Fault prediction method and device for excitation system of hydropower station generator set

Abstract

The invention provides a hydropower station generator set excitation system fault prediction method and device, and relates to the technical field of intelligent power grids, and the method comprises the steps: carrying out the continuous sampling of a generalized state observation vector through a sliding time window, so as to form an operation parameter observation tensor; fusing the operation parameter observation tensor and the boundary constraint type operation parameters of the generator, and constructing a fault observation tensor fused by multi-source operation parameters; introducing a structural causal atlas in a modeling layer, establishing a causal path network embedded based on causal reasoning and topological time sequence for a fault observation tensor, and dividing system interaction faults encountered by excitation system prediction into a plurality of typical modes; and performing graph attention modeling on the evolution trend and the instability boundary of each key parameter in the fault causal chains of different typical modes contained in the causal path network, and dynamically identifying the path characteristics of the excitation system entering the fault critical state. According to the method, the interaction type fault of the excitation system of the hydropower station generator set can be predicted.

Description

Technical Field

[0001] This application relates to the technical field of smart grids, and particularly to a method and device for predicting faults in the excitation system of a hydropower generator set. Background Art

[0002] As a key component for ensuring the quality of power output, maintaining voltage stability, and realizing the dynamic response of the power system, the excitation system of a hydropower generator set shows an increasingly complex evolution trend in the context of the current automation and intelligence of the energy system. The core function of the excitation system is to provide a DC excitation current to the rotor of the synchronous generator to establish the required magnetic field. The current mainstream forms of excitation systems include static excitation systems, rotary excitation systems, and compound excitation systems. Among them, the static excitation system is widely used in large hydropower units due to its advantages of high control accuracy, fast response speed, and convenient maintenance.

[0003] The interactive faults of the excitation system include complex phenomena such as poor coupling of voltage disturbances between the excitation system and the main power grid, improper cooperation between the excitation system and the power system stabilizer (PSS), synchronization failure, and abnormal excitation circulating current under the multi-machine parallel state. Such faults often occur during the unit switching mode, fast paralleling, or large power grid disturbances, with characteristics of suddenness and linkage. The fault diagnosis is difficult, and the risk boundary is uncertain. Especially in the high-impedance isolation structure of the static excitation system, once an abnormality occurs on the auxiliary power supply side, it may lead to the failure of the system's self-regulation ability, and then cause excitation disconnection. Therefore, a method is needed to predict the interactive faults of the excitation system of a hydropower generator set. Summary of the Invention

[0004] This application provides a method and device for predicting faults in the excitation system of a hydropower generator set, which can realize the prediction of interactive faults in the excitation system of a hydropower generator set.

[0005] In the first aspect of this application, a method for predicting faults in the excitation system of a hydropower generator set is provided. The method includes:

[0006] Based on the coupling relationship among the excitation system, synchronous generator, power system stabilizer, and main power grid, construct a generalized state observation vector including linkage parameters;

[0007] Continuously sample the generalized state observation vector through a sliding time window to form an operating parameter observation tensor;

[0008] Fuse the operating parameter observation tensor and the boundary constraint type operating parameters of the generator to construct a fault observation tensor for multi-source operating parameter fusion;

[0009] Introduce a structural causal graph at the modeling layer, establish a causal path network based on causal inference and topological temporal embedding for the fault observation tensor, and divide the system interaction faults predicted by the excitation system into multiple typical modes;

[0010] Perform graph attention modeling on the evolution trends and instability boundaries of the key parameters in the fault causal chains of different typical modes included in the causal path network to dynamically identify the path characteristics of the excitation system entering the fault critical state.

[0011] Based on the above technical solutions, preferably, the continuous sampling of the generalized state observation vector through a sliding time window to form an operating parameter observation tensor specifically includes:

[0012] Determine the time length and sliding step of the sliding time window;

[0013] In each sliding time window, sample the generalized state observation vector hourly at a globally unified sampling frequency for the system to obtain a sequence of generalized state observation vectors corresponding to multiple consecutive time points within the sliding time window. The generalized state observation vector at each time point is a vector of a fixed dimension and includes the linkage parameters from four subsystems: the excitation system, the synchronous generator, the power system stabilizer, and the main power grid;

[0014] In each sliding time window, form a two-dimensional matrix composed of the sequence of generalized state observation vectors arranged in time, where the rows represent the time evolution dimension and the columns represent different linkage parameter dimensions;

[0015] Advance the two-dimensional matrix in sequence by the sliding step in the time dimension, repeat the sampling process, and gradually construct a set of state observation sequences under multiple sliding time windows;

[0016] By stacking the set of state observation sequences sampled from all sliding time windows in the window sequence dimension, construct a three-dimensional operating parameter observation tensor, and the three-dimensional structure of the operating parameter observation tensor corresponds to the time window sequence dimension, the time evolution dimension, and the linkage parameter dimension respectively.

[0017] Based on the above technical solutions, preferably, the construction of a fault observation tensor by fusing the operating parameter observation tensor and the boundary constraint type operating parameters of the generator specifically includes:

[0018] For the typical boundary states faced by the synchronous generator during grid-connected operation, representative boundary constraint type operating parameters are extracted. The boundary constraint type operating parameters include the maximum excitation voltage limit factor of the synchronous generator excitation system, the minimum excitation current constraint value, the synchronous generator voltage reference switching state flag, the paralleling synchronization failure indication signal, the upper limit of the active load jump rate, the system frequency over-limit warning state, the duration of the instantaneous power-off of the auxiliary power supply, and the recovery delay identifier;

[0019] Based on the fact that most of the boundary constraint type operating parameters are trigger-based and irregular, boundary state information overlapping with the window timestamp is extracted within the time interval corresponding to the sliding time window and aggregated within the window. Further, time alignment processing is performed on the boundary constraint type operating parameters and the operating parameter observation tensor based on the master control system time synchronization mechanism;

[0020] After completing the time alignment processing, the boundary constraint type operating parameters are restructured to construct a tensor dimension compatible with the structure of the operating parameter observation tensor. By expanding in the direction of the linkage parameter dimension of the operating parameter observation tensor, multiple encoding channels representing the boundary constraint type operating parameters are added, so that the three-dimensional operating parameter observation tensor is expanded from the time window sequence dimension × time evolution dimension × linkage parameter dimension to the time window sequence dimension × time evolution dimension × fusion parameter dimension, where the fusion parameter dimension includes the joint expression of the linkage parameter and the boundary constraint type operating parameters;

[0021] A parameter self-attention mapping mechanism is introduced to weight-encode the importance of different boundary constraint type operating parameters. After completing the fusion encoding, a fault observation tensor of multi-source operating parameter fusion is output.

[0022] On the basis of the above technical solutions, preferably, a structural causal graph is introduced in the modeling layer to establish a causal path network for the fault observation tensor based on causal reasoning and topological time series embedding. The system interaction faults predicted by the excitation system are divided into multiple typical modes, specifically including:

[0023] In the causal relationship discovery stage, time series sub-vectors corresponding to each fusion parameter dimension are extracted from the fault observation tensor;

[0024] For the first time series vector and the second time series vector in multiple time series sub-vectors, it is checked whether the second time series vector can provide additional information to explain the current state of the second time series vector under the condition of the past state of the first time series vector within a fixed sliding window. If it holds, a directed edge from the first time series vector to the second time series vector is constructed, and at the same time, an information gain entropy criterion is introduced to filter out invalid edges with a causal strength lower than the statistical threshold;

[0025] In the topological graph structure construction stage, using the set of the fault observation tensors as the graph node set, and using the multiple directed edges as the edge set, construct the original graph topological structure of the structural causal graph;

[0026] Introduce a node attribute embedding mechanism to inject its operating physical attributes, measurement accuracy, boundary sensitivity, and the category of the subsystem it belongs to into each graph node, forming an attribute node graph with multiple semantics;

[0027] Individually construct the structural causal graph for each sliding window at any time, and establish horizontal connections between the sliding windows at any time to form a cross-time graph evolution sequence;

[0028] In the path embedding representation generation stage, use the graph attention network in the graph neural network as the path embedder to perform path characterization based on attention weights on all causal paths in the structural causal graph, and dynamically adjust the expression ability of the contribution of key nodes in different paths to the overall path;

[0029] According to the structural topology form, path activation intensity distribution, and edge weight evolution trajectory of the causal path set extracted from the graph embedding representation, aggregate all paths into stable domains through a clustering partition algorithm, and further map the stable domains to typical patterns of system interaction faults. The typical patterns include the coupling imbalance mode between the excitation system and the main grid voltage disturbance, the mutual interference mode between the excitation system and the power system stabilizer, the dominant mode of the synchronous failure path, and the multi-machine parallel excitation current offset and circulating current mode.

