6, the point-to-polyhedron distance algorithm is a spatial distance calculation method in computational geometry used to determine the perpendicular distances from a given point in space to the constituent faces of a convex polyhedron. This algorithm processes each boundary face of the convex polyhedron one by one. Each boundary face is typically represented by its vertex sequence or by the coefficients of its spatial plane equation. The algorithm determines the effective Euclidean distance from the point to the plane by solving for the perpendicular projection distance from the point to the plane and further determining whether the projected point falls within the effective region of the boundary face. The Euclidean distance in the shortest Euclidean distance refers to the straight-line distance between two points in N-dimensional state space. When N is two-dimensional, it is the straight-line distance between two points on a plane; when N is three-dimensional, it is the straight-line distance between two points in solid space; and so on for higher dimensions. It is calculated by taking the square root of the sum of the squares of the differences in coordinates across all dimensions. Understandably, the shortest Euclidean distance from the current operating point to a certain boundary reflects the state-space displacement the system needs to undergo to reach the critical boundary from its current state along the most dangerous direction, considering all stable dimensions. The shortest distance refers to the minimum calculated result of the shortest Euclidean distances across all boundary surfaces in the dynamic stable domain; the boundary surface corresponding to this minimum value is the closest stable boundary to be breached in the current operating state. The significance of taking the minimum value instead of other statistics lies in the fact that the system's stability margin depends on the weakest link in the dimension closest to the boundary, rather than the average distance across dimensions. This is similar to the principle of the "barrel effect." For example, if the current operating point has a sufficient distance to one boundary surface but a very small distance to another, the system is likely to become unstable first in the stable dimension corresponding to the latter. Taking the minimum value can sensitively capture this risk.

[0089] The elements within step S6, as well as the steps preceding and following S6, work closely together to achieve the transformation from the stability domain geometry to the quantitative margin index.

[0090] From the perspective of the coordination of various computational elements within S6, the point-to-polyhedron distance algorithm, the boundary faces of the convex polyhedron, and the current running point constitute the relationship between the computational tool, the computational object, and the evaluated object. The point-to-polyhedron distance algorithm requires the geometric definition of the boundary faces as input, which is explicitly provided by the polyhedron generated by the convex hull construction algorithm in step S5. Specifically, this can be a list of vertex coordinates for each face or the surface equation coefficients calculated from the vertices. The current running point serves as another input, obtained through coordinate mapping from the power state feature vector. After calculating the distance to each boundary face, the algorithm's minimum value operation essentially performs a risk profile in a multidimensional stable space, identifying the weakest point in the current stable dimension. The point-to-polyhedron distance calculation itself is a geometric metric; the minimum value operation imbues this geometric metric with the semantics of safety assessment, giving a purely spatial distance an engineering meaning of stability margin.

[0091] From the perspective of S6's role in the overall process, S6 builds upon the dynamic stability domain construction of S5 and connects to the early warning judgment of S7. The dynamic stability domain output by S5 is a geometric object represented by a convex polyhedron. This geometric object has clear boundary interfaces between its interior and exterior, but the geometry alone cannot determine the distance between the current running point and these boundaries. S6 uses distance calculation to associate the geometric domain of S5 with the running point representing the current state, transforming the boundaries of the geometric domain into quantifiable distance values. Subsequently, in S7, the early warning judgment requires a dimensionless margin index as a comparison object. S6 directly satisfies this requirement by normalizing the shortest distance and outputting a real-time stability margin index, enabling S7 to trigger control signals through simple threshold comparison. The entire process, from geometric construction to distance measurement to threshold comparison, is seamlessly connected and progressively builds upon each step. S6 acts as a bridge between the geometric results and numerical judgments.

[0092] In some embodiments of this application, the method further includes: preprocessing the real-time power data sequence, the preprocessing including filtering out measurement noise using a Kalman filter and repairing missing data using linear interpolation to obtain a cleaned power data stream; and segmenting and stacking the cleaned power data stream using a sliding time window to construct a multi-dimensional time-series power feature map.

[0093] In this application, the Kalman filter is a recursive estimation algorithm based on a state-space model. It uses the system state equation and observation equation, employing a weighted recursive calculation of the previous time step's state estimate and the current time step's measurement value to arrive at the optimal estimate of the current state, thereby effectively suppressing Gaussian white noise mixed into the measurement channel. Measurement noise refers to the random error component introduced during power acquisition by the measurement unit due to factors such as sensor accuracy limitations, signal transmission interference, and environmental electromagnetic coupling. It typically manifests as high-frequency, small-amplitude fluctuations superimposed on the true power value. Using a Kalman filter to filter out measurement noise allows the filtered power value to more closely approximate the true power change trajectory of the system without introducing significant phase delay. Linear interpolation is a data repair method that linearly estimates the value at the missing time step based on the proportional relationship between the preceding and following valid data points at the location of missing data. Understandably, data loss may be caused by communication interruptions, temporary malfunctions of measurement units, or packet loss. Linear interpolation, while completing the data, maintains the continuity and trend consistency of the power sequence in the time domain. For example, if the load power data at a certain moment is lost, and the load power at the two moments before and after it is 100 kW and 110 kW respectively, the interpolated value will be approximately 105 kW. The cleaned power data stream refers to the complete and smooth real-time power data sequence obtained after Kalman filter noise reduction and linear interpolation completion processing, which eliminates the interference of measurement noise and missing points on subsequent feature extraction.

