A method and device for predicting faults in the excitation system of a hydropower station generator unit

Abstract

This application provides a method and apparatus for predicting faults in the excitation system of a hydropower station generator unit, relating to the technical field of smart grids. The method includes: continuously sampling a generalized state observation vector through a sliding time window to form an operating parameter observation tensor; fusing the operating parameter observation tensor with the generator's boundary constraint-type operating parameters to construct a fault observation tensor fused from multiple operating parameters; introducing a structural causal graph at the modeling layer to establish a causal path network based on causal reasoning and topological temporal embedding for the fault observation tensor, classifying the predicted system interaction faults encountered by the excitation system into multiple typical modes; and performing graph attention modeling on the evolution trends and instability boundaries of key parameters in the fault causal chains of different typical modes contained in the causal path network to dynamically identify the path characteristics of the excitation system entering a critical fault state. This application can predict interaction-type faults in the excitation system of a hydropower station generator unit.

Description

Technical Field

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

[0002] The excitation system of a hydropower station generator unit is a key component for ensuring power output quality, maintaining voltage stability, and achieving dynamic response of the power system. Its design, control, and operation strategies are evolving with increasing complexity in the context of current energy system automation and intelligence. The core function of the excitation system is to provide DC excitation current to the rotor of the synchronous generator to establish the required magnetic field. Currently, the mainstream excitation system types include static excitation systems, rotating excitation systems, and composite excitation systems. Among these, static excitation systems are widely used in large hydropower units due to their advantages of high control precision, fast response speed, and convenient maintenance.

[0003] Interactive faults in the excitation system encompass complex phenomena such as poor coupling of voltage disturbances between the excitation system and the main grid, improper coordination between the excitation system and the power system stabilizer (PSS), synchronization failure, and abnormal excitation circulating current in multi-unit parallel operation. These faults often occur during unit switching, rapid paralleling, or large grid disturbances, exhibiting suddenness and interconnectivity, making fault diagnosis difficult and risk boundaries uncertain. Especially in high-impedance isolation structures employing static excitation systems, an anomaly on the auxiliary power supply side may lead to the failure of the system's self-regulation capability, thereby causing excitation disconnection. Therefore, a method is needed to predict interactive faults in the excitation system of hydropower station generator units. Summary of the Invention

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

[0005] The first aspect of this application provides a method for predicting faults in the excitation system of a hydropower station generator unit, the method comprising:

[0006] Based on the coupling relationship between the excitation system, synchronous generator, power system stabilizer and main grid, a generalized state observation vector including linkage parameters is constructed.

[0007] The generalized state observation vector is continuously sampled using a sliding time window to form the runtime parameter observation tensor.

[0008] By fusing the observed operating parameters tensor with the boundary constraint class operating parameters of the generator, a fault observation tensor based on multi-source operating parameters is constructed.

[0009] In the modeling layer, a structural causal graph is introduced to establish a causal path network based on causal reasoning and topological temporal embedding for the fault observation tensor, and the system interaction faults predicted by the excitation system are divided into multiple typical modes.

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

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

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

[0013] In each of the sliding time windows, the generalized state observation vector is sampled hourly at a globally unified sampling frequency 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 fixed dimensions, containing linkage parameters from four subsystems: the excitation system, the synchronous generator, the power system stabilizer, and the main grid.

[0014] Within each sliding time window, a two-dimensional matrix is ​​formed by arranging the generalized state observation vector sequence in time, where rows represent the time evolution dimension and columns represent different linkage parameter dimensions.

[0015] In the time dimension, the two-dimensional matrix is ​​advanced sequentially according to the sliding step size, and the sampling process is repeated to gradually form a set of state observation sequences under multiple sliding time windows;

[0016] By stacking the state observation sequences sampled by all sliding time windows in the window sequence dimension, a three-dimensional runtime parameter observation tensor is constructed. The three-dimensional structure of the runtime 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 step of fusing the operating parameter observation tensor and the boundary constraint class operating parameters of the generator to construct a fault observation tensor based on multi-source operating parameter fusion specifically includes:

[0018] For the typical boundary states faced by the synchronous generator during grid-connected operation, representative boundary constraint operation parameters are extracted. These boundary constraint operation parameters include the maximum excitation voltage limit factor, minimum excitation current constraint value, synchronous generator voltage reference switching status flag, parallel synchronization failure indication signal, active load jump rate limit, system frequency over-limit alarm status, auxiliary power supply instantaneous power outage duration and recovery delay indicator.

[0019] Since most boundary constraint-type operating parameters are triggered and irregular, boundary state information that overlaps with the window timestamp is extracted within the time interval corresponding to the sliding time window, and aggregation processing is performed within the window. Furthermore, based on the time synchronization mechanism of the main control system, the boundary constraint-type operating parameters and the operating parameter observation tensor are time aligned.

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

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

[0022] Based on the above technical solutions, preferably, the introduction of a structural causal graph at the modeling layer establishes a causal path network based on causal inference and topological temporal embedding for the fault observation tensor, classifying the predicted system interaction faults encountered by the excitation system into multiple typical modes, specifically including:

[0023] During the causal relationship discovery phase, 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 among the multiple time series sub-vectors, within 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 given the past state of the first time series vector. If it is true, a directed edge from the first time series vector to the second time series vector is constructed. At the same time, the information gain entropy criterion is introduced to filter out invalid edges with causal strength lower than the statistical threshold.

[0025] In the topology graph structure construction stage, the set of fault observation tensors is used as the graph node set, and the multiple directed edges are used as the edge set to construct the original graph topology structure of the structural causal graph.

[0026] A node attribute embedding mechanism is introduced to inject each graph node with its operational physical attributes, measurement accuracy, boundary sensitivity, and subsystem category, forming an attribute node graph with multiple semantics;

[0027] The structural causal graph for each sliding window is constructed separately, and lateral connections are established between sliding windows to form a time-series graph evolution sequence.

[0028] 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 expressive ability of key nodes in different paths to the overall path contribution.

[0029] Based on 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 structural domains through a clustering partitioning algorithm. Furthermore, the stable structural domains are mapped to typical modes of system interactive faults. The typical modes include the excitation system and main grid voltage disturbance coupling imbalance mode, the excitation system and power system stabilizer mutual interference mode, the synchronization failure path dominance mode, and the multi-machine parallel excitation current offset circulating current mode.

[0030] Based on the above technical solutions, preferably, the step of performing graph attention modeling on the evolution trends and instability boundaries of key parameters 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 critical fault state, specifically includes:

[0031] Based on the structural topology results of each typical fault causal chain, key parameters affecting the path activation intensity and response system stability changes are extracted from the fusion parameter set. The judgment principle is based on the graph betweenness centrality of nodes, cumulative weight of causal edges, and activation frequency statistics in historical evolution paths, forming a local path subgraph constructed with key parameters as the center.