[0030] Based on the above technical solutions, preferably, perform graph attention modeling on the evolution trends and instability boundaries of each key parameter in the fault causal chains of different typical patterns included in the causal path network, and dynamically identify the path characteristics of the excitation system entering the fault critical state, specifically including:

[0031] According to the structural topology results of each typical fault causal chain, extract the key parameters that affect the path activation intensity and respond to the change of system stability from the fusion parameter set. Its determination principle is based on the graph betweenness centrality of the node, the cumulative weight of the causal edge, and the activation frequency statistics in the historical evolution path, and form a local path subgraph constructed with the key parameters as the center;

[0032] In the path graph structure construction stage, based on the typical causal path, extract the sub-path structure containing the key parameter nodes, retain the effective causal connection and the actual activation path between the parameters, and at the same time introduce time dimension embedding, by attaching a timestamp code or a time interval variable to each node, form a dynamic graph structure with time sequence information;

[0033] The constructed state graph structure is trained using a graph attention network, and dynamic attention weights are assigned to the edges between each node and its adjacent nodes. During the training process, the graph attention network learns the key parameter nodes that contribute the most to the path activation probability under different operating stages and different perturbation environments by minimizing the path activation prediction error function;

[0034] In the instability boundary modeling stage, based on the evolution trend of the key parameter nodes, combined with historical fault data and prior knowledge of the boundary state, an instability boundary determination model is constructed. The instability boundary determination model uses a sliding statistical window and a dynamic threshold discrimination mechanism to evaluate multiple indicators such as the change amplitude, direction consistency, and trend reversal points of the key parameters within a continuous time period, extracts the critical state features, and embeds the critical state features into the node attributes;

[0035] Based on the path representation vectors output by the graph attention network, combined with the determination results of the instability boundary determination model, the activity levels of all paths are sorted, and the paths that are close to the instability boundary, have high attention aggregation, and historical frequent activation records within the current time period are extracted as critical path features. The critical path features include the causal chain relationship between parameters, the activation weight of the overall path, the starting and ending parameters of the path, and the evolution slope and direction change information of the key nodes in the path.

[0036] On the basis of the above technical solutions, preferably, a generalized state observation vector including coupling parameters is constructed based on the coupling relationship between the excitation system, the synchronous generator, the power system stabilizer, and the main power grid, specifically including:

[0037] The first operating parameters in the excitation system are collected in real time. The first operating parameters include excitation voltage, excitation current, rectifier bridge output voltage, excitation winding temperature, and the offset between the excitation voltage regulator reference signal and the actual output signal;

[0038] The second operating parameters in the synchronous generator are synchronously collected. The second operating parameters include the amplitude of the synchronous generator terminal voltage, the phase angle of the synchronous generator voltage, the rotor angular velocity, the change rate of the excitation magnetic flux, the three-phase components of the stator current, and the power factor factor;

[0039] The third operating parameters of the power system stabilizer are collected. The third operating parameters include the gain adjustment coefficient, the modulation signal frequency distribution, the output signal phase delay, and the input signal amplitude change slope;

[0040] The fourth operating parameters of the main power grid are introduced. The fourth operating parameters include the instantaneous voltage offset, the grid short-circuit ratio, the grid frequency volatility, the grid impedance angle, and the equivalent reactance value;

[0041] After aligning the first operating parameter, the second operating parameter, the fourth operating parameter, and the fourth operating parameter at a unified clock frequency, a generalized state observation vector is constructed.

[0042] Based on the above technical solution, preferably, after performing graph attention modeling on the evolution trends and instability boundaries of key parameters in the fault causal chains of different typical modes included in the causal path network, and dynamically identifying the path characteristics of the excitation system entering the fault critical state, the method further includes:

[0043] Taking the path characteristics as the input basis, jointly construct a state representation vector of the excitation system by combining the path activation weight, the evolution slope of key parameters in the path, the direction change characteristics, the proximity to the instability boundary, and the current operating parameters of the main control system;

[0044] In the action space design stage, an action combination set is constructed around the adjustable control variables of the excitation system. Each action is represented as a parameter combination vector, and all actions constitute a discrete or continuous action space, covering all possible regulation operations that may affect the stability of the excitation system. The action combination set includes the upper and lower limit settings of the excitation voltage, the target value of the excitation current, the gain factor of the excitation regulator, the voltage reference signal switching mode, the coordinated modulation coefficient of the power system stabilizer, and the excitation mode switching instruction;

[0045] With the goal of delaying or avoiding fault triggering, the reward is designed as a reward function weighted by the negative activation degree, the proximity to the instability boundary, and the adjustment cost. In each time step, if the system state is far from the critical path, the reward value increases; if the path activation increases or the system approaches the instability boundary, the reward value decreases; if the execution of an action causes control oscillation or exceeds the physically bearable range, a penalty term is given; if the action successfully returns the system state to the safe domain, the maximum positive reward is given.

[0046] In the second aspect of the present application, a fault prediction device for a hydroelectric generator unit excitation system is provided. The device is used to execute any one of the above-mentioned hydroelectric generator unit excitation system fault prediction methods. The device includes an acquisition module, a processing module, and an output module, where:

[0047] The acquisition module is used to construct a generalized state observation vector including linkage parameters based on the coupling relationship between the excitation system, the synchronous generator, the power system stabilizer, and the main power grid;

[0048] The processing module is used to continuously sample the generalized state observation vector through a sliding time window to form an operating parameter observation tensor;

[0049] The processing module is configured to fuse the operation parameter observation tensor and the boundary constraint type operation parameters of the generator, and construct a fault observation tensor for multi-source operation parameter fusion;

[0050] The processing module is configured to introduce a structural causal graph in the modeling layer, establish a causal path network based on causal reasoning and topological time series embedding for the fault observation tensor, and divide the system interaction faults that the excitation system is predicted to encounter into multiple typical modes;

[0051] The output module is configured to perform graph attention modeling on the evolution trends and instability boundaries of the key parameters in the fault causal chains of different typical modes included in the causal path network, and dynamically identify the path characteristics of the excitation system entering the fault critical state.

[0052] In the third aspect of the present application, an electronic device is provided, including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The processor is used to execute the instructions stored in the memory, so that the electronic device executes the method described in any one of the above.

[0053] In the fourth aspect of the present application, a computer-readable storage medium is provided. The computer-readable storage medium stores instructions, and when the instructions are executed, the method described in any one of the above is executed.

[0054] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0055] 1. By constructing a generalized state observation vector covering the coupling relationship between the excitation system, the synchronous generator, the power system stabilizer, and the main power grid, combining a sliding time window to form a high-dimensional operation parameter observation tensor, and further fusing boundary constraint type operation parameters to form a multi-source fusion expression, the present application realizes the full-dimensional dynamic modeling of the system state; introducing a structural causal graph in the modeling layer, constructing a causal path network by combining causal reasoning and topological time series embedding, and deeply analyzing the causal chain under typical system interaction fault modes; finally, modeling the evolution trends and instability boundaries of the key parameters in the fault path through a graph attention mechanism, and dynamically identifying the path characteristics in the fault critical state, so as to realize the accurate prediction and evolution trend identification of the interaction faults of the hydroelectric generator set excitation system.

[0056] 2. The present application continuously samples the generalized state observation vector by introducing a sliding time window mechanism, constructs a three-dimensional structure operation parameter observation tensor with dimensions of time window sequence, time evolution, and linkage parameter, and can completely capture the dynamic response process and linkage evolution characteristics of the excitation system and its coupled system at different time scales, thereby providing high time series resolution and multi-variable coupling expression ability for subsequent fault identification, causal modeling, and state prediction, and realizing high-precision dynamic modeling of the system operation state and abnormal trend perception.

[0057] 3. The present application fuses the boundary constraint type operation parameters with the operation parameter observation tensor, and introduces a parameter self-attention mapping mechanism to realize weighted encoding of the boundary state information. While maintaining the original system dynamic feature expression ability, it effectively enhances the ability to depict and identify the operation boundary of the synchronous generator under extreme conditions, thereby constructing a multi-source fusion fault observation tensor with higher semantic sensitivity and boundary perception ability, providing a solid data foundation and structural guarantee for the accurate modeling and critical state prediction of system interaction faults.

[0058] 4. The present application introduces a structural causal graph to explicitly model the dynamic causal relationship between multi-dimensional fusion parameters in the fault observation tensor, and constructs a causal path network by combining topological structure and time series features. Under the support of the graph attention mechanism of the graph neural network, it realizes high-precision embedding representation and semantic extraction of key causal paths, and then classifies and identifies the system interaction faults that the excitation system may encounter based on the path structure, activation intensity, and edge weight evolution trajectory, effectively dividing them into multiple typical fault modes, thereby significantly improving the modeling ability, interpretability, and pattern prediction effect of complex linkage type faults.