[0094] In the above data preprocessing steps, Kalman filtering and linear interpolation completion work together to address two common data quality issues in real-time power data sequences—measurement noise and missing data—and jointly output a high-quality cleaned power data stream, providing a reliable data foundation for the subsequent construction of multi-dimensional time-series power feature maps and network mapping.

[0095] From the complementary relationship between the two processing methods, Kalman filtering mainly addresses the problem of data existing but contaminated by noise. Its recursive filtering mechanism can remove noise components while maintaining the dynamic trend of power changes. Linear interpolation, on the other hand, addresses the problem of data not existing at all, reconstructing missing points based on adjacent valid data. In actual isolated network measurement environments, these two problems often coexist and intertwine. Noise interference may cause the signal-to-noise ratio of individual sampling points to be extremely low, rendering them almost unusable, while communication interruptions directly cause data gaps. Performing filtering alone without filling in missing values ​​may introduce structural defects in subsequent sliding time window operations due to sequence discontinuity; performing interpolation alone without filtering will still allow noise at interpolation points or valid points to propagate and be solidified in the feature map construction stage. The cleaned data stream after the two processing methods are combined possesses both integrity and smoothness, ensuring that the data extracted in each time segment of the subsequent sliding time window segmentation operation has a consistent quality level.

[0096] From the perspective of its role in connecting with subsequent processes, the high-quality cleaned power data stream directly affects the weight allocation effect of the multi-head attention mechanism in the subsequent S3 step. If the data contains significant noise or unpatched missing points, the attention mechanism may incorrectly identify noise impulses as important power fluctuation features, thus allocating attention weights to time steps with no actual physical meaning, interfering with the accuracy of the weighted fused power state feature vector. Through the pre-processing of Kalman filtering and linear interpolation, the cleaned data stream can more faithfully reflect the real dynamic changes in the source-load-storage power of the isolated network system, enabling the attention mechanism to learn time-step correlations based on real physical fluctuation patterns, providing a fundamental guarantee for the accuracy of subsequent network mapping.

[0097] In some embodiments of this application, after S7, a step S8 is further included, which includes: S81: Encapsulate the real-time stability margin index and its corresponding source-end power generation, load-end power, energy storage charging and discharging power, and dynamic stability domain boundary information into a historical operating state data packet; S82: Periodically feeds back historical running data packets to the training database, and when the preset model update triggering conditions are met, uses an incremental learning mechanism to fine-tune the parameters of the nonlinear mapping network online.

[0098] In this application, in S81, the historical operational data packet is a record unit formed by structurally encapsulating all inputs, outputs, and intermediate results involved in a single online evaluation process. This data packet contains four items: the real-time stability margin index is the normalized margin quantification result output in step S6; the source-end power generation, load-end power, and storage-end charging / discharging power at the corresponding time are real-time power data collected and preprocessed in step S1; and the dynamic stability domain boundary information is the geometric description data of the convex polyhedron generated in step S5, which, exemplarily, can be the vertex coordinate sequence or surface equation coefficient set of each boundary face constituting the convex polyhedron. Encapsulating the above information into a whole aims to completely preserve the correspondence between the system state, stability domain morphology, and safety margin at that operational moment, providing high-quality samples with self-labeled attributes for subsequent model updates. In S82, periodic feedback refers to the process of transmitting the historical operational data packets accumulated during this period to the training database in batches according to a preset time interval, such as hourly, daily, or weekly. The training database is a data warehouse used to store samples required for offline training and online fine-tuning. Its initial content consists of labeled samples generated by S31 and S32. The preset model update trigger conditions refer to the predetermined rules required to initiate model fine-tuning operations. These can be triggered by the number of newly added historical running data packets in the training database reaching a preset threshold, or by detecting a continuous upward trend in the recent mapping error statistics of the nonlinear mapping network. The incremental learning mechanism is a model update strategy. Unlike full training from scratch, it uses only newly added historical running data packets to make small, continuous adjustments to the network parameters based on the existing network parameters, allowing the model to gradually adapt to new data distributions while maintaining its learned knowledge. Online fine-tuning refers to updating the nonlinear mapping network parameters deployed in the stable domain intelligent mapping engine module in the background without interrupting the online evaluation service. The updated network parameters replace the original parameters and continue to perform online mapping tasks.