[0032] In the path graph structure construction stage, based on typical causal paths, sub-path structures containing key parameter nodes are extracted, valid causal connections between parameters and actual activation paths are preserved, and time dimension embedding is introduced. By attaching timestamp encoding or time interval variables to each node, a dynamic graph structure with time sequence information is formed.

[0033] The constructed state graph structure is trained using a graph attention network. Dynamic attention weights are assigned to the edges between each node and its neighboring nodes. During training, 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] The instability boundary modeling stage is based on the evolution trend of key parameter nodes. It combines historical fault data and prior knowledge of boundary states to construct an instability boundary judgment model. The instability boundary judgment model adopts a sliding statistical window and dynamic threshold discrimination mechanism to evaluate the change amplitude, direction consistency and trend reversal point of key parameters in a continuous time period, extract critical state features, and embed the critical state features into node attributes.

[0035] Based on the path representation vector output by the graph attention network, and combined with the judgment result of the instability boundary judgment model, the activity level of all paths is sorted, and paths that are close to the instability boundary, have high attention concentration and frequent historical activation records in the current time period are extracted as critical path features. The critical path features include the causal chain relationship between parameters, as well as the overall activation weight of the path, the path start and end parameters, and the evolution slope and direction change information of key nodes in the path.

[0036] Based on the above technical solutions, preferably, the construction of a generalized state observation vector, including linkage parameters, is based on the coupling relationship between the excitation system, synchronous generator, power system stabilizer, and main grid. Specifically, this includes:

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

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

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

[0040] A fourth operating parameter is introduced for the main power grid, which includes instantaneous voltage offset, grid short-circuit ratio, grid frequency fluctuation rate, grid impedance angle and equivalent reactance value;

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

[0042] Based on the above technical solutions, 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 contained in the causal path network, and dynamically identifying the path characteristics of the excitation system entering the critical fault state, the method further includes:

[0043] Using the path features as input, the path activation weights, the evolution slope of key parameters in the path, the direction change characteristics, the proximity of the instability boundary, and the current operating parameters of the main control system are jointly constructed into a state representation vector of the excitation system.

[0044] In the action space design phase, an action combination set is constructed around the adjustable control variables of the excitation system. Each action is represented as a parameter combination vector. All actions constitute a discrete or continuous action space, covering all control 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 power system stabilizer cooperative modulation coefficient, and the excitation mode switching command.

[0045] With the goal of delaying or avoiding fault triggering, the reward is designed as a reward function weighted by the degree of negative activation, the proximity value 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 action causes control oscillations or exceeds the physical tolerance range, a penalty is given; if the action successfully returns the system state to the safe domain, the maximum positive reward is given.

[0046] A second aspect of this application provides a fault prediction device for the excitation system of a hydropower station generator set. The device is used to execute any of the fault prediction methods for the excitation system of a hydropower station generator set described above. The device includes an acquisition module, a processing module, and an output module, wherein:

[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, synchronous generator, power system stabilizer and main grid.

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

[0049] The processing module is used to fuse the operating parameter observation tensor and the boundary constraint class operating parameters of the generator to construct a fault observation tensor fused from multiple operating parameters.

[0050] The processing module is used to introduce a structural causal graph at the modeling layer, establish a causal path network based on causal reasoning and topological temporal embedding for the fault observation tensor, and classify the system interaction faults predicted by the excitation system into multiple typical modes.

[0051] The output module is used to perform graph attention modeling on the evolution trend and instability boundary of each key parameter in the fault causal chain of different typical modes contained in the causal path network, and dynamically identify the path characteristics of the excitation system entering the fault critical state.

[0052] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the foregoing.

[0053] A fourth aspect of this application provides a computer-readable storage medium storing instructions that, when executed, perform the method described in any of the preceding descriptions.

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

[0055] 1. This application constructs a generalized state observation vector covering the coupling relationship between the excitation system, synchronous generator, power system stabilizer, and main grid. It combines this vector with a sliding time window to form a high-dimensional operating parameter observation tensor and further integrates boundary constraint-type operating parameters to form a multi-source fusion expression, thereby achieving full-dimensional dynamic modeling of the system state. At the modeling layer, a structural causal graph is introduced, and a causal path network is constructed by combining causal reasoning and topological temporal embedding to conduct in-depth analysis of the causal chains under typical system interactive fault modes. Finally, a graph attention mechanism is used to model the evolution trend and instability boundary of key parameters in the fault path, dynamically identifying the path characteristics under the critical fault state, thereby achieving accurate prediction and evolution trend identification of interactive faults in the excitation system of hydropower station generator units.

[0056] 2. This application introduces a sliding time window mechanism to continuously sample the generalized state observation vector, constructing a three-dimensional structural operation parameter observation tensor with time window sequence dimension, time evolution dimension, and linkage parameter dimension. This tensor can fully capture the dynamic response process and linkage evolution characteristics of the excitation system and its coupled system at different time scales, thereby providing high temporal resolution and multivariate coupling expression capability for subsequent fault identification, causal modeling, and state prediction, and realizing high-precision dynamic modeling and abnormal trend perception of the system's operating state.

[0057] 3. This application integrates boundary constraint class operating parameters with operating parameter observation tensors and introduces a parameter self-attention mapping mechanism to achieve weighted encoding of boundary state information. While maintaining the original system dynamic feature expression ability, it effectively enhances the ability to characterize and identify the operating boundary of synchronous generators under extreme conditions. This results in the construction of a multi-source fusion fault observation tensor with higher semantic sensitivity and boundary awareness, providing a solid data foundation and structural guarantee for accurate modeling of system interactive faults and critical state prediction.

[0058] 4. This application introduces a structural causal graph to explicitly model the dynamic causal relationships between multidimensional fusion parameters in the fault observation tensor. It also constructs a causal path network by combining topological structure and time series features. With the support of the graph attention mechanism of the graph neural network, it achieves high-precision embedding representation and semantic extraction of key causal paths. Then, based on the path structure, activation intensity and edge weight evolution trajectory, it classifies and identifies the system interaction faults that the excitation system may encounter, effectively classifying them into multiple typical fault modes. This significantly improves the modeling ability, interpretability and pattern prediction effect of complex linkage faults.