[0059] 5. The present application performs graph attention modeling on the key parameters in the fault causal path network, combines path structure features, time evolution information, and proximity to the boundary, and dynamically identifies high-risk path features that can characterize the critical state of the excitation system, realizing early warning and accurate characterization of the fault evolution trend; by introducing an instability boundary determination model, embedding the trend change and boundary characteristics of the key parameters into the graph node attributes, it significantly enhances the sensitivity and prediction discriminability of path representation, thereby effectively improving the recognition ability and pre-control value of the critical state before the occurrence of system interaction faults in the excitation system of hydropower generator sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a schematic flow chart of a method for predicting faults in the excitation system of a hydropower generator set disclosed in an embodiment of the present application;

[0061] Figure 2 is a schematic module diagram of a device for predicting faults in the excitation system of a hydropower generator set disclosed in an embodiment of the present application;

[0062] Figure 3 It is a schematic structural diagram of an electronic device disclosed in an embodiment of the present application.

[0063] Explanation of reference numerals in the drawings: 201, acquisition module; 202, processing module; 203, output module; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. Specific embodiments

[0064] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments.

[0065] In the description of the embodiments of the present application, words such as "for example" or "for illustration" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "for example" or "for illustration" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Exactly speaking, the use of words such as "for example" or "for illustration" is intended to present relevant concepts in a specific manner.

[0066] In the description of the embodiments of the present application, the meaning of the term "a plurality" refers to two or more. For example, a plurality of systems refers to two or more systems, and a plurality of screen terminals refers to two or more screen terminals. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. The terms "include", "comprise", "have" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0067] As the core module for maintaining voltage stability and ensuring the quality of power output in a hydropower generator unit, especially in the context of the accelerating advancement of automation and intelligence, its structure and control are becoming increasingly complex. Although the mainstream structure represented by the static excitation system has high-precision and high-response characteristics, in the actual operation of the system, it still faces risks of interactive faults such as coupling imbalance with the main grid voltage disturbance, abnormal cooperation with the power system stabilizer, synchronization failure, and abnormal excitation circulating current in the multi-machine parallel state. Moreover, these faults mostly occur in the mode switching, grid disturbance, or high-load paralleling stage, with characteristics of suddenness, linkage, and unpredictability. Especially in the high-impedance isolation structure, the abnormal auxiliary power supply is likely to lead to system regulation failure. Therefore, it is urgent to construct a fault prediction method for this type of interactive mechanism to enhance the forward-looking guarantee ability of the system's stable operation.

[0068] This embodiment discloses a fault prediction method for the excitation system of a hydropower generator unit. Referring to Figure 1 , it includes the following steps S110 - S150:

[0069] S110, based on the coupling relationship among the excitation system, synchronous generator, power system stabilizer, and main grid, construct a generalized state observation vector including linkage parameters.

[0070] A fault prediction method for the excitation system of a hydropower generator unit disclosed in an embodiment of this application is applied to a server. The server includes, but is not limited to, electronic devices such as mobile phones, tablet computers, wearable devices, and PCs (Personal Computers), or can also be a background server running a fault prediction method for the excitation system of a hydropower generator unit. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0071] In a possible implementation manner, based on the coupling relationship among the excitation system, the synchronous generator, the power system stabilizer, and the main power grid, a generalized state observation vector including linkage parameters is constructed, specifically including: real-time acquisition of the first operating parameters in the excitation system, where the first operating parameters include excitation voltage, excitation current, rectifier bridge output voltage, excitation winding temperature, and the offset between the excitation voltage regulator reference signal and the actual output signal; synchronous acquisition of the second operating parameters in the synchronous generator, where the second operating parameters include synchronous generator terminal voltage amplitude, synchronous generator voltage phase angle, rotor angular velocity, excitation magnetic flux change rate, three-phase components of stator current, and power factor factor; acquisition of the third operating parameters of the power system stabilizer, where the third operating parameters include gain adjustment coefficient, modulation signal frequency distribution, output signal phase delay, and input signal amplitude change slope; introduction of the fourth operating parameters of the main power grid, where the fourth operating parameters include instantaneous voltage offset, grid short-circuit ratio, grid frequency volatility, grid impedance angle, and equivalent reactance value; after aligning the first operating parameters, the second operating parameters, the fourth operating parameters, and the fourth operating parameters at a unified clock frequency, a generalized state observation vector is constructed.

[0072] Specifically, the real-time acquisition of the first operating parameters in the excitation system specifically means that by arranging high-precision sensing units and digital measurement and control devices at key positions inside the excitation system, the excitation voltage, excitation current, rectifier bridge output voltage, excitation winding temperature, and the offset between the excitation voltage regulator reference signal and the actual output signal are respectively acquired. The excitation voltage and excitation current are extracted in real time through a voltage transformer and a Hall current sensor, the rectifier bridge output voltage is sampled through a multi-channel analog interface linked with the main control system, the excitation winding temperature is measured non-contact by a thermal resistance or fiber optic temperature sensor, the reference signal and the actual signal of the excitation voltage regulator are directly output by the internal operation module of the regulator, and the offset is calculated in real time through differential logic to fully capture the dynamic regulation ability and thermal stability state of the excitation system.

[0073] The synchronous acquisition of the second operating parameters in the synchronous generator means that current, voltage, high-speed position, and magnetic flux density measurement devices are arranged on the outgoing line side and the stator winding end of the synchronous generator to obtain the synchronous generator terminal voltage amplitude, synchronous generator voltage phase angle, rotor angular velocity, excitation magnetic flux change rate, three-phase components of stator current, and power factor factor. The voltage amplitude and phase angle are extracted by a synchronous phasor measurement unit, the rotor angular velocity is accurately measured by a high-resolution rotary encoder, the excitation magnetic flux change rate is jointly estimated by the rotor induced voltage and the stator back electromotive force, the three-phase components of the stator current are synchronously acquired through three current transformers and then reconstructed by Clark transformation, and the power factor factor is calculated in real time by the ratio of active power and reactive power, so as to accurately reflect the electromagnetic behavior response characteristics of the synchronous generator under controlled excitation.

[0074] Collecting the third operating parameter of the power system stabilizer refers to extracting parameters such as its gain adjustment coefficient, modulation signal frequency distribution, output signal phase delay, and input signal amplitude change slope in the control link of the power system stabilizer. The gain adjustment coefficient is directly read through the built-in parameter register of the stabilizer. The modulation signal frequency distribution is obtained from the frequency domain conversion of the input signal. The output signal phase delay is automatically calculated through the input-output timing difference. The input signal amplitude change slope is extracted by the sliding window difference method to obtain the continuous amplitude change rate. The above parameters are used to describe the regulation ability of the power system stabilizer under dynamic small disturbances and low-frequency oscillation responses and its influence on the excitation system control path, and are important information sources for modeling the voltage regulation-frequency stability composite action mechanism.

[0075] Introducing the fourth operating parameter of the main power grid refers to obtaining the instantaneous voltage offset, grid short-circuit ratio, grid frequency volatility, grid impedance angle, and equivalent reactance value of the main power grid by extracting real-time data from the synchronous measurement unit, wide-area measurement and control node, and protection measurement device deployed at the grid connection interface. The instantaneous voltage offset is obtained by collecting the original signal at the voltage measurement terminal and comparing it with the reference value. The grid short-circuit ratio is determined by calculating the ratio of the online impedance assessment to the grid connection capacity. The frequency volatility is obtained by taking the derivative of the frequency change amount in a continuous time period. The grid impedance angle and equivalent reactance value are obtained by fitting and calculating through the impedance measurement model, which is used to describe the electrical load coupling effect of the dynamic environment of the main power grid on the excitation system of the synchronous generator.

[0076] Finally, the first operating parameter, the second operating parameter, the third operating parameter, and the fourth operating parameter are synchronously sampled under the drive of the main clock of the control system or the unified GPS time stamp, and a data synchronization interface and a timing correction module are used to achieve unified alignment of multi-source data, so as to form a complete, structurally consistent, and time-synchronized generalized state observation vector. This observation vector integrates the operating information of the four subsystems of the excitation system, synchronous generator, power system stabilizer, and main power grid in the parameter dimension and satisfies global consistency in the time dimension, providing a unified multi-source time series expression basis for subsequent construction of the operating parameter observation tensor, extraction of path evolution characteristics, and prediction of system interaction faults.

[0077] S120, Continuously sample the generalized state observation vector through a sliding time window to form an operating parameter observation tensor.

[0078] In a possible implementation, the generalized state observation vector is continuously sampled by sliding a time window to form an operating parameter observation tensor, which specifically includes: determining the time length and sliding step of the sliding time window; in each sliding time window, the generalized state observation vector is sampled hourly at a globally unified sampling frequency of the system to obtain a sequence of generalized state observation vectors corresponding to consecutive time points within the sliding time window. The generalized state observation vector at each time point is a vector with a fixed dimension, including the coupling parameters from four subsystems: the excitation system, the synchronous generator, the power system stabilizer, and the main power grid; in each sliding time window, a two-dimensional matrix is formed by arranging the sequence of generalized state observation vectors in time order, where the rows represent the time evolution dimension and the columns represent different coupling parameter dimensions; in the time dimension, the two-dimensional matrix is advanced in sequence according to the sliding step, and the sampling process is repeated to gradually form a set of state observation sequences under multiple sliding time windows; by stacking all the state observation sequence sets sampled in the sliding time windows in the window sequence dimension, a three-dimensional operating parameter observation tensor is constructed, and the three-dimensional structure of the operating parameter observation tensor corresponds to the time window sequence dimension, the time evolution dimension, and the coupling parameter dimension respectively.