[0099] Step S8 consists of two sub-steps that form a complete closed-loop collaboration from online operation to offline update and back to online operation. S81 is responsible for data encapsulation and record storage, while S82 is responsible for condition judgment and model update. S81 packages the complete information of each online evaluation into structured records, which accumulate continuously over time, gradually covering the real operating conditions of the isolated network system under different seasons and operating modes. S82 periodically reads the data packets accumulated by S81 and imports them into the training database, initiating incremental learning when the update trigger condition is met. The cooperation between the two lies in the fact that the continuous encapsulation of S81 provides a continuous supply of real operating samples for the model update of S82, while the update trigger condition of S82 provides reasonable constraints on the timing of model updates, avoiding frequent and meaningless updates that consume computational resources. The completeness of the encapsulated content ensures that each data packet simultaneously contains input features, output labels, and intermediate geometric information, which is equivalent to providing self-labeled training samples that can be used directly without manual annotation for incremental learning, reducing the dependence of model updates on simulation resources. In this application, embodiments S1 to S7 constitute a forward inference link from data acquisition to online evaluation, which relies on the parameter quality of the nonlinear mapping network. However, in actual operation, the topology, component parameters, or statistical characteristics of source load power in an isolated network system may change slowly over time, causing a deviation between the mapping rules learned during offline training and the current actual system. If not corrected, the mapping accuracy will gradually decline. S8, as a feedback loop, feeds back the real samples accumulated during online operation to the training stage, using an incremental learning mechanism to enable the network parameters to adjust synchronously with the drift of system characteristics, thereby connecting the forward inference link and the feedback learning link into a closed loop. This closed-loop collaboration enables the nonlinear mapping network to have the ability to autonomously adapt to the slow changes in the operating characteristics of the isolated network system, helping to maintain the stability of mapping accuracy in long-term operation.

[0100] In some embodiments of this application, it further includes: S51: Obtain the future scheduling plan of the isolated grid system, which includes the predicted power generation at the source end, the predicted power load at the load end, and the planned power charging and discharging at the storage end within a preset future time period; S52: Input the power prediction curve corresponding to the scheduling plan into the nonlinear mapping network to predict the predicted stability region corresponding to each time section within the future preset period. S53: Based on the predicted stability domain, assess whether there will be a risk period in the future where the real-time stability margin index is lower than the warning threshold; S54: If there is a risk period, calculate the minimum energy storage charging and discharging power adjustment required to make the system's stability margin index reach above the warning threshold during the risk period, and generate a forward-looking energy storage pre-scheduling instruction accordingly, which is then fed back to the isolated grid energy management system.

[0101] In this application, in S51, the dispatch plan refers to the operational arrangements pre-prepared by the isolated grid energy management system based on load forecasting, renewable energy output forecasting, and economic dispatch strategies, and prepared for execution or reference in future periods. The source-side power generation forecast is a predicted value of future output from intermittent power sources such as photovoltaics and wind power based on weather forecasts and historical output characteristics, typically carrying a certain degree of uncertainty. The load-side power generation forecast is a prediction of future load demand based on historical load curves, meteorological factors, and electricity consumption patterns. The energy storage charging and discharging plan power is the pre-arranged charging and discharging power command value of the energy storage system in various future periods within the dispatch plan; positive values ​​indicate discharging, and negative values ​​indicate charging. The future preset period can be understood as a time range looking forward from the current moment; for example, it could be 15 minutes, 1 hour, or 4 hours, with the specific duration set according to the needs of the isolated grid dispatch cycle. In S52, the power forecast curve is a sequence formed by arranging the values ​​of the three types of power in the dispatch plan obtained in S51 at different future time segments in chronological order. Each time segment corresponds to a three-dimensional power vector containing the source-side power generation forecast, the load-side power generation forecast, and the energy storage charging and discharging plan power. The predicted stability domain is the dynamic stability domain corresponding to each future time segment, output by a nonlinear mapping network after the power prediction curve undergoes the same data preprocessing, feature map construction, and network mapping process as the real-time online assessment. Its construction method is consistent with step S5. It is understood that the predicted stability domain is not the currently existing stability domain, but rather the model's deduction of the future stability boundary shape based on the predicted power input. In S53, the risk period refers to one or more continuous or discrete time intervals within a preset future period where the predicted real-time stability margin index is lower than the warning threshold; the system is predicted to have an instability risk during this period. In S54, the minimum energy storage charging and discharging power adjustment amount refers to the minimum adjustment range required to adjust the planned charging and discharging power of the energy storage end, under the constraint of raising the stability margin index within the risk period to above the warning threshold. This adjustment amount is obtained by solving an optimization problem. The forward-looking energy storage pre-scheduling instruction refers to the corrected charging and discharging power instruction issued to the energy storage system in advance before the risk period arrives. Its content includes the adjusted charging or discharging power value and its execution time. For example, if a risk period of 10 minutes is predicted to occur in the next 30 minutes, and it is calculated that the energy storage discharge power needs to be increased by 50 kilowatts from the original plan to meet the margin requirements, a forward-looking instruction will be issued to the energy storage system at the present moment, instructing it to discharge at the adjusted power value between 30 minutes and 40 minutes later.

[0102] The above four sub-steps form a complete forward link of prediction-mapping-evaluation-regulation in the time dimension, which extends the stability domain assessment from real-time monitoring to the field of forward-looking management, and the parts are closely connected.

