Intelligent fault analysis method and system for extra-high voltage direct current transmission system based on domestic artificial intelligence large model framework, and storage medium

Abstract

The invention discloses an ultra-high voltage direct current transmission system fault intelligent analysis method and system based on a domestic artificial intelligence large model framework, and a storage medium, and belongs to the field of electric power system artificial intelligence fault diagnosis. According to the method, multi-source heterogeneous fault data are accessed and standardized and aligned, the processed data are input into a domestic artificial intelligence large model adapted by knowledge in the electric power field, and the model is utilized to perform multi-modal feature fusion and intelligent analysis, so that automatic identification and accurate positioning of fault types are realized; and an analysis report and a disposal strategy are automatically generated. The system comprises a data input module, an intelligent analysis module and a decision output module, and the intelligent analysis module performs feature extraction and recognition by adopting a CNN-Transform hybrid neural network and realizes fault positioning based on a knowledge graph and a graph neural network. According to the method, full-process automatic intelligent analysis from data access to decision output is realized, and the efficiency, the accuracy and the engineering landing capability of fault diagnosis are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault analysis of high voltage direct current transmission systems, and in particular to an intelligent fault analysis method, system and storage medium for ultra-high voltage direct current transmission systems based on a domestic artificial intelligence large model framework. Background Technology

[0002] Ultra-high voltage direct current (UHVDC) transmission systems play an increasingly important role in the construction of the global energy internet due to their advantages such as large transmission capacity, long distance, and low loss. However, their complex system structure and sophisticated operating logic result in extremely complex fault modes and response mechanisms, far exceeding those of traditional AC systems. Currently, fault analysis of UHVDC transmission systems mainly relies on manual methods.

[0003] Manual analysis typically follows a process from macroscopic phenomena to microscopic essence: Analysts first need to macroscopically examine the fault waveform data, and combine it with monitoring messages and protection action information to preliminarily determine the fault type (such as ground fault, commutation failure, valve short circuit, etc.) and time window; then, they meticulously check the phase, amplitude, waveform distortion, and load angle changes of electrical quantities to assess the correctness of the protection action; next, they precisely correlate the timing of the switch quantity changes with the electrical quantity changes to verify the action logic of the protection and control system; finally, they manually write a detailed report covering the fault type, location, cause, action evaluation, and handling recommendations.

[0004] Traditional methods suffer from inefficiency, heavy reliance on expert experience, high subjectivity and error-proneness, and significant limitations in data processing capabilities. Specifically, when faced with a large volume of waveform data generated by a single fault, manual point-by-point inspection and waveform analysis are time-consuming and cannot meet the timeliness requirements for rapid power grid recovery and decision-making. The quality of analysis is strongly correlated with the knowledge level and experience of the analysts; different experts may reach different conclusions on the same fault, leading to a lack of consistency and standardization. The analysis process is susceptible to subjective judgment and preconceived notions, and visual fatigue and decreased attention can easily occur amidst the deluge of data, resulting in misjudgments or omissions of key information. With the expansion of the power grid and the surge in monitoring points, the amount of fault data is growing exponentially, and manual analysis methods are no longer sufficient to effectively address the challenges of this data explosion.

[0005] Although there have been some attempts to introduce artificial intelligence technology into power system fault analysis, general models still have significant shortcomings in understanding power technical terms, identifying DC-specific fault waveform characteristics, integrating multi-source heterogeneous data, and meeting real-time requirements. A complete, efficient, interpretable, and applicable intelligent fault analysis solution for UHVDC transmission systems has not yet been formed. Summary of the Invention