[0079] Specifically, determining the time length and sliding step of the sliding time window means that before constructing the operating parameter observation tensor, according to the typical dynamic response periods of the four subsystems of the excitation system, the synchronous generator, the power system stabilizer, and the main power grid, the sliding time window parameters for observing the system state evolution are reasonably selected. The time length determines the continuous time interval covered by each sampling period, usually set to a time period that can completely contain a small disturbance or an excitation regulation response period, such as 5 seconds to 30 seconds, to capture the dynamic change process of the coupling parameters during the disturbance conduction process; the sliding step is defined as the advancing distance between adjacent sliding time windows, and its setting should be between milliseconds and seconds to balance the dynamic tracking ability and the utilization rate of computing resources, ensuring effective information coverage without losing high-frequency features.

[0080] In each of the sliding time windows, sampling the generalized state observation vector hourly at a globally unified sampling frequency of the system specifically means performing periodic sampling operations according to the main control clock of the excitation system or the GPS clock synchronized with it, and discretely recording the system state at each time point within the entire sliding time window. Each time point corresponds to a complete generalized state observation vector with a fixed vector dimension, including all the coupling parameters from the four subsystems of the excitation system, the synchronous generator, the power system stabilizer, and the main power grid, ensuring that each sampling can restore the information structure of the entire system state space and providing high-time-resolution support for the state evolution trajectory within the sliding time window.

[0081] Within each of the sliding time windows, a two-dimensional matrix is formed by arranging the sequence of the generalized state observation vectors in time order. Specifically, all the generalized state observation vectors sampled within the sliding time window are arranged from top to bottom in time order to form a two-dimensional structured matrix with the dimension of time evolution × linkage parameter. Each row of this matrix represents the state set of the linkage parameters of all subsystems of the system at a certain time point, and each column represents the evolution trajectory of a certain linkage parameter within the entire time window, thereby constructing a structured data entity that can describe the time series change pattern and providing basic data segments for subsequent path modeling.

[0082] Advance the two-dimensional matrix in sequence by the sliding step length in the time dimension and repeat the sampling process. That is, after the sampling of the initial sliding time window is completed, keep the length of the sliding time window unchanged, slide forward by one sliding step length unit, resample the generalized state observation vectors in the new time period, and construct the next two-dimensional matrix. Repeat this step until the entire target time interval is covered. Each two-dimensional matrix generated by the advancement of the sliding time window retains the time series structure of the system state within that time period, enabling the system state to form a dynamic observation sequence with a high overlap rate under consecutive sliding time windows.

[0083] By stacking the state observation sequence sets sampled from all sliding time windows in the window sequence dimension, a three-dimensional operating parameter observation tensor is constructed. That is, the two-dimensional matrices generated within all the aforementioned sliding time windows are arranged in sequence and stacked along the time window sequence dimension direction to generate a complete three-dimensional tensor structure. The three dimensions of this operating parameter observation tensor are the time window sequence dimension, the time evolution dimension, and the linkage parameter dimension, which have the ability to perform long-term and short-term linkage modeling of the fault evolution path, can capture the multi-scale feature evolution process of the system state on the continuous time axis, and are the only input data structure basis for subsequent construction of the structural causal graph and identification of the causal path.

[0084] S130. Integrate the operating parameter observation tensor and the boundary constraint type operating parameters from the synchronous generator to construct a fault observation tensor with multi-source operating parameter integration.

[0085] In a possible implementation, the fault observation tensor of multi-source operating parameters is constructed by fusing the operating parameter observation tensor and the boundary constraint type operating parameters of the generator, which specifically includes: for the typical boundary states faced by the synchronous generator during grid-connected operation, extracting representative boundary constraint type operating parameters, where the boundary constraint type operating parameters include the maximum excitation voltage limit factor of the synchronous generator excitation system, the minimum excitation current constraint value, the synchronous generator voltage reference switching state flag, the paralleling synchronization failure indication signal, the upper limit of the active load jump rate, the system frequency over-limit warning state, the auxiliary power instantaneous power-off duration and the recovery delay identifier; based on the fact that most of the boundary constraint type operating parameters are trigger-based and irregular, extracting the boundary state information overlapping with the window timestamp within the time interval corresponding to the sliding time window, and performing in-window aggregation processing, and further performing time alignment processing on the boundary constraint type operating parameters and the operating parameter observation tensor based on the master control system time synchronization mechanism; after completing the time alignment processing, restructuring the boundary constraint type operating parameters to construct a tensor dimension compatible with the structure of the operating parameter observation tensor, and by expanding in the direction of the linkage parameter dimension of the operating parameter observation tensor, adding multiple coding channels representing the boundary constraint type operating parameters, so that the operating parameter observation tensor with a three-dimensional structure is expanded from the time window sequence dimension × time evolution dimension × linkage parameter dimension to the time window sequence dimension × time evolution dimension × fusion parameter dimension, where the fusion parameter dimension includes the joint expression of the linkage parameter and the boundary constraint type operating parameters; introducing a parameter self-attention mapping mechanism to perform weighted coding on the importance of different boundary constraint type operating parameters, and after completing the fusion coding, outputting the fault observation tensor of multi-source operating parameters fusion.

[0086] Specifically, extracting representative boundary constraint type operating parameters for the typical boundary states faced by the synchronous generator during grid-connected operation means identifying and selecting the operating state indicators that can characterize its operating stability boundary and limit constraints based on the system modeling of the operating limit behavior of the synchronous generator. The boundary constraint type operating parameters mainly include the maximum excitation voltage limit factor and the minimum excitation current constraint value of the synchronous generator excitation system, which are used to reflect the upper limit and the safety lower limit of the excitation capacity; the voltage reference switching state flag and the paralleling synchronization failure indication signal, which are used to identify the regulation mode switching and the grid-connected synchronization abnormality; the upper limit of the active load jump rate and the system frequency over-limit warning state, which are used to indicate the direct impact degree of the load dynamic disturbance on the synchronous generator; the auxiliary power instantaneous power-off duration and its recovery delay identifier, which are used to capture the change of the power supply support ability boundary under the high-impedance isolation structure. Most of these boundary constraint type operating parameters are derived from the protection device records, the control system event logs or the status register, and have high sensitivity and state segmentation characteristics in the fault prediction task.

[0087] Based on the fact that most of the running parameters of the boundary constraint class are triggering and irregular, extracting the boundary state information overlapping with the window timestamp within the time interval corresponding to the sliding time window and performing in-window aggregation processing means that each sliding time window is time-overlapped and matched with the boundary event sequence, extracting the corresponding boundary state flag during the successfully matched time period, and converting it into a continuous encoding expression through the in-window aggregation strategy. For example, for discrete event boundary constraint parameters, methods such as trigger frequency normalization, duration weighting, or trigger moment centering are used to map them into a continuous numerical representation between 0 and 1; for quantitative constraint parameters, methods such as the maximum value, average value, or slope trend within the window are used for summary encoding to enhance the continuity of their expression in the time dimension. The above operations ensure that the running parameters of the boundary constraint class have the same granularity as the sliding time window in the time domain.

[0088] Further, based on the master control system time synchronization mechanism, performing time alignment processing on the running parameters of the boundary constraint class and the running parameter observation tensor means using the system's unified clock signal as the time reference, and precisely aligning the sampling points in the running parameters of the boundary constraint class and the running parameter observation tensor through the synchronous timing interpolation mechanism, ensuring that within each sliding time window, the sampling points of the boundary parameters and the linkage parameters have a one-to-one correspondence on the time axis, thereby eliminating the sequence asynchrony problem caused by sampling frequency differences, clock drift, or timestamp misalignment brought by different sampling devices.

[0089] After completing the time alignment processing, restructuring the running parameters of the boundary constraint class to construct a tensor dimension compatible with the structure of the running parameter observation tensor means embedding the dimension of the aligned running parameters of the boundary constraint class in the three-dimensional tensor structure. The specific implementation method is to perform dimension expansion in the direction of the linkage parameter dimension of the running parameter observation tensor, adding several channels as the expression channels for the running parameters of the boundary constraint class. The dimension of each channel is kept consistent with the time window sequence dimension and the time evolution dimension of the running parameter observation tensor, so that the original tensor of time window sequence dimension × time evolution dimension × linkage parameter dimension is expanded to time window sequence dimension × time evolution dimension × fusion parameter dimension, where the fusion parameter dimension is the combined sum of the linkage parameter dimension and the running parameter dimension of the boundary constraint class, realizing the unified fusion of multi-source running information structurally.