[0103] S51 and S52 collaboratively complete the projection and prediction of future stability domain forms from future planning information. S51 obtains time-series information on various power types for future periods from the energy management system. Although this information is not measured in real time, it reflects the planned operating trajectory of the system in the future. S52 inputs this series of planning information into a nonlinear mapping network whose mapping accuracy has been verified in real-time operation, and generates the predicted stability domains for each future time segment through the same calculation process. The significance of these two steps is that S51 provides a power trajectory input that fully covers the future period, while S52 uses the learned power-stability domain mapping relationship to project each future time segment one by one, enabling operators or automated systems to foresee the future evolution trend of the stability domain under the planned power trajectory. S53 and S54 then collaboratively complete risk assessment and proactive intervention decisions based on the prediction results. S53 performs margin calculations and threshold comparisons on a segment-by-segment basis on the predicted stability domain sequence output by S52 to identify specific periods where instability risks may exist in the future. If the assessment result of S53 indicates the existence of a risk period, then S54 uses the identified risk period as a constraint and the energy storage power adjustment amount as the optimization variable to calculate the minimum adjustment amount required to eliminate the risk and generate a forward-looking instruction. S53 outputs a risk existence judgment, providing a decision-making basis for the initiation of S54. The solution process of S54 further utilizes the margin gap information in this judgment to accurately calculate the adjustment amount, avoiding blind or over-adjustment. From the overall process perspective, these four sub-steps work together to complete a closed loop from predictive perception to predictive control, advancing the timing of control intervention from after the instability risk actually occurs to when the risk is still in the prediction stage, providing the isolated grid system with more ample response preparation time.

[0104] In some embodiments of this application, step S6 calculates the shortest distance from the current running point to each boundary surface of the dynamic stable domain and normalizes it to obtain the real-time stability margin index, specifically including: S61: In the N-dimensional state space, determine the coordinates of the current running point and identify the set of vertices of the boundary surface closest to the current running point; S62: Construct the spatial plane equation of the boundary surface based on the vertex set, and obtain the shortest distance by calculating the perpendicular distance from the current running point to the spatial plane; S63: Calculate the ratio of the shortest distance to the preset baseline distance value to obtain the normalized stability margin index; S64: Based on the stability margin index, generate the stability margin gradient direction vector of the current running point in the N-dimensional state space. The gradient direction vector points to the direction of the fastest growth of the stability margin, which is used to provide adjustment direction guidance for subsequent prevention and control decisions.

[0105] In this application, S61, determining the coordinate position of the current running point in the N-dimensional state space means converting the weighted fused power state feature vector obtained in step S3 into a point with definite coordinate values ​​in the state space spanned by N key state indicators through a mapping relationship before the decoder output layer. Each dimension of the coordinate value is consistent with the dimension and scale of the corresponding key state indicator. Identifying the vertex set of the boundary face closest to the current running point means finding all vertices of the boundary face with the shortest spatial distance to the current running point among all the boundary faces of the convex polyhedron constituting the dynamic stable domain, through preliminary distance estimation or spatial partitioning indexing. It can be understood that each boundary face of the convex polyhedron is uniquely determined by its vertex set; identifying this vertex set provides the geometric elements for subsequently constructing the plane equation. In S62, the spatial plane equation refers to a mathematical expression describing the position of a hyperplane in N-dimensional space. Its form can be a linear combination of coordinates in each dimension equal to a constant, with coefficients obtained by solving the vertex set of the boundary face through normal vectors. For example, in three-dimensional space, this equation is a planar equation; in spaces higher than three dimensions, it represents a hyperplane equation, whose coefficients can be obtained through singular value decomposition of the vertex coordinate matrix or by solving a system of linear equations. The vertical distance refers to the normal projection length from the current running point to the plane of this space, i.e., the shortest distance from the point to the hyperplane represented by the planar equation. It is calculated by substituting the coordinates of the current running point into the normalized expression of the planar equation. In S63, the preset reference distance value is a pre-defined reference distance. Its value can be determined based on the typical range of distances from stable running points to the boundary under various running scenarios obtained from offline simulation statistics, or it can be taken as the equivalent radius or average scale of the dynamic stability domain in multi-dimensional space. The normalized stability margin exponent after ratio calculation eliminates the absolute dimension of the distance, making the margins under different operating conditions comparable. In S64, the stability margin gradient direction vector refers to the vector pointing from the current running point to the direction of the fastest growth of the stability margin exponent in the N-dimensional state space. Understandably, this gradient direction can be obtained by solving for the partial derivatives of the stability margin index with respect to each key state index. Geometrically, it is usually perpendicular to the boundary surface closest to the current operating point and points towards the interior of the convex polyhedron. The physical meaning of this gradient direction vector is that if the key state indices are adjusted along this vector direction, the system will move away from the critical boundary at the fastest speed. This provides a directional basis for the allocation of adjustment amounts among controllable devices when formulating subsequent preventive control strategies.

[0106] The four sub-steps S61 to S64 proceed progressively along three processing levels: distance quantization, normalization, and direction guidance, gradually enriching and deepening the transformation process from stable domain geometry to usable decision information.