[0059] 5. This application uses graph attention modeling on key parameters in the fault causal path network, combined with path structure features, time evolution information, and boundary proximity, to dynamically identify high-risk path features that can characterize the critical state of the excitation system, achieving early warning and accurate characterization of fault evolution trends. By introducing an instability boundary judgment model, the trend changes of key parameters and boundary features are embedded into the graph node attributes, significantly enhancing the sensitivity and predictive power of path characterization, thereby effectively improving the ability to identify and pre-control the critical state before the triggering of interactive faults in the excitation system of hydropower station generator units. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a method for predicting faults in the excitation system of a hydropower station generator unit, as disclosed in an embodiment of this application.

[0061] Figure 2 This is a schematic diagram of a module of a fault prediction device for the excitation system of a hydropower station generator unit disclosed in an embodiment of this application;

[0062] Figure 3 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application.

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

[0064] 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 with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0065] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0066] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0067] The excitation system of a hydropower station generator unit is a core module for maintaining voltage stability and ensuring power output quality. Especially with the rapid advancement of automation and intelligence, its structure and control are becoming increasingly complex. While mainstream structures, represented by static excitation systems, possess high precision and high response characteristics, they still face interactive fault risks in actual operation. These risks include voltage imbalance coupling with the main grid, abnormal coordination with power system stabilizers, synchronization failures, and abnormal excitation circulating currents in multi-unit parallel operation. These faults often occur during mode switching, grid disturbances, or high-load parallel operation, exhibiting suddenness, interconnectedness, and unpredictability. Particularly in high-impedance isolation structures, auxiliary power supply anomalies can easily lead to system regulation failures. Therefore, it is urgent to develop fault prediction methods for these interactive mechanisms to improve the system's forward-looking assurance capabilities for stable operation.

[0068] This embodiment discloses a method for predicting faults in the excitation system of a hydropower station generator set, referring to... Figure 1 This includes the following steps S110-S150:

[0069] S110 constructs a generalized state observation vector that includes linkage parameters, based on the coupling relationship between the excitation system, synchronous generator, power system stabilizer and main grid.

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

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

[0072] Specifically, the real-time acquisition of the primary operating parameters of the excitation system involves deploying high-precision sensing units and digital measurement and control devices at key locations within the excitation system to collect excitation voltage, excitation current, rectifier bridge output voltage, excitation winding temperature, and the offset between the reference signal and the actual output signal of the excitation voltage regulator. The excitation voltage and current are extracted in real-time using voltage transformers and Hall effect current sensors. The rectifier bridge output voltage is obtained through sampling via a multi-channel analog interface linked to the main control system. The excitation winding temperature is measured non-contactly using a resistance temperature detector (RTD) or fiber optic temperature sensor. The reference and actual signals of the excitation voltage regulator are directly output from the regulator's internal processing module. The offset is calculated in real-time using differential logic to comprehensively capture the dynamic adjustment capability and thermal stability of the excitation system.

[0073] Synchronous acquisition of the second operating parameter in a synchronous generator refers to the installation of current, voltage, high-speed position, and magnetic flux density measuring devices on the output side and stator winding terminals of the synchronous generator. These devices are used to acquire the synchronous generator terminal voltage amplitude, voltage phase angle, rotor angular velocity, excitation flux linkage rate of change, three-phase components of stator current, and power 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 flux linkage rate of change is estimated by combining the rotor induced voltage and stator back electromotive force. The three-phase components of stator current are synchronously acquired by three current transformers and then reconstructed using Clark transformation. The power factor is calculated in real time from the ratio of active to reactive power, thus accurately reflecting the electromagnetic behavior response characteristics of the synchronous generator under controlled excitation.

[0074] The third set of operating parameters for power system stabilizers refers to parameters extracted from the stabilizer's control link, including its gain adjustment coefficient, modulation signal frequency distribution, output signal phase delay, and input signal amplitude change slope. The gain adjustment coefficient is directly read from the stabilizer's built-in parameter register. The modulation signal frequency distribution is derived from the frequency domain transformation of the input signal. The output signal phase delay is automatically calculated based on the input-output timing difference. The input signal amplitude change slope is extracted using the sliding window differential method to obtain the continuous amplitude change rate. These parameters describe the power system stabilizer's control capability under dynamic small disturbances and low-frequency oscillation responses, and its impact on the excitation system control path. They are crucial information sources for modeling the voltage regulation-frequency stabilization composite mechanism.

[0075] The introduction of a fourth operating parameter for the main power grid refers to the acquisition of instantaneous voltage deviation, short-circuit ratio, frequency fluctuation rate, impedance angle, and equivalent reactance value of the main power grid through real-time data extraction from synchronous measurement units, wide-area monitoring and control nodes, and protection measurement devices deployed at the grid connection interface. The instantaneous voltage deviation is obtained by collecting the raw signal from the voltage measurement terminal and comparing it with a reference value. The short-circuit ratio is determined through online impedance assessment and calculation of the grid connection capacity ratio. The frequency fluctuation rate is obtained by differentiating the frequency change over a continuous time period. The impedance angle and equivalent reactance value are obtained through impedance measurement model fitting calculations, used to characterize the electrical load coupling effect of the main power grid's dynamic environment on the synchronous generator excitation system.

[0076] Finally, the first, second, third, and fourth operating parameters are synchronously sampled under the drive of the control system's master clock or a unified GPS timestamp. A data synchronization interface and a timing correction module are used to achieve unified alignment of multi-source data, forming a complete, structurally consistent, and time-synchronized generalized state observation vector. This observation vector integrates operating information from four subsystems—the excitation system, synchronous generator, power system stabilizer, and main grid—in the parameter dimension, and satisfies global consistency in the time dimension. This provides a unified multi-source time-series representation foundation for subsequent construction of operating parameter observation tensors, extraction of path evolution features, and prediction of system interaction faults.

[0077] S120 continuously samples the generalized state observation vector through a sliding time window to form the operating parameter observation tensor.

[0078] In one possible implementation, a sliding time window is used to continuously sample the generalized state observation vector to form an operating parameter observation tensor. Specifically, this includes: determining the time length and sliding step size of the sliding time window; within each sliding time window, sampling the generalized state observation vector hourly at a globally unified sampling frequency to obtain a sequence of generalized state observation vectors corresponding to multiple consecutive time points within the sliding time window. Each time point's generalized state observation vector is a fixed-dimensional vector containing linkage parameters from four subsystems: the excitation system, synchronous generator, power system stabilizer, and main grid; forming a two-dimensional matrix within each sliding time window, composed of the generalized state observation vector sequence arranged chronologically, where rows represent the time evolution dimension and columns represent different linkage parameter dimensions; advancing the two-dimensional matrix sequentially along the time dimension according to the sliding step size, repeating the sampling process to gradually construct a set of state observation sequences under multiple sliding time windows; and stacking the set of state observation sequences sampled from all sliding time windows along the window sequence dimension to 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, the time evolution dimension, and the linkage parameter dimension, respectively.