[0090] Introduce a parameter self-attention mapping mechanism to perform weighted encoding on the importance of the operating parameters of different boundary constraint classes, which means constructing a lightweight parameter attention allocation module. Taking the overall state of the system within the current sliding time window as the context, a weight coefficient is assigned to each operating parameter of the boundary constraint class to express its importance for fault discrimination at that window moment. This attention allocation module can be implemented using a soft attention mechanism based on parameter correlation measurement or a graph attention mechanism based on path activation patterns. Finally, a set of normalized weights are output to perform weighted summation on the boundary constraint class operating parameter channels, further forming a fusion expression feature with time-varying adaptive ability.

[0091] After completing the above weighted fusion encoding, the output fault observation tensor of the multi-source operating parameter fusion not only realizes the joint expression of multi-subsystem parameters and boundary constraint class operating information in terms of structure, retains the complete dynamic evolution characteristics in the time dimension, but also has the ability to sensitively reflect the changes in key operating boundary states at the semantic level, providing an input expression basis with high identifiability, high representation ability, and high generalization ability for subsequent causal path modeling and system interaction fault prediction tasks.

[0092] S140, introduce a structural causal graph in the modeling layer to establish a causal path network for the fault observation tensor based on causal reasoning and topological time series embedding, and divide the system interaction faults encountered in the excitation system prediction into multiple typical modes.

[0093] In a possible implementation, a structural causal graph is introduced in the modeling layer to establish a causal path network for the fault observation tensor based on causal reasoning and topological temporal embedding. The system interaction faults predicted by the excitation system are divided into multiple typical modes, specifically including: In the causal relationship discovery stage, time series sub-vectors corresponding to each fusion parameter dimension are extracted from the fault observation tensor; for the first time series vector and the second time series vector among multiple time series sub-vectors, it is checked whether the second time series vector can provide additional information to explain the current state of the second time series vector under the condition of the past state of the given first time series vector within a fixed sliding window. If it holds, a directed edge is constructed from the first time series vector to the second time series vector, and at the same time, an information gain entropy criterion is introduced to filter out invalid edges with causal strength lower than the statistical threshold; in the topological graph structure construction stage, the set of fault observation tensors is used as the graph node set, and multiple directed edges are used as the edge set to construct the original graph topological structure of the structural causal graph; a node attribute embedding mechanism is introduced to inject its operating physical attributes, measurement accuracy, boundary sensitivity, and the category of the subsystem to which it belongs into each graph node, forming an attribute node graph with multiple semantics; a structural causal graph for each sliding window is constructed separately, and horizontal connections are established between sliding windows to form a cross-time graph evolution sequence; in the path embedding representation generation stage, the graph attention network in the graph neural network is used as the path embedder to perform path characterization based on attention weights for all causal paths in the structural causal graph, dynamically adjusting the expression ability of the contribution of key nodes in different paths to the overall path; according to the structural topological form, path activation intensity distribution, and edge weight evolution trajectory of the causal path set extracted from the graph embedding representation, all paths are aggregated into stable domains through a clustering division algorithm, and further the stable domains are mapped to the typical modes of system interaction faults. The typical modes include the coupling imbalance mode between the excitation system and the main grid voltage disturbance, the mutual interference mode between the excitation system and the power system stabilizer, the dominant mode of the synchronization failure path, and the multi-machine parallel excitation current offset and circulating current mode.

[0094] Specifically, in the causal relationship discovery stage, extracting the time series sub-vectors corresponding to each fusion parameter dimension from the fault observation tensor specifically means slicing the fault observation tensor according to the fusion parameter dimension, and extracting the complete observation trajectories of each fusion parameter in all time window sequences and time evolution dimensions, so as to obtain multiple independent time series sub-vector sets. Each time series sub-vector represents the dynamic evolution process of a certain fusion parameter under a sliding time window, retaining the trends, jumps, lags, and steady-state characteristics of the system state changing with time, providing an original input structure with temporality and variable independence for subsequent causality analysis.

[0095] For the first time series vector and the second time series vector among multiple said time series sub-vectors, in a fixed sliding window, it is examined whether the second time series vector can provide additional information to explain the current state of the second time series vector under the condition of the past state of the first time series vector. If it holds, a directed edge from the first time series vector to the second time series vector is constructed. Meanwhile, an information gain entropy criterion is introduced to filter out invalid edges with a causal strength lower than the statistical threshold. The specific implementation is as follows: The Granger causality test method is used to construct a conditional prediction model in each fixed sliding time window, and the change in the prediction residuals of the current state of the second time series vector before and after adding the historical information of the first time series vector is compared. If the residuals are significantly reduced, it is considered that there is an effective causal relationship, and a directed edge with the direction from the first vector to the second vector is constructed; To improve the sparsity and discriminability of the causal graph, the information gain entropy criterion is simultaneously introduced for all potential causal edges, the contribution degree of the causal edge to the overall state uncertainty is calculated, and the weak causal relationships below the threshold are removed, and the key influence paths with strong discriminability are retained.

[0096] In the stage of constructing the topological graph structure, the set of fusion parameters of the fault observation tensor is used as the set of graph nodes, and multiple said directed edges are used as the edge set to construct the original graph topological structure of the structural causal graph. Specifically, it means that each fusion parameter is used as a graph node with a unique identifier, and each directed edge established based on causal reasoning is used as the connection relationship between nodes. Under the joint action of the node set and the edge set, a directed graph structure is generated to express the dynamic causal propagation relationship between multiple parameters. This graph structure is consistent with the time distribution of the fault observation tensor and forms a local static topology in each sliding time window to capture the variable driving mechanism inside the system during a specific time period.

[0097] The node attribute embedding mechanism is introduced to inject its operating physical attributes, measurement accuracy, boundary sensitivity, and the category of the subsystem to which it belongs into each said graph node, forming an attribute node graph with multiple semantics. It means that on the basis of the topological structure, multiple auxiliary information of each graph node and its corresponding fusion parameter are embedded and integrated. For example, the unit, dimension, standard deviation of the measurement error, response sensitivity coefficient under the boundary state, and the subsystem identifier from the excitation system, synchronous generator, power system stabilizer, or main power grid of the parameter represented by the node are used to form a composite node representation with physical, measurement, and structural multi-semantic information. This embedding mechanism provides semantic prior support for the graph neural network in subsequent path modeling, improving the expression ability and interpretability of path learning.

[0098] The structural causal graph for each time-sliding window is constructed independently, and horizontal connections are established between time-sliding windows to form a graph evolution sequence across time. Specifically, it means generating the structural causal graph independently within each sliding time window, and taking the window sequence as the time reference, connecting the same nodes in adjacent windows according to the time evolution relationship to construct a graph structure sequence with a time-axis dimension. This graph evolution sequence forms a graph time series model in structure, supports the modeling of the dynamic change trend of causal paths over time, and provides continuous modeling support for identifying the path activation intensity, fault propagation chain, and critical state formation mechanism.

[0099] In the path embedding representation generation stage, the graph attention network in the graph neural network is used as a path embedder to perform path representation based on attention weights for all causal paths in the structural causal graph, dynamically adjusting the expression ability of the contribution of key nodes in different paths to the overall path. Specifically, it means using the graph attention network to assign dynamically learnable attention weights to each node in each path, strengthening the expression ability of nodes with high influence or high boundary sensitivity under the current system state, and generating a path-level embedding vector by aggregating the weighted representation vectors of different nodes, thereby forming a set of embedding path representations with state awareness, path feature distinguishability, and fault evolution directionality.

[0100] According to the structural topology, path activation intensity distribution, and edge weight evolution trajectory of the causal path set extracted from the graph embedding representation, all paths are aggregated into stable domains through a clustering algorithm, and further mapping the stable domains to typical patterns of system interaction faults. It means performing unsupervised clustering analysis on the set of embedded paths, identifying subsets of causal paths with similar topological structures, activation patterns, and dynamic evolution laws, and defining this subset as a stable domain to characterize the causal path features of the system operating state under a certain potential interaction fault type. Each stable domain is further mapped to a typical pattern of a class of system interaction faults through pattern labeling operations, including the coupling imbalance pattern between the excitation system and the main grid voltage disturbance, the mutual interference pattern between the excitation system and the power system stabilizer, the dominant pattern of the synchronous failure path, and the multi-machine parallel excitation current offset and circulating current pattern, realizing the path-driven modeling closed-loop of system interaction faults from original observation to structural cognition.

[0101] S150, perform graph attention modeling on the evolution trend and instability boundary of each key parameter in the fault causal chains of different typical patterns included in the causal path network, and dynamically identify the path features of the excitation system entering the fault critical state.