[0107] S61 and S62 work together to calculate the distance from the geometry. S61 locates the current operating point and locks the nearest boundary surface, essentially performing a risk ranking among the faces of the polyhedron and identifying the stable dimensional direction that poses the most pressing threat to the current working condition. S62 then takes the results identified by S61 and converts the spatial distance into an analytically calculable vertical distance by constructing a precise plane equation for the boundary surface. The combination of these two technologies ensures that the distance calculation is both directionally targeted, using the distance to the most dangerous boundary surface as the basis for the margin metric, rather than calculating the distance to the centroid or average surface of the polyhedron in general terms, and computationally feasible, with clear and precise calculation of the vertical distance between a point and a surface, stably outputting a unique shortest distance value.

[0108] S63 normalizes the shortest distance output by S62. The overall scale and shape of the dynamic stability domain may differ under different operating conditions. Directly using absolute distance values ​​is not comparable across different conditions and makes it difficult to establish a unified risk criterion. By introducing a preset benchmark distance value for ratio calculation, S63 transforms the absolute distance into a normalized stability margin index, ensuring a consistent evaluation standard for assessment results at different times and under different operating conditions. This also facilitates the unified setting and automatic comparison of warning thresholds in step S7.

[0109] S64 further generates directional guidance information based on the evaluation results of the first three steps. While the first three steps answer the question of how safe the system is currently, S64 answers the question of in which direction to adjust to improve the safety margin. The generation of the gradient direction vector extends margin assessment from passive perception to proactive decision support, providing a clear direction for subsequent preventative control strategies and helping to improve the targeting and efficiency of control decisions. In summary, the four sub-steps progress step by step, transforming the geometric domain generated by S5 into a dual information output containing both quantitative margin values ​​and optimized adjustment directions, providing information support for S7's emergency warning and higher-level preventative control coordination.

[0110] In some embodiments of this application, the calculation in step S54 of the minimum energy storage charging and discharging power regulation required to keep the system stable during risky periods specifically includes: S541: Obtain the predicted stability domain corresponding to the risk period, and determine the most dangerous critical state indicator corresponding to the boundary closest to the predicted operating point in the predicted stability domain; S542: Using the sensitivity matrix of the most dangerous critical state index to the energy storage charging and discharging power as a constraint, and with the goal of minimizing the adjustment of the energy storage charging and discharging power, a linear programming model is constructed. S543: Solve the linear programming model to obtain the minimum energy storage charging and discharging power adjustment required to bring the most dangerous critical state index back to the predicted stability domain. S544: Based on the minimum energy storage charging and discharging power adjustment amount and the rated power and capacity constraints of the energy storage system, the feasibility of the forward-looking energy storage pre-dispatch command is verified. The pre-dispatch command is only fed back to the isolated grid energy management system when the adjustment amount is within the adjustable range of the energy storage system.

[0111] In this application, in S541, the predicted operating point refers to the coordinate point obtained by mapping the weighted fused power state feature vector output by the decoder to the N-dimensional state space after inputting the power prediction curve of a future time segment into the nonlinear mapping network in step S52. This point represents the expected stable state of the system at that future moment. The most dangerous critical state indicator refers to the critical state indicator corresponding to the boundary surface that is spatially closest to the predicted operating point among multiple boundary dimensions of the predicted stability domain. It reflects the stable dimension most likely to be breached first during future risk periods. For example, if the distance from the predicted operating point to the boundary surface corresponding to the lowest voltage point indicator is significantly smaller than the distance to other boundary surfaces, then the lowest voltage point is determined as the most dangerous critical state indicator. In S542, the sensitivity matrix refers to the matrix composed of coefficients describing the influence of the change in charging and discharging power of each energy storage device on the value of the most dangerous critical state indicator. Its elements can be understood as the magnitude and direction of the change in the most dangerous critical state indicator caused by a unit power adjustment of a certain energy storage device. The sensitivity matrix can be obtained by numerical perturbation calculation of the electromechanical transient simulation model of the power system near the predicted operating condition, or it can be approximated by local linearization of the nonlinear mapping network at that operating condition point. The linear programming model is a mathematical optimization model whose objective function is to minimize the sum or weighted sum of the absolute values ​​of the charging and discharging power adjustments of each energy storage device. Constraints include the requirement that the adjustment effect represented by the sensitivity matrix must achieve a preset margin to move the most dangerous critical state index within the predicted stability domain. In S543, solving the linear programming model refers to using linear programming algorithms such as the simplex method or interior-point method to calculate the values ​​of the charging and discharging power adjustments of each energy storage device that satisfy the constraints and minimize the objective function; this value is the required minimum adjustment. In S544, the rated power of the energy storage system refers to the maximum active power value that the energy storage converter is allowed to continuously output, and the capacity constraint refers to the remaining available electricity or charging space of the energy storage system at the current moment, usually defined by the upper and lower limits of the state of charge. Feasibility verification refers to comparing the minimum adjustment amount obtained from S543 with the rated power and capacity constraints of energy storage. If the adjustment amount simultaneously meets the conditions that the power does not exceed the limit and the capacity can support the required duration, the pre-scheduling instruction is considered feasible and sent. Otherwise, the instruction may need to be modified or an alarm may be issued to indicate that energy storage resources are insufficient.