[0079] Specifically, determining the length and step size of the sliding time window refers to rationally selecting sliding time window parameters for observing the system state evolution based on the typical dynamic response cycles of the four subsystems: the excitation system, synchronous generator, power system stabilizer, and main grid, before constructing the operational parameter observation tensor. The length determines the continuous time interval covered by each sampling period, typically set to a time period that can completely encompass a small disturbance or an excitation regulation response cycle, such as 5 to 30 seconds, to capture the dynamic changes of the linkage parameters during disturbance propagation. The step size is defined as the advance distance between adjacent sliding time windows, and its setting should be between milliseconds and seconds to balance dynamic tracking capability and computational resource utilization, ensuring effective information coverage without losing high-frequency characteristics.

[0080] Within each sliding time window, the generalized state observation vector is sampled hourly at a globally unified sampling frequency. Specifically, this means performing periodic sampling according to the excitation system's master control clock or a GPS clock synchronized with it, discretizing and 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 dimension, containing all linkage parameters from the four subsystems: the excitation system, synchronous generator, power system stabilizer, and main grid. This ensures that each sampling can reconstruct the information structure of the system's full state space, providing high temporal resolution support for the state evolution trajectory within the sliding time window.

[0081] Within each sliding time window, a two-dimensional matrix is ​​formed, consisting of the generalized state observation vector sequence arranged chronologically. Specifically, all generalized state observation vectors sampled within the sliding time window are arranged from top to bottom in chronological order, forming a two-dimensional structure matrix with a time evolution dimension multiplied by a linkage parameter dimension. Each row of this matrix represents the set of states of all subsystem linkage parameters at a certain time point, and each column represents the evolution trajectory of a certain linkage parameter within the entire time window. This constructs a structured data entity capable of describing temporal change patterns, providing basic data fragments for subsequent path modeling.

[0082] The two-dimensional matrix is ​​advanced sequentially along the sliding step size in the time dimension. The repeated sampling process means that after the initial sliding time window sampling is completed, the sliding time window length remains unchanged, and the matrix is ​​slid forward by one sliding step unit. The generalized state observation vectors for the new time period are then resampled, and the next two-dimensional matrix is ​​constructed. This process is repeated until the entire target time interval is covered. Each two-dimensional matrix generated by advancing the sliding time window retains the temporal structure of the system state within that time period, resulting in a dynamic observation sequence with high overlap of system states under continuous sliding time windows.

[0083] A three-dimensional operational parameter observation tensor is constructed by stacking the state observation sequences sampled within all sliding time windows along the window sequence dimension. This involves sequentially arranging the two-dimensional matrices generated within all the aforementioned sliding time windows and stacking them along the time window sequence dimension to generate a complete three-dimensional tensor structure. The three dimensions of this operational parameter observation tensor are the time window sequence dimension, the time evolution dimension, and the linkage parameter dimension. It has the ability to model the long-term and short-term linkages of fault evolution paths and can capture the multi-scale feature evolution process of system state on a continuous time axis. It serves as the sole input data structure foundation for subsequent structural causal graph construction and causal path identification.

[0084] S130, integrates the operating parameter observation tensor and the boundary constraint class operating parameters from the synchronous generator to construct a fault observation tensor fused from multi-source operating parameters.

[0085] In one possible implementation, a fault observation tensor based on multi-source operating parameters is constructed by fusing the operating parameter observation tensor and the generator's boundary constraint operating parameters. Specifically, this includes: extracting representative boundary constraint operating parameters for typical boundary states faced by synchronous generators during grid-connected operation. These boundary constraint operating parameters include the maximum excitation voltage limit factor, minimum excitation current constraint value, synchronous generator voltage reference switching status flag, parallel synchronization failure indication signal, active load jump rate limit, system frequency over-limit alarm status, duration of auxiliary power supply instantaneous power outage, and recovery delay identifier. Given that boundary constraint operating parameters are mostly triggerable and irregular, boundary state information overlapping with the window timestamp is extracted within the time interval corresponding to the sliding time window, and in-window aggregation processing is performed. Further, based on the main control system... A unified time synchronization mechanism performs time alignment processing on the boundary constraint class operating parameters and the operating parameter observation tensor. After completing the time alignment processing, the boundary constraint class operating parameters are structurally reorganized to construct a tensor dimension compatible with the structure of the operating parameter observation tensor. By expanding the linkage parameter dimension of the operating parameter observation tensor, multiple encoding channels representing the boundary constraint class operating parameters are added, so that the three-dimensional structure's 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. The fusion parameter dimension includes the joint expression of linkage parameters and boundary constraint class operating parameters. A parameter self-attention mapping mechanism is introduced to weight the encoding of the importance of different boundary constraint class operating parameters. After completing the fusion encoding, the fault observation tensor of multi-source operating parameters is output.

[0086] Specifically, for the typical boundary states faced by synchronous generators during grid-connected operation, representative boundary constraint operating parameters are extracted. This refers to identifying and selecting operating state indicators that characterize the operational stability boundaries and limit constraints based on system modeling of the synchronous generator's operational limit behavior. Boundary constraint operating parameters mainly include the maximum excitation voltage limit factor and minimum excitation current constraint value of the synchronous generator's excitation system, used to reflect the upper and lower limits of excitation capability; voltage reference switching status flags and parallel synchronization failure indication signals, used to identify regulation mode switching and grid synchronization anomalies; active load jump rate upper limit and system frequency over-limit alarm status, used to indicate the direct impact of load dynamic disturbances on the synchronous generator; and auxiliary power supply instantaneous power outage duration and its recovery delay indicator, used to capture changes in the power supply support capability boundary under high-impedance isolation structures. Most of these boundary constraint operating parameters originate from protection device records, control system event logs, or status registers, and possess high sensitivity and state segmentation characteristics in fault prediction tasks.

[0087] Since most boundary constraint parameters are trigger-dependent and irregular, extracting boundary state information overlapping with the window timestamp within the time interval corresponding to the sliding time window and performing in-window aggregation processing involves matching each sliding time window with the boundary event sequence for temporal overlap, extracting the corresponding boundary state flags within the successfully matched time period, and transforming them into a continuous encoded representation through an in-window aggregation strategy. For example, for discrete event boundary constraint parameters, methods such as trigger frequency normalization, duration weighting, or trigger time centering are used to map them into a continuous numerical representation between 0 and 1; for quantitative constraint parameters, methods such as in-window maximum value, average value, or slope trend are used for aggregation encoding to enhance their temporal continuity. The above operations ensure that boundary constraint parameters have the same granularity in the time domain as the sliding time window.