[0102] In a possible implementation, graph attention modeling is performed on the evolution trends and instability boundaries of key parameters in the fault causal chains of different typical patterns included in the causal path network to dynamically identify the path characteristics of the excitation system entering the fault critical state. Specifically, based on the structural topology results of each typical fault causal chain, key parameters that affect the path activation intensity and respond to changes in system stability are extracted from the fusion parameter set. The determination principle is based on the graph betweenness centrality of nodes, the cumulative weight of causal edges, and the activation frequency statistics in the historical evolution path, forming a local path subgraph constructed with the key parameters as the center. In the path graph structure construction stage, based on the typical causal path, the sub-path structure containing key parameter nodes is extracted, and the effective causal connections and actual activation paths between parameters are retained. At the same time, the time dimension embedding is introduced by attaching a timestamp code or a time interval variable to each node, forming a dynamic graph structure with temporal information. The constructed dynamic graph structure is trained using a graph attention network, and dynamic attention weights are assigned to the edges between each node and its adjacent nodes. During the training process, the graph attention network learns the key parameter nodes that contribute the most to the path activation probability under different operating stages and different perturbation environments by minimizing the path activation prediction error function. In the instability boundary modeling stage, based on the evolution trends of the key parameter nodes, combined with historical fault data and prior knowledge of the boundary state, an instability boundary determination model is constructed. The instability boundary determination model uses a sliding statistical window and a dynamic threshold discrimination mechanism to perform multi-index evaluations on the change amplitude, direction consistency, and trend reversal points of the key parameters within a continuous time period, extracts the critical state characteristics, and embeds the critical state characteristics into the node attributes. Based on the path representation vectors output by the graph attention network and combined with the determination results of the instability boundary determination model, the activity levels of all paths are sorted, and the paths that are close to the instability boundary, have high attention aggregation, and have a historical frequent activation record in the current time period are extracted as the critical path characteristics. The critical path characteristics include the causal chain relationship between parameters, the overall activation weight of the path, the starting and ending parameters of the path, and the evolution slope and direction change information of the key nodes in the path.

[0103] Specifically, based on the structural topology results of each typical fault causal chain, key parameters that affect the activation strength of the influence path and the change in the stability of the response system are extracted from the fusion parameter set. The specific implementation is as follows: In the constructed causal path network, all typical fault causal chains are traversed, and the graph betweenness centrality measure of each node is calculated to identify the parameter nodes with high information flow control in path transmission; at the same time, the weights of all causal edges associated with this node are accumulated to reflect its direct driving force for the overall path activation; combined with the activation frequency of this node in the historical evolution path, its probability of participating in the system instability process under multiple operation disturbances is evaluated. The comprehensive score of the three is used as the key parameter determination index, thereby forming a local path subgraph constructed with key parameters as the center to ensure that the extracted key parameters have structural influence, behavior driving force, and evolutionary representativeness.

[0104] In the path graph structure construction stage, based on the typical causal path, a sub-path structure containing key parameter nodes is extracted. Specifically, in each typical fault causal chain, all key parameter nodes are located, and their precursor nodes and precursor edges are traced back, and the complete activation path segment where they are located is intercepted. At the same time, the directed causal connection relationship between the parameters in the path is retained to construct the smallest path subgraph covering the key parameters; on this basis, a time dimension embedding variable is introduced for each node, and its sampling timestamp or the time interval between the last activation state is encoded as a time embedding feature, which together with the physical attributes of the node constitutes the node embedding vector, thereby generating a dynamic graph structure with temporal attributes, completely retaining the dynamic evolution trajectory and temporal dependence relationship of the causal path.

[0105] The constructed state graph structure is trained using a graph attention network. Specifically, the key parameter path graph with time embedding is input into the graph attention network model. In each graph neural network propagation layer, a learnable attention weight based on the similarity of node embedding vectors is applied to the edges between each node and its adjacent nodes, and multi-source information in the path is integrated through the multi-head attention mechanism; during the training process, the path activation prediction error is used as the loss function target, and supervised training is carried out using the historical labeled path activation state labels to optimize the parameter update process, so that the model gradually learns the key parameter nodes and their temporal evolution characteristics that contribute the most to the path activation probability under different disturbance types and operation stages, thereby improving the model's ability to identify high-risk paths and the accuracy of path dynamic representation.

[0106] Select key parameters in multiple time windows in the historical operation data that are close to or enter the pre-fault state. Extract statistical features of their change amplitude, change direction consistency, and trend reversal points within a continuous time period. Calculate indicators such as dynamic standard deviation, slope change, and number of turning points in multiple time periods using a sliding window method to construct a multi-dimensional state vector reflecting whether the parameter enters the critical state. Based on the boundary state features extracted from historical fault samples, construct a dynamic threshold model to determine whether the key parameter reaches the critical state at the current moment. Convert the judgment result into a binary or probabilistic label and embed it into the corresponding graph node attributes, enabling the graph attention network to identify and amplify the importance of the current critical state node during information propagation.

[0107] Based on the path representation vectors output by the graph attention network and combined with the judgment results of the instability boundary judgment model, sort the activity levels of all paths, and extract the paths that are close to the instability boundary, have high attention aggregation, and historical frequent activation records during the current time period as critical path features. Specifically, in each prediction cycle, the embedded vectors of all key parameter paths and the instability boundary state are jointly input into the path activation sorting module to jointly evaluate the activity intensity, instability probability, and importance of the paths. According to the sorting results, select the path set in the activation critical state or high-risk evolution state as the critical path feature. Each critical path feature contains the path start parameter and end parameter identifiers, the list of key nodes in the path, the path activation weight value, the slope change and direction reversal information of the key parameters, which are used to accurately describe the potential fault evolution path of the current system and serve as a structured input for fault prediction and control strategy generation.

[0108] In a possible implementation, after performing graph attention modeling on the evolution trends of key parameters and the instability boundaries in the fault causal chains of different typical patterns included in the causal path network to dynamically identify the path features when the excitation system enters the fault critical state, the method further includes: using the path features as the input basis, jointly constructing a state representation vector of the excitation system by combining the path activation weight, the evolution slope of the key parameters in the path, the direction change feature, the proximity to the instability boundary, and the current operating parameters of the main control system; in the action space design stage, constructing an action combination set around the adjustable control variables of the excitation system, each action is represented as a parameter combination vector, and all actions form a discrete or continuous action space, covering all possible regulation operations that may affect the stability of the excitation system. The action combination set includes the upper and lower limit settings of the excitation voltage, the target value of the excitation current, the gain factor of the excitation regulator, the voltage reference signal switching mode, the coordinated modulation coefficient of the power system stabilizer, and the excitation mode switching instruction; aiming at delaying or avoiding fault triggering, designing the reward as a reward function weighted by the negative activation degree, the proximity to the instability boundary, and the adjustment cost. At each time step, if the system state is far from the critical path, the reward value increases; if the path activation increases or the system approaches the instability boundary, the reward value decreases; if the execution of an action causes control oscillation or exceeds the physically bearable range, a penalty term is given; if the action successfully returns the system state to the safe domain, the maximum positive reward is given.

[0109] Specifically, using the path features as the input basis and jointly constructing a state representation vector of the excitation system by combining the path activation weight, the evolution slope of the key parameters in the path, the direction change feature, the proximity to the instability boundary, and the current operating parameters of the main control system means that after the graph attention modeling stage is completed, all the path features identified as critical paths are extracted from the current sliding time window, including the activation weight value of the overall path, the change slope of the key parameters in the path over time, the trend factor of the parameter value change direction, and the distance index between the current state and the preset instability boundary, and jointly with the control variables such as the excitation voltage, excitation current, rotor angular velocity, voltage reference source state, and power system stabilizer gain collected in real time in the main control system, a high-dimensional state vector is constructed after unified normalization processing. This state representation vector has structural drivability and physical interpretability, can accurately express the operating environment, risk level, and dynamic trend of the current excitation system, and provides state-aware input for policy generation in reinforcement learning.

[0110] In the action space design stage, an action combination set is constructed around the adjustable control variables of the excitation system. Specifically, according to the adjustment ability boundary and actual control channels of the excitation system, the control variables that can directly affect its response ability and stability performance are selected as action elements to construct a multi-dimensional parameter combination. Each action is represented as a control parameter vector, including the upper limit setting value of the excitation voltage, the lower limit setting value of the excitation voltage, the target value of the excitation current, the proportional gain factor of the excitation regulator, the switching mode of the voltage reference signal between the main reference and the standby reference, the adjustment ratio of the coordinated modulation coefficient of the power system stabilizer, and the switching instruction of the excitation mode between the automatic control mode, the constant current mode, and the constant voltage mode. The above action combination includes both the fine adjustment of continuous variables and the mode switching of discrete states, constituting a complete action set that covers the entire system adjustment space, enabling the policy network to have sufficient control selection capabilities during the policy generation process to cope with various fault critical states.