[0112] The four sub-steps S541 to S544 work together to form a complete decision optimization chain from problem identification to solution finding and then to feasibility verification. The steps show a progressive relationship of gradual convergence and layer-by-layer verification.

[0113] S541 and S542 work together to transition from risk assessment to control modeling. Building upon the risk periods identified and the predicted stability domain determined in S53, S541 further pinpoints the root causes of the risk—specifically, which critical state indicator is closest to the boundary. This provides a clear target for subsequent control object selection. S542, utilizing the most dangerous critical state indicator identified by S541 and combining it with the system's sensitivity information under that operating condition, transforms the adjustment requirements for stability margin into a quantitative constraint relationship between specific energy storage power adjustment and changes in critical state indicators. This formalizes the control problem into a solvable mathematical optimization model. Together, they transform the qualitative results of the risk assessment into quantitative inputs for control decisions.

[0114] S543 and S544 collaborated to complete the process from optimization solution to engineering feasibility verification. S543 solved the linear programming model constructed by S542, deriving the mathematically optimal minimum adjustment scheme. However, this mathematical solution did not consider the physical limitations faced by the energy storage equipment in actual operation, and directly issuing it as a command might lead to execution failure or equipment damage. Based on the output of S543, S544 introduced rated power and capacity constraints for feasibility verification, acting as an engineering safety valve, issuing only feasible commands that passed the verification to the energy management system. The existence of this checkpoint ensures that the entire decision-making chain pursues both economical regulation and reliable execution, avoiding secondary disturbances to system operation caused by issuing infeasible commands. In summary, the four sub-steps, through a layer-by-layer locking approach, from identifying the most dangerous indicators to constructing a sensitivity constraint model, and then to optimization solution and physical constraint verification, form a complete decision-making chain from risk root cause location to feasible control command output.

[0115] In some embodiments of this application, the calculation process of the multi-head attention mechanism further includes: S335: Perform sparsity constraint processing on the attention weight matrix corresponding to each attention head. The sparsity constraint processing includes: setting an attention weight threshold and setting the weight values ​​below the attention weight threshold to zero in order to suppress interference from irrelevant time steps. S336: Based on the attention weight matrix after sparsification constraint processing, extract the source-load-storage power features corresponding to time steps with non-zero weight values, and construct a sparse temporal feature subset that is strongly correlated with the dominant dynamic characteristics of the isolated network. S337: Perform a residual connection between the sparse temporal feature subset and the attention head output features obtained in S333 to obtain enhanced attention head output features. Use the enhanced attention head output features as input for concatenation in S334 to enhance the nonlinear mapping network's ability to identify key time segments in long temporal dependencies.

[0116] In this application, S335, the sparsity constraint processing refers to a structural constraint operation applied to the attention weight matrix. This aims to ensure that only a few elements in the weight matrix corresponding to the time steps deemed important maintain non-zero values, while the remaining elements are forced to zero, thus forming a sparse weight distribution. The attention weight threshold is a preset numerical limit used to determine the importance of weights. Its specific value can be dynamically set according to the statistical distribution of the weight matrix, for example, taking a multiple or quantile of the mean of all elements in the weight matrix, or a fixed empirical value. Setting weight values ​​below the attention weight threshold to zero means that only those time steps with sufficiently high correlation contribute substantially to subsequent feature fusion, while time steps with weak correlation are directly excluded from the calculation. Understandably, this sparsity processing helps suppress interference from irrelevant time steps, which refer to sampling moments where the power features have low correlation with the current stability domain evaluation task or may even introduce noise. In S336, the sparse temporal feature subset refers to a subset of features extracted from the original multidimensional temporal power feature map based on the time steps corresponding to the non-zero weights retained after sparsification constraint processing. This subset only contains the power features of the time steps that the attention mechanism identifies as having a strong correlation with the dominant dynamic characteristics of the isolated network system. The dominant dynamic characteristics can be understood as the power fluctuation patterns that have a decisive influence on the boundary position of the system's stable domain, such as the power response features corresponding to the dominant mode of frequency or voltage oscillation after a large disturbance. In S337, residual connection refers to a network structure design that directly transmits the input signal across layers to the output and adds it to the transformed signal. Its mathematical implementation is to add the feature representation of the sparse temporal feature subset to the corresponding elements of the attention head output features calculated in S333. It can be understood that residual connection provides the network with a direct path to transmit the original feature information, allowing the original features of the key time steps selected by sparsification to bypass the weighted fusion transformation of the attention mechanism and be directly superimposed with the fused features. The enhanced attention head output feature is the feature representation after residual connection stacking. This feature simultaneously contains global temporal dependency information fused by attention weighting and the original detailed information of key time segments retained after sparse filtering.

[0117] The three sub-steps S335 to S337, based on the multi-head attention computation process constructed in the original S331 to S334, embed a parallel processing branch for sparsity filtering and residual enhancement, which works in conjunction with the original steps to strengthen the network's ability to identify key time segments in long-term dependencies.