[0006] To address the problems existing in the prior art, the main objective of this invention is to provide an intelligent fault analysis method, system, and storage medium for ultra-high voltage direct current (UHVDC) transmission systems based on a domestically developed large-scale artificial intelligence model framework. This enables a paradigm shift from being driven by human experience to being driven by data and knowledge collaboration, and constructs an intelligent analysis system that can automatically complete the entire process from raw data access to the generation of handling suggestions. At the same time, the general data analysis and artificial intelligence technologies are deeply adapted to the specific scenario of UHVDC transmission, which is highly specialized and has extremely stringent requirements for safety and real-time performance.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: S1. Acquire multi-source heterogeneous data of the UHVDC transmission system, including fault recording data, protection device event messages and synchronous phasor measurement data; perform standardization conversion and spatiotemporal alignment processing on the multi-source heterogeneous data to obtain preprocessed multimodal time series data. S2. Input the spatiotemporally aligned multimodal time-series data into a large-scale artificial intelligence model adapted with knowledge from the power sector, and perform multimodal feature extraction, including spatial features of transient signals, temporal dependency features, and topological correlation features, to generate a fused feature vector. S3. Input the fused feature vector into the fault diagnosis model and output the fault diagnosis result; the fault diagnosis result includes at least fault type determination, fault location and diagnosis confidence. S4. Based on the fault diagnosis results, generate a fault analysis report and suggestions for handling strategies.

[0008] Optionally, step S2 in the above method includes: The large-scale artificial intelligence model uses a hybrid neural network composed of convolutional neural network modules and Transformer modules; The preprocessed multimodal time series data is input into a pre-constructed convolutional neural network module, and the transient spatial features characterizing the local morphology and spatial correlation of the signal in the fault recording data are extracted through the convolution operation of the convolutional neural network module. The multimodal time series data is input into a pre-built Transformer module, and long-term time-series dependency features characterizing the evolution of electrical quantities before and after a fault are extracted through the self-attention mechanism of the Transformer module. The transient spatial features are fused with the long-term temporal dependency features to generate the fused feature vector.

[0009] Optionally, step S3 in the above method specifically includes: locating the fault location in the output fault diagnosis result, which is achieved through the following steps: Pre-construct a heterogeneous knowledge graph of equipment, protection devices, and electrical connections; The fusion feature vector generated in step S2 is mapped to the state information of nodes in the heterogeneous knowledge graph; By using graph neural networks to perform message propagation and node state updates in the heterogeneous knowledge graph, the propagation path of electrical disturbances can be inferred. By combining the action timing information contained in the protection device event messages in the multi-source heterogeneous data, the faulty component or the smallest fault section is determined on the propagation path, and the fault location is output.

[0010] Optionally, step S3 of the above method specifically includes: the adaptation method of the large-scale artificial intelligence model adapted with knowledge from the power field includes: Construct a structured prompt word template, which includes descriptive fields of typical fault phenomena of UHVDC systems, numerical fields of key equipment parameters, and logical fields of protection action logic; Obtain training sample pairs consisting of historical failure case data and their corresponding standard expert analysis reports; Using historical fault case data from the training sample pairs as model input, and using diagnostic text generated by the structured prompt word template that matches the standard expert analysis report as training target, supervised fine-tuning training is performed on the basic artificial intelligence large model.

[0011] Optionally, step S1 of the above method specifically includes: performing standardization transformation and spatiotemporal alignment processing on the multi-source heterogeneous data, which is achieved through the following steps: Extract the continuous time-series electrical quantity curves of the fault recording data from the multi-source heterogeneous data, as well as the discrete action event sequences in the event messages of the protection device; The optimal time warping path between the continuous time sequence curve of the electrical quantity and the discrete action event sequence is calculated using a dynamic time warping algorithm. Based on the optimal time curvature path, the continuous time sequence curve of the electrical quantity or the discrete action event sequence is calibrated on the time axis to align their timestamps.

[0012] Optionally, the above method further includes a closed-loop iteration step: The method also includes a continuous iterative optimization step for the model: Implement an active learning strategy to automatically trigger manual annotation for fault cases where the confidence level of the AI ​​large model prediction is lower than a set threshold, and add the newly annotated data to the training set for incremental learning. Validated and effective handling cases are transformed into structured rules and dynamically updated into the knowledge graph used for strategy generation.