[0111] With the goal of delaying or avoiding fault triggering, the reward is designed as a reward function composed of the weighted sum of the negative activation degree, the proximity to the instability boundary, and the adjustment cost. Specifically, in each state transition step, the change between the current state and the previous state of the system is evaluated. If, after performing an action, the current state of the system deviates in a direction away from the critical path, the reward function increases, reflecting that the control operation successfully realizes state fallback; if the activation degree of the path increases or the distance to the instability boundary decreases, and the system approaches the fault area, a negative reward is given, indicating that the action exacerbates the risk; if the execution of the action causes the excitation voltage or excitation current to exceed the set upper and lower limits, or generates overshoot oscillations, an additional penalty term is added to avoid aggressive control; conversely, if the action significantly reduces the activation intensity of the activated path and increases the stable slope of the key parameters, and the operation is within the physical constraints, the maximum positive reward is given. This reward function comprehensively considers the system safety objective, response smoothness, and control cost, and realizes the precise guidance and stable optimization of the learning direction of the policy network.

[0112] This embodiment also discloses a fault prediction device for the excitation system of a hydropower generator set. The device is used to execute any one of the above-mentioned fault prediction methods for the excitation system of a hydropower generator set, referring to Figure 2 , and includes an acquisition module 201, a processing module 202, and an output module 203, where:

[0113] The acquisition module 201 is used to construct a generalized state observation vector including linkage parameters based on the coupling relationship between the excitation system, the synchronous generator, the power system stabilizer, and the main power grid.

[0114] The processing module 202 is used to continuously sample the generalized state observation vector through a sliding time window to form an operating parameter observation tensor.

[0115] The processing module 202 is configured to fuse the operation parameter observation tensor and the boundary constraint type operation parameters of the generator, and construct a fault observation tensor for multi-source operation parameter fusion.

[0116] The processing module 202 is configured to introduce a structural causal graph in the modeling layer, establish a causal path network based on causal reasoning and topological time series embedding for the fault observation tensor, and divide the system interaction faults predicted by the excitation system into multiple typical modes.

[0117] The output module 203 is configured to perform graph attention modeling on the evolution trends and instability boundaries of the key parameters in the fault causal chains of different typical modes included in the causal path network, and dynamically identify the path characteristics of the excitation system entering the fault critical state.

[0118] It should be noted that when the device provided in the above embodiment realizes its functions, only the above-mentioned division of each functional module is used for illustration. In practical applications, the above functions can be allocated to different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process can be found in the method embodiment, which will not be elaborated here.

[0119] This embodiment also discloses an electronic device. Referring to Figure 3 , the electronic device may include: at least one processor 301, at least one communication bus 302, a user interface 303, a network interface 304, and at least one memory 305.

[0120] Among them, the communication bus 302 is used to realize the connection and communication between these components.

[0121] Among them, the user interface 303 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 303 may further include a standard wired interface and a wireless interface.

[0122] Among them, the network interface 304 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0123] Among them, the processor 301 may include one or more processing cores. The processor 301 connects various parts within the entire server through various interfaces and lines, and executes various functions of the server and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 305, and by calling the data stored in the memory 305. Optionally, the processor 301 may be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), or programmable logic array (PLA). The processor 301 may integrate one or several combinations of a central processing unit (CPU), a graphics processing unit (GPU), and a modem, etc. Among them, the CPU mainly processes the operating system, user interface, application programs, etc.; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 301 and may be implemented separately through a single chip.

[0124] Among them, the memory 305 may include random access memory (RAM) and may also include read-only memory. Optionally, the memory includes a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, codes, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area. Among them, the program storage area may store instructions for implementing the operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store the data involved in the above-mentioned various method embodiments. Optionally, the memory 305 may also be at least one storage device located far from the aforementioned processor 301. As a computer storage medium, the memory 305 may include an operating system, a network communication module, a user interface 303 module, and an application program for a fault prediction method of a hydroelectric generator excitation system.

[0125] In Figure 3In the electronic device shown, the user interface 303 is mainly used to provide an interface for the user to input and obtain the data input by the user; and the processor 301 can be used to call the application program stored in the memory 305 for a method for predicting faults in the excitation system of a hydropower generator set. When executed by one or more processors 301, the electronic device is caused to execute the method of one or more of the above embodiments.

[0126] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that this application is not limited by the described action sequence, because according to this application, certain steps can be adopted in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0127] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0128] In the several embodiments provided by this application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is only a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection to each other can be through some service interfaces. The indirect coupling or communication connection of the device or unit can be in an electrical or other form.

[0129] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place, or they can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0130] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated units can be implemented in the form of hardware or in the form of software functional units.

[0131] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 305 and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of this application. And the aforementioned memory 305 includes: various media such as USB flash drives, mobile hard disks, magnetic disks, or optical discs that can store program codes.

[0132] This application also discloses a computer-readable storage medium that stores instructions. When executed by one or more processors 301, it enables the electronic device to execute the method as described in one or more of the above embodiments.

[0133] The foregoing are only exemplary embodiments of the present disclosure and should not be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure still fall within the scope covered by the present disclosure. After considering the specification and the practice of the disclosure, those skilled in the art will easily think of other embodiments of the present disclosure. This application aims to cover any variations, uses, or adaptive changes of the present disclosure, and these variations, uses, or adaptive changes follow the general principles of the present disclosure and include the common general knowledge or conventional technical means in the technical field not recorded in the present disclosure. The specification and the embodiments are only regarded as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A fault prediction method for the excitation system of a hydropower generating unit, characterized in that The method includes: Based on the coupling relationship among the excitation system, synchronous generator, power system stabilizer, and main power grid, construct a generalized state observation vector including coupling parameters; Continuously sample the generalized state observation vector through a sliding time window to form an operating parameter observation tensor; Fuse the operating parameter observation tensor and the boundary constraint type operating parameters of the generator to construct a fault observation tensor with multi-source operating parameter fusion; Introduce a structural causal graph in the modeling layer, establish a causal path network based on causal reasoning and topological time series embedding for the fault observation tensor, and divide the system interaction faults predicted by the excitation system into multiple typical modes; Perform graph attention modeling on the evolution trends and instability boundaries of each key parameter in the fault causal chains of different typical modes included in the causal path network, and dynamically identify the path characteristics of the excitation system entering the fault critical state.

2. A fault prediction method for the excitation system of a hydropower generator set according to claim 1, characterized in that, The continuous sampling of the generalized state observation vector through a sliding time window to form an operating parameter observation tensor specifically includes: Determine the time length and sliding step of the sliding time window; In each sliding time window, sample the generalized state observation vector hourly at a globally unified sampling frequency for the system to obtain a sequence of generalized state observation vectors corresponding to multiple consecutive time points within the sliding time window. Each generalized state observation vector at each time point is a vector with a fixed dimension and includes coupling parameters from four subsystems: the excitation system, synchronous generator, power system stabilizer, and main power grid; Form a two-dimensional matrix within each sliding time window, which is composed of the sequence of generalized state observation vectors arranged in time. Here, the rows represent the time evolution dimension, and the columns represent different coupling parameter dimensions; Advance the two-dimensional matrix in sequence at the sliding step in the time dimension, repeat the sampling process, and gradually form a set of state observation sequences under multiple sliding time windows; By stacking the set of state observation sequences sampled from all sliding time windows in the window sequence dimension, construct a three-dimensional operating parameter observation tensor. The three-dimensional structure of the operating parameter observation tensor corresponds to the time window sequence dimension, time evolution dimension, and coupling parameter dimension respectively.

3. A method for predicting faults in the excitation system of a hydropower generator set according to claim 2, characterized in that The fusion of the operating parameter observation tensor and the boundary constraint type operating parameters of the generator to construct a fault observation tensor with multi-source operating parameter fusion specifically includes: For the typical boundary states faced by the synchronous generator during grid-connected operation, extract representative boundary constraint type operating parameters. The boundary constraint type operating parameters include the maximum excitation voltage limit factor of the synchronous generator excitation system, minimum excitation current constraint value, synchronous generator voltage reference switching state flag, paralleling synchronization failure indication signal, upper limit of active load jump rate, system frequency over-limit alarm state, instantaneous power-off duration and recovery delay identification of the auxiliary power supply; Based on the fact that most boundary constraint operating parameters are triggered and irregular, the boundary state information overlapping with the window timestamp is extracted within the time interval corresponding to the sliding time window, and the in-window aggregation processing is performed. Further, the boundary constraint operating parameters and the operating parameter observation tensor are time-aligned based on the time synchronization mechanism of the main control system. After completing the time alignment process, the boundary constraint type operating parameters are restructured to construct a tensor dimension compatible with the operating parameter observation tensor structure, and multiple encoding channels representing the boundary constraint type operating parameters are added by expanding the operating parameter observation tensor in the direction of the linkage parameter dimension, so that the operating parameter observation tensor of the three-dimensional structure is expanded from the time window sequence dimension × time evolution dimension × linkage parameter dimension to the time window sequence dimension × time evolution dimension × fusion parameter dimension, wherein the fusion parameter dimension includes the joint expression of the linkage parameter and the boundary constraint type operating parameter; A parameter self-attention mapping mechanism is introduced to perform weighted encoding on the importance of operating parameters of different boundary constraint classes. After fusion encoding is completed, the fault observation tensor of multi-source operating parameter fusion is output.