[0088] Further time alignment processing of the boundary constraint class operating parameters and the operating parameter observation tensor based on the main control system time synchronization mechanism refers to using the system's unified clock signal as the time reference and using a synchronous timing interpolation mechanism to accurately align the sampling points in the boundary constraint class operating parameters and the operating parameter observation tensor. This ensures that within each sliding time window, the sampling points of the boundary parameters and linkage parameters on the time axis have a one-to-one correspondence, thereby eliminating the asynchronous sequence problem caused by sampling frequency differences, clock drift, or timestamp misalignment due to different sampling devices.

[0089] After completing the time alignment process, the boundary constraint class operating parameters are structurally reorganized to construct a tensor dimension compatible with the operating parameter observation tensor structure. This refers to embedding the aligned boundary constraint class operating parameters into a three-dimensional tensor structure. Specifically, this involves expanding the dimension of the operating parameter observation tensor along the linkage parameter dimension, adding several channels as expression channels for the boundary constraint class operating parameters. The dimension of each channel remains consistent with the time window sequence dimension and time evolution dimension of the operating parameter observation tensor. This expands the tensor from its original form (time window sequence dimension × time evolution dimension × linkage parameter dimension) 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 boundary constraint class operating parameter dimension. This structurally achieves a unified fusion of multi-source operating information.

[0090] Introducing a parameter self-attention mapping mechanism to weighted encode the importance of different boundary constraint class operating parameters involves constructing a lightweight parameter attention allocation module. Using the overall system state within the current sliding time window as context, a weight coefficient is assigned to each boundary constraint class operating parameter to express its importance for fault detection at that time. 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. It ultimately outputs a set of normalized weights, which are used to perform weighted summation on the boundary constraint class operating parameter channels, further forming a fusion representation feature with time-varying adaptive capabilities.

[0091] After completing the above weighted fusion encoding, the output fault observation tensor of 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, and retains 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. This provides a highly identifiable, highly representative, and highly generalizable input expression foundation for subsequent causal path modeling and system interaction fault prediction tasks.

[0092] S140 introduces a structural causal graph in the modeling layer, establishes a causal path network based on causal reasoning and topological temporal embedding for the fault observation tensor, and classifies the system interaction faults predicted by the excitation system into multiple typical modes.

[0093] In one possible implementation, a structural causal graph is introduced at the modeling layer. A causal path network based on causal inference and topological temporal embedding is established for the fault observation tensor. The predicted system interaction faults encountered by the excitation system are divided into several typical patterns. Specifically, this includes: in the causal relationship discovery stage, extracting time-series sub-vectors corresponding to each fusion parameter dimension from the fault observation tensor; for the first and second time-series vectors among the multiple time-series sub-vectors, examining 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; if so, constructing directed edges from the first time-series vector to the second time-series vector, while introducing an information gain entropy criterion to filter invalid edges with causal strength below a statistical threshold; in the topology graph construction stage, using the set of fault observation tensors as the graph node set and multiple directed edges as the edge set, constructing the original graph topology of the structural causal graph; introducing a node attribute embedding mechanism for each... Each graph node is injected with its operational physical attributes, measurement accuracy, boundary sensitivity, and subsystem category, forming an attribute node graph with multiple semantics. A separate structural causal graph is constructed for each sliding window, and lateral connections are established between sliding windows to form a time-series graph evolution sequence. In the path embedding representation generation stage, a graph attention network from a graph neural network is used as the path embedder to perform attention-weight-based path representation on all causal paths in the structural causal graph, dynamically adjusting the expressive ability of key nodes in different paths to contribute to the overall path. Based on the structural topology, path activation intensity distribution, and edge weight evolution trajectory of the causal path set extracted from the graph embedding representation, a clustering algorithm is used to aggregate all paths into stable structural domains. These stable structural domains are further mapped to typical modes of system interactive faults, including the excitation system and main grid voltage disturbance coupling imbalance mode, the excitation system and power system stabilizer mutual interference mode, the synchronization failure path-dominated mode, and the multi-machine parallel excitation current offset circulating current mode.

[0094] Specifically, in the causal relationship discovery phase, time-series sub-vectors corresponding to each fusion parameter dimension are extracted from the fault observation tensor. This involves slicing the fault observation tensor according to the fusion parameter dimensions, extracting the complete observation trajectory of each fusion parameter across all time window sequences and the time evolution dimension, thus obtaining multiple independent sets of time-series sub-vectors. Each time-series sub-vector represents the dynamic evolution process of a certain fusion parameter within a sliding time window, preserving the trend, jumps, lags, and steady-state characteristics of the system state over time, providing a raw input structure with temporal sequence and variable independence for subsequent causal analysis.

[0095] For the first and second time series vectors among the multiple time series sub-vectors, within 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 given the past state of the first time series vector. If so, a directed edge from the first time series vector to the second time series vector is constructed. At the same time, an information gain entropy criterion is introduced to filter invalid edges with causal strength below a statistical threshold. Specifically, the Granger causality test method is used to construct a conditional prediction model within each fixed sliding time window. The changes in the prediction residuals of the current state of the second time series vector before and after adding historical information of the first time series vector are compared. If the residuals are significantly reduced, a valid causal relationship is considered to exist, and a directed edge is constructed from the first vector to the second vector. To improve the sparsity and discriminativeness of the causal graph, the information gain entropy criterion is introduced on all potential causal edges to calculate the contribution of the causal edge to the uncertainty of the overall state. Weak causal relationships below the threshold are eliminated, and key influence paths with strong discriminative power are retained.

[0096] In the topology graph construction phase, the fusion parameter set of the fault observation tensor is used as the graph node set, and the multiple directed edges are used as the edge set to construct the original graph topology of the structural causal graph. Specifically, each fusion parameter is used as a uniquely identified graph node, and each directed edge established based on causal reasoning is used as a connection relationship between nodes. Under the combined 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 maintains consistency with the temporal distribution of the fault observation tensor, forming a local static topology within each sliding time window to capture the variable-driven mechanism within the system during a specific time period.

[0097] A node attribute embedding mechanism is introduced to inject each graph node with its operational physical attributes, measurement accuracy, boundary sensitivity, and subsystem category, forming an attribute node graph with multiple semantics. This involves embedding and integrating various auxiliary information of each graph node and its corresponding fusion parameters based on the topology structure. Examples include the unit, dimension, standard deviation of measurement error, response sensitivity coefficient under boundary conditions, and the subsystem identifier originating from the excitation system, synchronous generator, power system stabilizer, or main grid. This constitutes a composite node representation with multiple semantic information encompassing physics, measurement, and structure. This embedding mechanism provides semantic prior support for graph neural networks in subsequent path modeling, improving the expressive and interpretive capabilities of path learning.