[0013] Optionally, the above method further includes a real-time inference optimization step: The knowledge of the large artificial intelligence model is transferred to a lightweight reasoning model using knowledge distillation technology. The lightweight inference model is optimized by operator fusion and precision quantization using the TensorRT framework to improve diagnostic response speed.

[0014] In a second aspect, the present invention provides an intelligent fault analysis system for ultra-high voltage direct current transmission systems, comprising: The data input module is used to acquire multi-source heterogeneous data from the UHVDC transmission system, including fault recording data, protection device event messages, and synchronous phasor measurement data; and to perform standardization conversion and spatiotemporal alignment processing on the multi-source heterogeneous data to obtain preprocessed multimodal time series data. An intelligent analysis module, connected to the data input module, is used to receive the multimodal time-series data. The intelligent analysis module includes a large-scale artificial intelligence model adapted with knowledge of the power industry, used to perform multimodal feature extraction on the multimodal time-series data to generate a fused feature vector, and includes a fault diagnosis model, used to output fault diagnosis results based on the fused feature vector. The fault diagnosis results include at least fault type determination, fault location, and diagnostic confidence. The decision output module, connected to the intelligent analysis module, is used to generate a fault analysis report and handling strategy recommendations based on the fault diagnosis results.

[0015] Optionally, the intelligent analysis module in the above system includes a fault type identification unit and a fault location unit; The input end of the fault type identification unit is connected to the data input module to receive the preprocessed multimodal time series data. It is equipped with a convolutional neural network, a Transformer module and a random forest classifier. The unit extracts spatial features from the multimodal time series data through the convolutional neural network, extracts the temporal dependencies through the Transformer module, and uses the random forest classifier to generate fault type determination results and their confidence levels. The input end of the fault location unit is connected to the fault type identification unit and the data input module, and is used to receive the fault type determination result and the multimodal time series data. The unit stores a heterogeneous knowledge graph of power grid equipment, protection devices and topological connection relationships, and based on the knowledge graph, uses graph neural network to infer the fault propagation path, and combines the protection action timing in the multimodal time series data to output the location result of the faulty equipment or faulty section.

[0016] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In this invention, multi-source heterogeneous data from the UHVDC transmission system are standardized and spatiotemporally aligned to obtain preprocessed multimodal time-series data, replacing the manual identification, conversion, and alignment steps and providing a unified input for subsequent analysis. Subsequently, based on the multimodal time-series data, a large-scale artificial intelligence model adapted to power sector knowledge extracts multimodal features and generates fused feature vectors. Then, a fault diagnosis model outputs fault diagnosis results including fault type determination, fault location, and diagnostic confidence, and automatically generates a fault analysis report and disposal strategy recommendations. This process achieves full automation from data preprocessing and intelligent analysis to report generation, significantly shortening the response time for fault diagnosis and disposal, and improving the consistency of results.

[0018] (2) The artificial intelligence model adapted to power industry knowledge adopted in this invention internalizes professional knowledge such as equipment parameters, fault logic, and protection principles of UHVDC systems into the model through targeted training. This model can perform feature extraction on the multimodal time-series data, comprehensively consider the transient characteristics of electrical quantities, time-series dependencies, and topological correlation features, and generate a fused feature vector, thereby completing the accurate identification and location of typical faults, including commutation failure and DC line grounding. This realizes a mode leap from rule-based matching to multi-dimensional deep feature understanding, improving the analysis accuracy and generalization ability in professional scenarios.

[0019] (3) This invention constructs a complete technical system from data preprocessing and intelligent analysis to decision output. Specifically, this system is manifested in: data acquisition and standardization alignment through the data input module; feature fusion and diagnosis through the adapted large model and fault diagnosis model in the intelligent analysis module; and the generation of structured reports and strategies through the decision output module. Each module collaborates through standardized interfaces, forming a systematic solution that can be integrated into existing power information platforms. This solution systematically solves the problems of low efficiency, strong reliance on experience, and insufficient processing capacity in traditional manual analysis in a standardized, automated, and intelligent manner, providing reliable technical support for the safe and stable operation of ultra-high voltage direct current transmission systems.