[0118] From the perspective of the synergistic relationship within the three sub-steps, S335 acts as an information filtering gate by thresholding the attention weight matrix through sparsity constraint processing. The attention weight matrix, after Softmax normalization, may contain a large number of weights with small but not zero values. These weights often correspond to power features during stable or noisy periods in the dynamic process, contributing negligibly to the determination of the stable domain boundary, but potentially introducing background interference during accumulation. S335 removes these weights from weakly correlated time steps by setting a hard threshold to zero, making the attention distribution more concentrated on a few strongly correlated time steps. Following the filtering result of S335, S336 extracts the power features corresponding to the retained weights from the original feature map, forming a sparse temporal feature subset. This is equivalent to using the index filtered by S335 to perform selective feature extraction from the original data, preserving the original power information at strongly correlated time steps. S337 then performs a residual connection between the sparse feature subset extracted by S336 and the attention head output feature of S333. The synergistic value of this operation lies in the fact that the attention head output feature is a global fusion representation obtained by weighted summation, which may have a certain smoothing or dilution effect on the detailed information at different time steps during the fusion process; while the sparse temporal feature subset retains the original feature details at the time steps that are determined to be key. The residual connection allows these detailed information to be directly superimposed on the fusion feature as supplementary signals, making up for the information loss during the fusion process.

[0119] From the perspective of the overall synergy between this enhanced branch and the original multi-head attention process, the original process focuses on discovering and weighting the information from all time steps through global attention, while the newly added sparsification and residual connection branch focuses on strengthening the original feature representation of a few key moments based on global fusion. The two are not substitutes for each other but complement each other. Global attention provides a perception of overall temporal dependencies, while the sparse residual branch ensures that detailed information at key moments is not overwhelmed by global fusion. This synergy allows the enhanced attention head output features to retain the ability to grasp long temporal dependencies globally while exhibiting a more sensitive response to key dynamic events that determine the boundaries of the stable domain. This helps improve the mapping accuracy of nonlinear mapping networks under complex conditions with long input sequences and a large number of redundant time steps.

[0120] A second aspect of this application provides an islanded grid stability domain nonlinear mapping system based on source-load-storage real-time power, used to perform the above method. The system includes: The data acquisition module is configured to execute step S1. The feature map construction module is configured to execute step S2. The encoder module is configured to execute step S3; The decoder module is configured to perform step S4. The stability region generation module is configured to execute step S5. The margin assessment module is configured to execute step S6; The early warning feedback module is configured to execute step S7.

[0121] A third aspect of this application provides an electronic device, including: one or more processors, one or more input devices, one or more output devices, and one or more memories. The processors, input devices, output devices, and memories communicate with each other via a communication bus. The memories store computer programs, including program instructions. The processors execute the program instructions stored in the memories. The processors are configured to invoke the program instructions to execute the aforementioned nonlinear mapping method for the stable domain of an isolated grid based on real-time power of source-load-storage systems.

[0122] It should be understood that, in the embodiments of this application, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0123] Input devices may include touchpads, fingerprint sensors (for collecting the user's fingerprint information and fingerprint orientation information), microphones, etc., while output devices may include displays (LCDs, etc.), speakers, etc.

[0124] The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store information about the device type.

[0125] In specific implementations, the processor, input device, and output device described in the embodiments of this application can execute the implementation methods described in any embodiment of the nonlinear mapping method for the stable domain of an isolated grid based on real-time power of source-load storage provided in the embodiments of this application, or can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0126] In another embodiment of this application, an electronic device is provided. The electronic device stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the above-described method for nonlinear mapping of the stable domain of an isolated network based on real-time power of source-load-storage. Alternatively, the computer program can instruct related hardware to implement these processes. The computer program can be stored in an electronic device, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0127] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0128] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0129] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0130] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0131] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0132] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0133] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for nonlinear mapping of stability region of isolated network based on real-time power of source-load-storage, characterized in that, Includes the following steps: S1: By installing measurement units at each node and branch of the isolated network, the power generation at the source end, the load power at the load end, and the charging and discharging power at the storage end within a preset time window are collected to form a real-time power data sequence characterizing the system's operating status. S2: The real-time power data sequence is segmented and stacked using a sliding time window to construct a multi-dimensional time-series power feature map containing time-series dependencies, which serves as the input to the trained nonlinear mapping network. S3: Input the multi-dimensional time-series power feature map into the nonlinear mapping network. The encoder in the nonlinear mapping network uses a multi-head attention mechanism to assign differentiated weights to power features of different types and times in the feature map, and calculates a weighted fused power state feature vector. S4: The decoder in the nonlinear mapping network performs regression mapping on the power state feature vector and outputs N key state indicators used to describe the boundary of the stable domain under the current operating state, where N is an integer greater than 1; S5: Based on the values ​​of the N key state indicators, in the N-dimensional state space with the N key state indicators as coordinate axes, a minimum convex polyhedron with the values ​​as vertices is generated using the convex hull construction algorithm, which serves as the dynamic stability domain at the current moment. S6: Map the weighted fused power state feature vector to the N-dimensional state space to obtain a current operating point; based on the preset stability margin evaluation function, calculate the shortest distance from the current operating point to each boundary surface of the dynamic stability domain, and normalize the shortest distance to obtain the real-time stability margin index. S7: When it is determined that the real-time stability margin index is lower than the preset warning threshold, an alarm control trigger signal is fed back to the isolated grid energy management system.