[0098] A structural causal graph is constructed separately for each time-sliding window, and lateral connections are established between these windows to form a cross-time graph evolution sequence. Specifically, a structural causal graph is generated separately within each sliding time window, and the same node in adjacent windows is connected across windows according to their temporal evolution relationship, using the window sequence as a time reference, thus constructing a graph structure sequence with a time axis dimension. This graph evolution sequence structurally forms a graph time series model, supporting the modeling of the dynamic changes in causal paths over time, and providing continuous modeling support for identifying path activation intensity, fault propagation chains, and critical state formation mechanisms.

[0099] 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. The expressive ability of key nodes in different paths to contribute to the overall path is dynamically adjusted. Specifically, the graph attention network is used to assign dynamically learnable attention weights to each node in each path, which strengthens the expressive ability of nodes with high influence or high boundary sensitivity in the current system state. The path-level embedding vector is generated by aggregating the weighted representation vectors of different nodes, thereby forming an embedded path representation set with state awareness, path feature distinguishability and fault evolution direction.

[0100] Based on the structural topology, path activation intensity distribution, and edge weight evolution trajectory of the causal path set extracted from the graph embedding representation, a clustering algorithm is used to aggregate all paths into stable structural domains. These stable structural domains are then mapped to typical patterns of system interactive faults. This involves performing unsupervised clustering analysis on the embedded path set to identify subsets of causal paths with similar topological structures, activation patterns, and dynamic evolution laws. These subsets are defined as stable structural domains to characterize the causal path features of the system's operating state under a certain potential interactive fault type. Each stable structural domain is further mapped to a typical pattern of system interactive faults through pattern labeling, including the excitation system and main grid voltage disturbance coupling imbalance mode, the excitation system and power system stabilizer mutual interference mode, the synchronization failure path-dominated mode, and the multi-machine parallel excitation current offset circulating current mode. This achieves a closed-loop path-driven modeling of system interactive faults from initial observation to structural understanding.

[0101] S150 performs graph attention modeling on the evolution trend and instability boundary of each key parameter in the fault causal chain of different typical modes contained in the causal path network, and dynamically identifies the path characteristics of the excitation system entering the critical state of fault.

[0102] In one possible implementation, graph attention modeling is performed on the evolution trends and instability boundaries of key parameters in fault causal chains of different typical modes contained in the causal path network to dynamically identify the path characteristics of the excitation system entering the critical fault state. Specifically, this includes: extracting key parameters affecting the path activation intensity and response system stability changes from the fused parameter set based on the structural topology results of each typical fault causal chain. The determination principle is based on the graph betweenness centrality of nodes, the cumulative weight of causal edges, and the activation frequency statistics in historical evolution paths, forming a local path subgraph centered on the key parameters; in the path graph structure construction stage, based on the typical causal paths, extracting the sub-path structure containing key parameter nodes, retaining the effective causal connections between parameters and the actual activation paths, and introducing time dimension embedding by adding timestamp encoding or time interval variables to each node to form a dynamic graph structure with time sequence information; training the constructed graph structure using a graph attention network, assigning dynamic attention weights to the edges between each node and its neighboring nodes. During training, the network learns the key parameter nodes that contribute most to the path activation probability under different operating stages and perturbation environments by minimizing the path activation prediction error function. In the instability boundary modeling stage, based on the evolution trend of key parameter nodes and combined with historical fault data and prior knowledge of boundary states, an instability boundary judgment model is constructed. The instability boundary judgment model adopts a sliding statistical window and dynamic threshold discrimination mechanism to evaluate the magnitude of change, directional consistency, and trend reversal points of key parameters in continuous time periods using multiple indicators, extracting critical state features, and embedding the critical state features into node attributes. Based on the path representation vector output by the graph attention network and combined with the judgment results of the instability boundary judgment model, the activity level of all paths is ranked, and paths that are close to the instability boundary, have high attention concentration, and have frequent historical activation records in the current time period are extracted as critical path features. Critical path features include the causal chain relationship between parameters, the overall activation weight of the path, the path start and end parameters, and the evolution slope and directional change information of key nodes in the path.

[0103] Specifically, based on the structural topology of typical fault causal chains, key parameters influencing path activation intensity and response system stability changes are extracted from the fused parameter set. This is implemented as follows: In the constructed causal path network, all typical fault causal chains are traversed. The graph betweenness centrality metric of each node is calculated to identify parameter nodes with high information flow control in path propagation. Simultaneously, the weights of all causal edges associated with that node are accumulated to reflect its direct driving force on overall path activation. Combined with the node's activation frequency in historical evolution paths, its probability of participating in system instability under multiple operational disturbances is assessed. The comprehensive score of these three factors is used as the key parameter judgment index, thereby forming a local path subgraph centered on the key parameters, ensuring that the extracted key parameters possess structural influence, behavioral driving force, and evolutionary representativeness.

[0104] In the path graph structure construction phase, based on typical causal paths, sub-path structures containing key parameter nodes are extracted. Specifically, in each typical fault causal chain, all key parameter nodes are located, and their predecessor nodes and predecessor edges are traced back to extract the complete activation path segment in which they are located. At the same time, the directed causal connections between parameters in the path are preserved, and the minimum path subgraph covering the key parameters is constructed. On this basis, a time dimension embedding variable is introduced for each node, and its sampling timestamp or the time interval between the previous activation state is encoded as a time embedding feature. Together with the node's physical attributes, it constitutes the node embedding vector, thereby generating a dynamic graph structure with temporal attributes, which fully preserves the dynamic evolution trajectory and temporal dependencies of the causal path.

[0105] The constructed state graph structure is trained using a graph attention network. Specifically, the constructed key parameter path graph with temporal embedding is input into the graph attention network model. In each propagation layer of the graph neural network, learnable attention weights based on the similarity of node embedding vectors are applied to the edges between each node and its neighboring nodes. Multi-head attention mechanism is used to integrate multi-source information in the path. During training, the path activation prediction error is used as the target loss function. The historically labeled path activation state labels are used for supervised training to optimize the parameter update process. This allows the model to gradually learn the key parameter nodes that contribute the most to the path activation probability under different perturbation types and operating stages, as well as their temporal evolution characteristics. This improves the model's ability to identify high-risk paths and the accuracy of path dynamic representation.