[0020] The invention will now be further described with reference to the accompanying drawings. Attached Figure Description

[0021] Figure 1This is a flowchart illustrating an intelligent fault analysis method for an ultra-high voltage direct current transmission system according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the process for extracting fault recording features in a method according to an embodiment of the present invention.

[0022] Figure 3 This is a functional block diagram of an intelligent fault analysis system for an ultra-high voltage direct current transmission system according to an embodiment of the present invention. Detailed Implementation

[0023] To better illustrate the objectives, technical solutions, and advantages of the present invention, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0024] like Figure 1 The diagram illustrates an intelligent fault analysis method for ultra-high voltage direct current transmission systems based on a domestically developed large-scale artificial intelligence model framework, according to an embodiment of the present invention. This method can be implemented through the following specific implementation steps, which completely construct a closed-loop intelligent analysis process from data access to report generation. The implementation of this method mainly includes the following steps: Step S1: Standardization and Spatiotemporal Alignment of Multi-Source Heterogeneous Data. Multi-source heterogeneous data from the UHVDC transmission system is acquired, including fault recording data, protection device event messages, and synchronous phasor measurement data. The data undergoes standardization format conversion, noise filtering, and dimensional normalization. Specifically, continuous time-series curves of electrical quantities are extracted from fault recordings, and discrete action sequences from protection event messages. A dynamic time warping algorithm is used to calculate the optimal time curvature path between the two to compensate for sampling delays, and time axis calibration is performed accordingly. This technique aims to eliminate analysis errors caused by multi-source asynchrony and output spatiotemporally strictly aligned, uniformly formatted multimodal time-series data, laying a high-quality data foundation for subsequent analysis.

[0025] Specifically, in step S1, a dedicated data input layer is used to achieve unified collection and management of information from underlying devices. This layer is compatible with heterogeneous communication protocols of various manufacturers' devices, and aggregates multi-source signals such as raw waveform recordings, protection device event messages, and PMU synchronization phasor measurement data. A unified data cleaning framework is used to parse and convert the raw binary or specific format waveform recordings, and perform noise filtering and dimension normalization operations. Based on synchronization phasor measurement standards such as IEEE C37.118, precise time-stamping of PMU data is performed, thereby constructing a standardized data baseline with spatiotemporal alignment. A streaming pipeline based on the Flinks (a stream processing framework) real-time computing engine and Kafka (a distributed message queue) is adopted. The implementation uses Flink for millisecond-level window calculations and complex event processing, while Kafka is responsible for high-throughput data buffering and load balancing.

[0026] Step S2 involves deep extraction and fusion of multimodal features based on a domain-adapted large-scale model. The multimodal time-series data obtained in Step S1 is input into an AI large-scale model adapted to power domain knowledge. The core of this model is a hybrid neural network architecture consisting of a Convolutional Neural Network (CNN) module and a Transformer module (a neural network model based on a self-attention mechanism). Its technical role is to collaboratively capture the local spatial patterns and global temporal dependencies of signals. The CNN module performs convolution operations on the recorded waveform data to extract spatial features of transient signals such as voltage drops and current surges; the Transformer module analyzes the evolution of electrical quantities before and after a fault through a self-attention mechanism, extracting long-term temporal dependency features. Furthermore, the large-scale model undergoes domain fine-tuning using pre-constructed structured cue word templates (containing typical fault description fields such as commutation failure and DC line grounding, as well as equipment parameters) to improve its understanding of specialized scenarios. Finally, the extracted spatial features and temporal dependency features are fused to generate a high-dimensional fused feature vector that comprehensively represents the fault state.

[0027] Specifically, in step S2, an analysis and decision layer is employed to perform intelligent diagnosis and decision generation. This analysis and decision layer uses a hierarchical reasoning framework to decompose the fault analysis task into sub-problems such as type identification, precise location, and response generation. In the hybrid neural network architecture, convolutional neural network modules are used to extract local spatial features from voltage and current transient waveforms, while the self-attention mechanism of the Transformer model is used to capture long-term dependencies across time. The importance weights of the input features are quantified using the SHAP (SHapley Additive exPlanations, a feature importance interpretation method) interpretability tool.