4. A fault prediction method for the excitation system of a hydropower generator set according to claim 1, characterized in that, The structural causal graph is introduced at the modeling layer, and a causal path network based on causal reasoning and topological time series embedding is established for the fault observation tensor, and the system interaction faults predicted to be encountered by the excitation system are divided into multiple typical modes, specifically including: In the causal relationship discovery stage, a time series sub-vector corresponding to each fusion parameter dimension is extracted from the fault observation tensor; For a first time series vector and a second time series vector in the plurality of time series sub-vectors, checking within a fixed sliding window whether the second time series vector can provide additional information to explain the current state of the second time series vector given the past state of the first time series vector, and if so, constructing a directed edge from the first time series vector to the second time series vector, while introducing an information gain entropy criterion to filter out invalid edges whose causal strength is lower than a statistical threshold; In the topology graph structure construction stage, the original graph topology structure of the structural causal graph is constructed by taking the set of the fault observation tensors as the graph node set and the plurality of directed edges as the edge set; A node attribute embedding mechanism is introduced to inject the running physical attributes, measurement accuracy, boundary sensitivity and subsystem category of each graph node into each graph node, so as to form an attribute node graph with multiple semantics; The structural causal graph of each sliding window at any time is constructed separately, and horizontal connections are established between the sliding windows at any time to form a graph evolution sequence across time; In the path embedding representation generation stage, the graph attention network in the graph neural network is used as the path embedder to perform path representation based on attention weights for all causal paths in the structural causal graph, and dynamically adjust the expression ability of key nodes in different paths to the overall path contribution; According to the structural topological form, path activation intensity distribution, and edge weight evolution trajectory of the set of causal paths extracted from the graph embedding representation, all paths are aggregated into stable domains through a clustering algorithm, and further, the stable domains are mapped to typical patterns of system interaction failures. The typical patterns include the coupling imbalance pattern between the excitation system and the main grid voltage disturbance, the mutual interference pattern between the excitation system and the power system stabilizer, the dominant pattern of the synchronous failure path, and the multi-machine parallel excitation current offset circulation pattern.

5. A method for predicting faults in the excitation system of a hydropower generator set according to claim 1, characterized in that, Performing graph attention modeling on the evolution trends and instability boundaries of key parameters in the fault causal chains of different typical patterns included in the causal path network, and dynamically identifying the path characteristics of the excitation system entering the fault critical state, specifically including: Based on the structural topological results of each typical fault causal chain, key parameters that affect path activation intensity and respond to system stability changes are extracted from the fusion parameter set. The determination principle is based on the graph betweenness centrality of nodes, the cumulative weight of causal edges, and the activation frequency statistics in the historical evolution path, forming a local path subgraph constructed with key parameters as the center; In the path graph structure construction stage, based on the typical causal paths, a sub-path structure containing key parameter nodes is extracted, and the effective causal connections and actual activation paths between parameters are retained. At the same time, a time dimension embedding is introduced by attaching a timestamp code or a time interval variable to each node, forming a dynamic graph structure with temporal information; Using a graph attention network to train the constructed dynamic graph structure, and assigning dynamic attention weights to the edges between each node and its adjacent nodes. The graph attention network learns the key parameter nodes that contribute the most to the path activation probability in different operating stages and different disturbance environments by minimizing the path activation prediction error function during the training process; In the instability boundary modeling stage, based on the evolution trends of key parameter nodes, combined with historical fault data and prior knowledge of boundary states, an instability boundary determination model is constructed. The instability boundary determination model uses a sliding statistical window and a dynamic threshold discrimination mechanism to evaluate multiple indicators such as the change amplitude, direction consistency, and trend reversal points of key parameters in a continuous time period, extracts critical state characteristics, and embeds the critical state characteristics into the node attributes; Based on the path representation vectors output by the graph attention network, combined with the determination results of the instability boundary determination model, the activity levels of all paths are sorted, and paths that are close to the instability boundary, have high attention aggregation, and have a history of frequent activation records in the current time period are extracted as critical path characteristics. The critical path characteristics include the causal chain relationship between parameters, the activation weight of the path as a whole, the starting and ending parameters of the path, and the evolution slope and direction change information of the key nodes in the path.

6. A method for predicting faults in the excitation system of a hydropower generator set according to claim 1, characterized in that, Based on the coupling relationship between the excitation system, synchronous generator, power system stabilizer, and the main grid, a generalized state observation vector including linkage parameters is constructed, specifically including: Collect the first operating parameters in the excitation system in real time. The first operating parameters include excitation voltage, excitation current, rectifier bridge output voltage, excitation winding temperature, and the offset between the excitation voltage regulator reference signal and the actual output signal; Synchronously collect the second operating parameters in the synchronous generator. The second operating parameters include the amplitude of the synchronous generator terminal voltage, the phase angle of the synchronous generator voltage, the rotor angular velocity, the change rate of the excitation magnetic flux, the three-phase components of the stator current, and the power factor factor; Collect the third operating parameters of the power system stabilizer. The third operating parameters include the gain adjustment coefficient, the modulation signal frequency distribution, the output signal phase delay, and the input signal amplitude change slope; Introduce the fourth operating parameters of the main power grid. The fourth operating parameters include the instantaneous voltage offset, the grid short-circuit ratio, the grid frequency volatility, the grid impedance angle, and the equivalent reactance value; After aligning the first operating parameters, the second operating parameters, the fourth operating parameters, and the fourth operating parameters at a unified clock frequency, construct the generalized state observation vector.

7. A fault prediction method for the excitation system of a hydropower generator set according to claim 1, characterized in that, After performing graph attention modeling on the evolution trends and instability boundaries of the key parameters in the fault causal chains of different typical modes included in the causal path network, and dynamically identifying the path characteristics of the excitation system entering the fault critical state, the method further includes: Based on the path characteristics as the input basis, jointly construct the state representation vector of the excitation system by combining the path activation weight, the evolution slope of the key parameters in the path, the direction change characteristics, the proximity to the instability boundary, and the current operating parameters of the main control system; In the action space design stage, construct an action combination set around the adjustable control variables of the excitation system. Each action is represented as a parameter combination vector, and all actions constitute a discrete or continuous action space, covering all possible regulation operations that may affect the stability of the excitation system. The action combination set includes the upper and lower limit setting values of the excitation voltage, the target value of the excitation current, the gain factor of the excitation regulator, the voltage reference signal switching mode, the coordinated modulation coefficient of the power system stabilizer, and the excitation mode switching instruction; With the goal of delaying or avoiding fault triggering, design the reward as a reward function weighted by the negative activation degree, the proximity to the instability boundary, and the adjustment cost. At each time step, if the system state is far from the critical path, the reward value increases; if the path activation increases or the system approaches the instability boundary, the reward value decreases; if the execution of an action causes control oscillation or exceeds the physically bearable range, a penalty term is given; if the action successfully returns the system state to the safe domain, the maximum positive reward is given.

8. A fault prediction device for the excitation system of a hydropower generator set, characterized in that, The device is used to execute a method for predicting faults in the excitation system of a hydropower generator set according to any one of claims 1-7. The device includes an acquisition module (201), a processing module (202), and an output module (203), where: The acquisition module (201) is used to construct a generalized state observation vector including linkage parameters based on the coupling relationship between the excitation system, the synchronous generator, the power system stabilizer, and the main power grid; The processing module (202) is configured to continuously sample the generalized state observation vector through a sliding time window to form an operating parameter observation tensor; The processing module (202) is configured to fuse the operating parameter observation tensor and the boundary constraint type operating parameters of the generator to construct a fault observation tensor for multi-source operating parameter fusion; The processing module (202) is configured to introduce a structural causal graph in the modeling layer, establish a causal path network based on causal reasoning and topological time series embedding for the fault observation tensor, and divide the system interaction faults predicted by the excitation system into multiple typical modes; The output module (203) is configured to perform graph attention modeling on the evolution trends and instability boundaries of the key parameters in the fault causal chains of different typical modes included in the causal path network, and dynamically identify the path features for the excitation system to enter the fault critical state.

9. An electronic device, characterized in that, It includes a processor (301), a communication bus (302), a user interface (303), a network interface (304), and a memory (305). The memory (305) is used to store instructions. Both the user interface (303) and the network interface (304) are used to communicate with other devices. The communication bus (302) is used to realize the connection and communication between the components in the electronic device. The processor (301) is used to execute the instructions stored in the memory (305) so that the electronic device executes the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions, and when the instructions are executed, the method according to any one of claims 1-7 is executed.

Images