[0106] By selecting multiple time windows in historical operational data where key parameters approach or enter the period before a failure, statistical features are extracted regarding the magnitude of change, consistency of change direction, and trend reversal points within consecutive time periods. A sliding window approach is used to calculate indicators such as dynamic standard deviation, slope change, and number of inflection points across multiple time periods, constructing a multi-dimensional state vector reflecting whether the parameter has entered a critical state. Based on the boundary state features extracted from historical failure samples, a dynamic threshold model is constructed to determine whether the key parameter has reached a critical state at the current moment. The judgment result is converted into a binary or probabilistic label and embedded 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 vector output by the graph attention network, and combined with the judgment results of the instability boundary determination model, the activity level of all paths is ranked, and paths that are close to the instability boundary, have high attention concentration, and have frequent historical activation records are extracted as critical path features within the current time period. Specifically, in each prediction period, the embedding vectors of all key parameter paths and the instability boundary state are jointly input to the path activation ranking module to jointly evaluate the activity intensity, instability probability, and importance of the paths. Based on the ranking results, the set of paths in the critical activation state or high-risk evolution state is selected as critical path features. Each critical path feature includes the path start parameter and end parameter identifiers, the list of key nodes in the path, the path activation weight value, and the slope change and direction reversal information of the key parameters, which are used to accurately describe the potential failure evolution path of the current system and serve as a structured input for the generation of failure prediction and control strategies.

[0108] In one possible implementation, after performing graph attention modeling on the evolution trends and instability boundaries of key parameters in different typical fault causal chains contained in the causal path network, and dynamically identifying the path characteristics of the excitation system entering the critical fault state, the method further includes: using the path characteristics as input, jointly constructing a state representation vector of the excitation system by combining the path activation weights, the evolution slope of key parameters in the path, the direction change characteristics, the proximity of 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, with each action represented as a parameter combination vector, and all actions constituting a discrete or continuous action space, covering all possible influences on the excitation system. The control operations for magnetic system stability include a set of actions such as upper and lower limit settings for excitation voltage, target value for excitation current, excitation regulator gain factor, voltage reference signal switching mode, power system stabilizer cooperative modulation coefficient, and excitation mode switching command. Aiming to delay or avoid fault triggering, the reward is designed as a reward function weighted by the degree of negative activation, proximity to the instability boundary, and regulation cost. At each time step, if the system state moves away from the critical path, the reward value increases; if path activation increases or the system approaches the instability boundary, the reward value decreases; if the action triggers control oscillations or exceeds the physical tolerance range, a penalty is applied; 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 input, the path activation weights, the evolution slope of key parameters in the path, directional change characteristics, proximity to the instability boundary, and the current operating parameters of the main control system are jointly constructed into a state representation vector for the excitation system. Specifically, after the graph attention modeling stage, all path features identified as critical paths are extracted from the current sliding time window. These features include the overall activation weight values ​​of the path, the slope of key parameters changing over time, the trend factor of parameter value changes, and the distance index between the current state and the preset instability boundary. These features are then combined with control variables such as excitation voltage, excitation current, rotor angular velocity, voltage reference source state, and power system stabilizer gain collected in real time from the main control system. After unified normalization, a high-dimensional state vector is constructed. This state representation vector possesses structural drive and physical interpretability, accurately expressing the current operating environment, risk level, and dynamic trends of the excitation system, providing state-aware input for policy generation in reinforcement learning.

[0110] In the action space design phase, an action combination set is constructed around the adjustable controllable variables of the excitation system. Specifically, based on the regulation capability boundary of the excitation system and the actual control channel, control variables that directly affect its response capability 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 setpoint of the excitation voltage, the lower limit setpoint 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 primary and backup references, the adjustment ratio of the power system stabilizer's cooperative modulation coefficient, and the switching instructions of the excitation mode between automatic control mode, constant current mode, and constant voltage mode. The above action combination includes both fine adjustment of continuous variables and mode switching of discrete states, forming a complete action set covering the entire regulation space of the system. This enables the strategy network to have sufficient control and selection capabilities during the strategy 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 weighted by the degree of negative activation, the proximity to the instability boundary, and the adjustment cost. Specifically, in each state transition step, the change between the current system state and the previous state is evaluated. If, after executing an action, the current system state shifts away from the critical path, the reward function increases, reflecting that the control operation has successfully achieved state regression. If the path activation degree increases or the distance to the instability boundary decreases, and the system approaches the fault region, a negative reward is given, indicating that the action exacerbates the risk. If the execution of an action causes the excitation voltage or excitation current to exceed the set upper or lower limits, or generates overshoot oscillations, a penalty term is added to avoid aggressive control. Conversely, if the action significantly reduces the activation intensity of the activation path and increases the stability slope of 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, achieving precise guidance and stable optimization of the policy network's learning direction.

[0112] This embodiment also discloses a fault prediction device for the excitation system of a hydropower station generator set. The device is used to execute any of the fault prediction methods for the excitation system of a hydropower station generator set as described above, with reference to... Figure 2 It includes an acquisition module 201, a processing module 202, and an output module 203, wherein:

[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, synchronous generator, power system stabilizer and main grid.

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

[0115] Processing module 202 is used to fuse the operating parameter observation tensor and the generator boundary constraint class operating parameters to construct a fault observation tensor fused from multiple operating parameters.

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

[0117] Output module 203 is used to perform graph attention modeling on the evolution trend and instability boundary of each key parameter in the fault causal chain of different typical modes contained 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 the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

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

[0120] The communication bus 302 is used to enable communication between these components.

[0121] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0122] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0123] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 305, and by calling data stored in the memory 305. Optionally, the processor 301 may be implemented using 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 a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 301 and may be implemented as a separate chip.

[0124] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory may include a non-transitory computer-readable storage medium. The memory 305 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory 305 may also be at least one storage device located remotely 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 method of predicting faults in the excitation system of a hydropower station generator set.

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

[0126] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0127] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0128] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

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

[0130] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0131] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). 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 the 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 cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory 305 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0132] This application also discloses a computer-readable storage medium storing instructions. When executed by one or more processors 301, these instructions cause an electronic device to perform one or more of the methods described in the above embodiments.