[0028] Step S3 integrates knowledge graphs and graph neural networks for intelligent diagnosis and precise location. The fused feature vector generated in step S2 is input into the fault diagnosis model, and the output includes fault type determination (e.g., identifying commutation failure, DC line grounding, valve short circuit, etc.) and diagnostic confidence. Simultaneously, to achieve precise location, this method pre-constructs a heterogeneous knowledge graph of equipment, protection devices, and electrical connections. The fused feature vector is mapped to the state of graph nodes, and graph neural networks are used for message propagation and node state updates to simulate and infer the propagation path of fault electrical disturbances in the power grid topology. Its technical advantage lies in interpreting isolated signal features within a system context. Finally, by combining the specific action timing information in the protection device event messages, the faulty component or the smallest fault segment is located on the inferred propagation path, completing the fault location.

[0029] Specifically, in step S3, the analysis and decision-making layer constructs a topological correlation map between equipment and protection for precise fault location, and integrates the temporal logic of protection actions with the abrupt change patterns of electrical quantities (such as amplitude, phase, and harmonics) for analysis. The hybrid reasoning mechanism employed is as follows: for near-field faults, rapid area locking is primarily based on the correlation of protection action signals; for complex faults with significant electrical distances, multi-round iterative calculation algorithms are initiated to gradually converge and narrow down the possible range of the fault point. Regarding the generation of handling strategies, based on reinforcement learning technology, the fault recovery process is modeled as a Markov decision process. By defining objective functions such as equipment operation costs and system safety risks, and using algorithms such as Monte Carlo tree search to explore the strategy space, a step-by-step optimal isolation and recovery operation sequence that satisfies N-1 safety constraints is automatically generated.

[0030] Specifically, this analytical decision-making layer deeply integrates knowledge-driven and data-driven methodologies. In achieving intelligent fault type determination, its technical approach first involves multimodal feature extraction, specifically employing a joint time-frequency domain analysis method. Wavelet transforms or short-time Fourier transforms are performed on the fault waveform signals to extract features such as amplitude, phase, and harmonic content at different time scales. Simultaneously, the SOE (Sequence of Events) information of the protection device is analyzed to obtain its specific timing and logical state during startup and operation. Finally, continuous waveform features and discrete event features are fused to construct a unified multi-dimensional fault feature matrix. In the model decision-making stage, an integrated architecture is deployed for collaborative decision-making. This architecture consists of a CNN-Transformer hybrid network and a random forest classifier: the CNN is responsible for capturing local spatial patterns and high-frequency transient features from the waveform data; the Transformer model encodes the entire fault time series to capture long-range temporal dependencies; and the random forest classifier works in parallel, providing classification predictions and ranking the importance of each input feature. Finally, a weighted voting mechanism is used to synthesize the outputs of all models to generate a preliminary diagnostic result. To enhance reliability, the module introduces an uncertainty quantification and verification mechanism. The entropy of the probability distribution output by each base classifier is calculated to measure the uncertainty of its predictions. This is combined with the prior distribution obtained from historical fault statistics to construct an evidence confidence matrix, resulting in a comprehensive confidence score. When this score falls below a preset threshold (e.g., 85%), the system automatically triggers a rule-based expert system for backtracking verification, such as verifying whether the timing logic of protection actions conforms to established procedures.