2. The method of claim 1, wherein, The training process of the nonlinear mapping network specifically includes: S31: Using a preset electromechanical transient simulation model of the power system, a large number of operating scenarios are generated by random combination within the preset fluctuation range of power generation at the source end, load power at the load end, and charging and discharging power at the storage end. S32: For each operating scenario, perform time-domain simulation, input the simulation trajectory data into the stability criterion analysis module, and automatically mark a set of key state indicators of the stability domain boundary corresponding to the scenario to form a labeled sample. S33: Using the multi-dimensional temporal power feature map as input and the labeled samples as the desired output, the encoder-decoder network containing the multi-head attention mechanism is trained offline using the gradient descent method until the preset loss function converges, thereby obtaining the nonlinear mapping network.

3. The method of claim 2, wherein, During the offline training process of S33, the encoder calculates a weighted and fused power-state feature vector through a multi-head attention mechanism, specifically including: S331: Perform a linear transformation on the input multidimensional time-series power feature map to generate a query matrix, a key matrix, and a value matrix corresponding to multiple attention heads; S332: For each attention head, calculate the dot product of the query matrix and the key matrix and scale it to obtain an attention score matrix that reflects the degree of correlation between power features at different time steps; S333: Use the Softmax function to convert the attention score matrix into an attention weight matrix, and then perform a weighted summation of the attention weight matrix and the value matrix to obtain the output features of the attention head; S334: Concatenate the output features of all attention heads and fuse them through a fully connected layer to generate the weighted fused power state feature vector.

4. The method of claim 2, wherein, The N key state indicators output in step S4 to describe the boundary of the stable domain are as follows: The decoder of the nonlinear mapping network outputs the coordinates of M vertices. The coordinates of each vertex are composed of the values ​​of N key state indicators, including the critical cut-off time, the lowest frequency point, the lowest voltage point, and the damping ratio of the dominant oscillation mode. The M vertices are used to jointly describe the boundary of the stability region. Wherein, N=4, and M is an integer greater than N.

5. The method of claim 4, wherein, In step S5, the dynamic stable region is generated using a convex hull construction algorithm, specifically as follows: In the N-dimensional state space with the N key state indicators as coordinate axes, the coordinates of the M vertices output by the decoder are used as a vertex set. All vertices in the vertex set are connected by the fast convex hull algorithm to form a minimal convex polyhedron. The internal region of this convex polyhedron is the dynamic stability domain that ensures the transient stability of the system.

6. The method of claim 1, wherein, The calculation of the shortest distance from the current running point to each boundary surface of the dynamic stable domain in S6 includes: calculating the shortest Euclidean distance from the current running point to each boundary surface of the dynamic stable domain using a point-to-polyhedron distance algorithm, and taking the minimum value among them as the shortest distance.

7. The method of claim 1, wherein, Also includes: The real-time power data sequence is preprocessed, including filtering out measurement noise using a Kalman filter and repairing missing data using linear interpolation, to obtain a cleaned power data stream. The cleaned power data stream is segmented and stacked using a sliding time window to construct the multidimensional time-series power feature map.

8. The method according to claim 1, characterized in that, Following step S7, step S8 is further included, which includes: S81: Encapsulate the real-time stability margin index and the corresponding source-end power generation, load-end power, storage-end charging and discharging power, and dynamic stability domain boundary information into a historical operating state data packet; S82: Periodically feed the historical running state data packets back to the training database, and when the preset model update triggering conditions are met, use the incremental learning mechanism to fine-tune the parameters of the nonlinear mapping network online.

9. The method according to claim 1, characterized in that, Also includes: S51: Obtain the future scheduling plan of the isolated grid system, the scheduling plan including the predicted power generation at the source end, the predicted power load at the load end, and the planned power charging and discharging at the storage end within a preset future time period; S52: Input the power prediction curve corresponding to the scheduling plan into the nonlinear mapping network to predict the predicted stability region corresponding to each time segment within a future preset period. S53: Based on the predicted stability region, assess whether there will be a risk period in the future where the real-time stability margin index is lower than the warning threshold; S54: If the risk period exists, calculate the minimum energy storage charging and discharging power adjustment required to make the stability margin index of the system reach above the warning threshold during the risk period, and generate a forward-looking energy storage pre-scheduling instruction accordingly, which is then fed back to the isolated grid energy management system.

10. A system for real-time power-based nonlinear mapping of stability region of an isolated grid based on source-load-storage, characterized in that, The system for performing the method as described in any one of claims 1 to 9 includes: Data acquisition module, the data acquisition module being configured to execute step S1; A feature map construction module, configured to execute step S2; The encoder module is configured to perform step S3; Decoder module, the decoder module being configured to perform step S4; A stability region generation module, configured to execute step S5; A margin assessment module, configured to perform step S6; The early warning feedback module is configured to execute step S7.

Images