[0133] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and practical application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for predicting faults in the excitation system of a hydropower station generator unit, characterized in that, The method includes: Based on the coupling relationship between the excitation system, synchronous generator, power system stabilizer and main grid, a generalized state observation vector including linkage parameters is constructed. The generalized state observation vector is continuously sampled using a sliding time window to form the runtime parameter observation tensor. By fusing the observed operating parameters tensor with the boundary constraint class operating parameters of the generator, a fault observation tensor based on multi-source operating parameters is constructed. In the modeling layer, a structural causal graph is introduced to establish a causal path network based on causal reasoning and topological temporal embedding for the fault observation tensor, and the system interaction faults predicted by the excitation system are divided into multiple typical modes. Graph attention modeling is performed on the evolution trend and instability boundary of each key parameter in the fault causal chain of different typical modes contained in the causal path network to dynamically identify the path characteristics of the excitation system entering the critical state of fault. The introduction of a structural causal graph at the modeling layer establishes a causal path network based on causal inference and topological temporal embedding for the fault observation tensor, classifying the predicted system interaction faults encountered by the excitation system into multiple typical modes, specifically including: During the causal relationship discovery phase, 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 the multiple time series sub-vectors, within 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 given the past state of the first time series vector. If it is true, a directed edge from the first time series vector to the second time series vector is constructed. At the same time, the information gain entropy criterion is introduced to filter out invalid edges with causal strength lower than the statistical threshold. In the topology graph structure construction stage, the set of fault observation tensors is used as the graph node set, and the multiple directed edges are used as the edge set to construct the original graph topology structure of the structural causal graph. A node attribute embedding mechanism is introduced to inject each graph node with its operational physical attributes, measurement accuracy, boundary sensitivity, and subsystem category, forming an attribute node graph with multiple semantics; The structural causal graph for each sliding window is constructed separately, and lateral connections are established between sliding windows to form a time-series 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 representation based on attention weights for all causal paths in the structural causal graph, and dynamically adjust the expressive ability of key nodes in different paths to the overall path contribution. Based on 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 structural domains through a clustering partitioning algorithm. Furthermore, the stable structural domains are mapped to typical modes of system interactive faults. The typical modes include the excitation system and main grid voltage disturbance coupling imbalance mode, the excitation system and power system stabilizer mutual interference mode, the synchronization failure path dominance mode, and the multi-machine parallel excitation current offset circulating current mode.

2. The method for predicting faults in the excitation system of a hydropower station generator unit according to claim 1, characterized in that, The step of continuously sampling the generalized state observation vector through a sliding time window to form the runtime parameter observation tensor specifically includes: Determine the time length and sliding step size of the sliding time window; In each of the sliding time windows, the generalized state observation vector is sampled hourly at a globally unified sampling frequency 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 fixed dimensions, containing linkage parameters from four subsystems: the excitation system, the synchronous generator, the power system stabilizer, and the main grid. Within each sliding time window, a two-dimensional matrix is ​​formed by arranging the generalized state observation vector sequence in time, where rows represent the time evolution dimension and columns represent different linkage parameter dimensions. In the time dimension, the two-dimensional matrix is ​​advanced sequentially according to the sliding step size, and the sampling process is repeated to gradually form a set of state observation sequences under multiple sliding time windows; By stacking the state observation sequences sampled by all sliding time windows in the window sequence dimension, a three-dimensional runtime parameter observation tensor is constructed. The three-dimensional structure of the runtime parameter observation tensor corresponds to the time window sequence dimension, the time evolution dimension, and the linkage parameter dimension, respectively.

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

4. The method for predicting faults in the excitation system of a hydropower station generator unit according to claim 1, characterized in that, The method involves performing graph attention modeling on the evolution trends and instability boundaries of key parameters 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 critical fault state. Specifically, this includes: Based on the structural topology results of each typical fault causal chain, key parameters affecting the path activation intensity and response system stability changes are extracted from the fusion parameter set. The judgment principle is based on the graph betweenness centrality of nodes, cumulative weight of causal edges, and activation frequency statistics in historical evolution paths, forming a local path subgraph constructed with key parameters as the center. In the path graph structure construction stage, based on typical causal paths, sub-path structures containing key parameter nodes are extracted, valid causal connections between parameters and actual activation paths are preserved, and time dimension embedding is introduced. By attaching timestamp encoding or time interval variables to each node, a dynamic graph structure with time sequence information is formed. The constructed state graph structure is trained using a graph attention network. Dynamic attention weights are assigned to the edges between each node and its neighboring nodes. During training, 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. The instability boundary modeling stage is based on the evolution trend of key parameter nodes. It combines historical fault data and prior knowledge of boundary states to construct an instability boundary judgment model. The instability boundary judgment model adopts a sliding statistical window and dynamic threshold discrimination mechanism to evaluate the change amplitude, direction consistency and trend reversal point of key parameters in a continuous time period, extract critical state features, and embed the critical state features into node attributes. Based on the path representation vector output by the graph attention network, and combined with the judgment result of the instability boundary judgment model, the activity level of all paths is sorted, and paths that are close to the instability boundary, have high attention concentration and frequent historical activation records in the current time period are extracted as critical path features. The critical path features include the causal chain relationship between parameters, as well as the overall activation weight of the path, the path start and end parameters, and the evolution slope and direction change information of key nodes in the path.

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

6. The method for predicting faults in the excitation system of a hydropower station generator unit according to claim 1, characterized in that, After performing graph attention modeling on the evolution trends and instability boundaries of key parameters 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 critical fault state, the method further includes: Using the path features as input, the path activation weights, the evolution slope of key parameters in the path, the direction change characteristics, the proximity of the instability boundary, and the current operating parameters of the main control system are jointly constructed into a state representation vector of the excitation system. In the action space design phase, an action combination set is constructed around the adjustable control variables of the excitation system. Each action is represented as a parameter combination vector. All actions constitute a discrete or continuous action space, covering all control 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 power system stabilizer cooperative modulation coefficient, and the excitation mode switching command. With the goal of delaying or avoiding fault triggering, the reward is designed as a reward function weighted by the degree of negative activation, the proximity value 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 action causes control oscillations or exceeds the physical tolerance range, a penalty is given; if the action successfully returns the system state to the safe domain, the maximum positive reward is given.

7. A fault prediction device for the excitation system of a hydropower station generator set, characterized in that, The device is used to execute a fault prediction method for the excitation system of a hydropower station generator unit as described in any one of claims 1-6. The device includes an acquisition module (201), a processing module (202), and an output module (203), wherein: 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, synchronous generator, power system stabilizer and main grid. The processing module (202) is used to continuously sample the generalized state observation vector through a sliding time window to form the operating parameter observation tensor. The processing module (202) is used to fuse the operating parameter observation tensor and the boundary constraint class operating parameters of the generator to construct a fault observation tensor fused from multiple operating parameters. The processing module (202) is used to introduce a structural causal graph in the modeling layer, establish a causal path network based on causal reasoning and topological temporal 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 used to perform graph attention modeling on the evolution trend and instability boundary of each key parameter in the fault causal chain of different typical modes contained in the causal path network, and dynamically identify the path characteristics of the excitation system entering the fault critical state.

8. An electronic device, characterized in that, The device 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. The user interface (303) and the network interface (304) are both used to communicate with other devices. The communication bus (302) is used to realize the connection and communication between the components within the electronic device. The processor (301) is used to execute the instructions stored in the memory (305) so that the electronic device performs the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-6.

Images