[0031] In terms of fault location and handling suggestion generation, the technical framework of this module is based on a collaborative working mechanism of knowledge graph reasoning and rule engine. First, causal reasoning using the knowledge graph is performed. A heterogeneous knowledge graph covering "equipment-protection-topology" is constructed. Nodes in the graph represent physical equipment and protection devices such as circuit breakers, transformers, and lines. Node attributes embed information such as nameplate parameters and protection settings, while edges represent electrical connections and protection affiliations. Graph neural network technology is used to perform reasoning on this graph, simulating the propagation path of fault current and combining it with the timing constraints of actual received protection actions to effectively narrow down the range of candidate faulty equipment. Second, a multi-objective optimization decision model is established. This model takes the shortest system recovery time and the fewest on-site operation steps as core optimization objectives, while using the current health status of the equipment and the priority of the load as hard constraints. Through evolutionary algorithms such as genetic algorithms, a set of Pareto optimal (i.e., the non-dominated solution set in multi-objective optimization) candidate handling schemes is searched and generated in the vast solution space. Finally, the rule engine performs dynamic adaptation and final decision-making. The system loads a structured rule base consisting of operating procedures and anti-accident measures, and uses the efficient RETE algorithm (an efficient pattern matching algorithm for production systems) to perform pattern matching and select the optimal handling strategy from the candidate solutions.

[0032] Step S4: Automatic generation of structured report and handling strategy. Based on the complete fault diagnosis results obtained in Step S3, a structured fault analysis report and handling strategy recommendations are automatically generated. The report content covers the fault time, type, location, analysis conclusions, and confidence level assessment, and supports exporting to standardized formats such as XML and PDF. The generation of handling strategy recommendations is based on the diagnostic results and built-in safety procedures, and can call upon a validated handling case library (the case library is dynamically updated by converting high-scoring handling cases into structured rules) for optimization and recommendation. Its technical role is to provide decision support that can directly guide on-site operations.

[0033] Specifically, in step S4, by configuring the output and service layer, the structured conclusions output by the analysis and decision-making layer are transformed into standardized control commands or alarm information that the scheduling system can understand and execute. These commands are then pushed through standardized interfaces such as RESTful APIs (Representational State Transfer Application Programming Interfaces) or message middleware. Simultaneously, this layer provides a visual human-machine collaboration interface. The system can automatically generate compliant accident analysis reports, organized in XML format, fully including key fields such as fault timescale, type determination, affected equipment scope, recommended handling steps, and reasoning basis. These reports can be exported as PDF or Word documents as needed. The human-machine collaboration interface design provides functions such as full fault waveform playback, simulation and comparison of different handling strategies, and optimization of key model parameters.

[0034] Step S5: Iterative optimization of the model based on active learning and feedback. Establish a continuous learning mechanism for the model. Implement an active learning strategy: when the diagnostic confidence level falls below a set threshold, automatically trigger a manual review process and add newly labeled high-value samples to the training set. Use incremental learning algorithms combined with knowledge distillation techniques to update the parameters of the fault diagnosis model, absorbing new knowledge while retaining historical experience. This technique enables the model to adapt to changes in power grid operation and new types of faults. Simultaneously, the actually implemented and verified effective handling strategies are parsed into condition-action rules using natural language processing technology and dynamically updated to the knowledge graph and rule base used for strategy generation, forming a reinforced closed loop of data, model, and knowledge.

[0035] Step S5 describes the sample closure and model iteration mechanism for achieving continuous model optimization. Regarding data sample acquisition and intelligent calibration, this embodiment designs a dual-channel data acquisition framework combined with an active learning strategy. A multimodal data acquisition matrix covering normal operating conditions, typical faults, and various edge scenarios is constructed. The system captures real-time waveform files from the field, timing records of protection device actions, and PMU synchronization phasor data; simultaneously, a fault injection simulation system is integrated to generate synthetic data covering various fault modes and severity levels. An intelligent calibration strategy is adopted. First, a semi-supervised clustering algorithm (such as DBSCAN, a density-based spatial clustering algorithm applying noise) is used to initially group a large number of unlabeled samples. Then, high-quality pseudo-labels are generated for each cluster using domain expert knowledge. Furthermore, contrastive learning techniques are used to narrow the distance between high-confidence samples in the feature space. To efficiently manage these samples, a dynamic sample library is constructed. Real-time data entry and efficient retrieval of samples are achieved using a Redis (a high-performance key-value pair in-memory database) in-memory database, and composite indexes are established according to multiple dimensions such as fault type, equipment type, and timestamp.

[0036] Regarding model performance monitoring and online iteration strategies, this embodiment deploys a multi-dimensional, real-time monitoring indicator system and establishes an automated model update process. The real-time performance monitoring component constructs a monitoring system covering core indicators such as diagnostic accuracy, recall, inference latency, and computational resource utilization. Prometheus (an open-source system monitoring and alerting toolkit) is used for real-time indicator collection and aggregation, and Grafana (an open-source data visualization and monitoring platform) is used to achieve minute-level data visualization and anomaly threshold alerts. Once the monitoring system detects that a key performance indicator (such as the recall rate of a certain type of fault) is consistently below a preset threshold, an online fine-tuning mechanism is automatically triggered. This mechanism employs an incremental learning framework, using only newly added fault samples to update the model's local parameters, and constrains the new model through knowledge distillation technology, ensuring it retains important knowledge learned from historical data. The entire iteration process follows a strict version management strategy. Model version control is established based on a Git system (a distributed version control system), meticulously recording the hyperparameters used in each iteration, training dataset snapshots, and performance evaluation results.

[0037] In the feedback and knowledge base update stages of handling effectiveness, this embodiment constructs a closed-loop feedback system. First, the handling effectiveness is quantitatively evaluated. A reward function based on reinforcement learning is designed to comprehensively score automatically generated or manually executed handling plans from multiple dimensions, including total fault recovery time, number of execution steps, and system volatility. The scoring results are then structured and written into the event log of that fault. Subsequently, the dynamic update process of the knowledge base is initiated. Handling cases evaluated as efficient are transformed into structured rules in the form of "IF (condition), THEN (action)" using natural language processing technology, and these rules are injected as new knowledge into the system's knowledge graph and rule engine. For example, the BERT model (Bidirectional Encoder Representations from Transformers) is used for key information extraction and relationship recognition. To maintain the consistency and reliability of the knowledge base, a causal reasoning enhancement mechanism is also introduced. The causal chain of fault propagation is explicitly embedded in the knowledge graph, and statistical methods such as Granger causality test (a statistical method for analyzing causal relationships between time series variables) are used periodically to verify the significance of newly discovered associations in the data.

[0038] Step S6: Lightweight Real-Time Inference Deployment Based on Knowledge Distillation and Hardware Acceleration. To meet the real-time requirements of engineering applications, the model is lightweighted and accelerated. Knowledge distillation technology is used, employing the domain-adapted large model from Step S2 as the teacher model to train a simplified student model with significantly reduced parameters, which serves as the final deployment model. Subsequently, the TensorRT inference optimization framework is used to perform operator fusion, inter-layer fusion, and INT8 quantization optimizations on this student model. This significantly reduces computational complexity and resource consumption. The optimized lightweight model is deployed on edge computing devices or servers, enabling fault diagnosis results to be generated and responded to within seconds or milliseconds.

[0039] This embodiment designs and executes system experiments and results analysis to comprehensively verify the proposed intelligent fault analysis method for UHVDC transmission systems.

[0040] Firstly, the focus was on building a simulation testing environment and standardized evaluation benchmarks that closely resemble actual engineering conditions. A high-performance experimental platform integrating hardware and software was constructed. At the hardware level, a heterogeneous computing cluster including multi-core CPUs and high-performance GPUs was deployed. In terms of software environment configuration, the mainstream deep learning frameworks TensorFlow and PyTorch were integrated for algorithm development and model training; simultaneously, professional power system simulation software such as OpenDSS (Open Distribution System Simulator) was selected to construct a detailed electromagnetic transient model of UHVDC transmission. To reproduce the data flow processing in the actual system, a Kafka distributed message queue was configured as a bus for data access and distribution, and a Redis in-memory database was used to achieve real-time data caching and high-speed access. Based on this platform, simulation datasets covering various typical and edge fault scenarios were generated. By systematically adjusting key parameters such as fault point resistance (e.g., from metallic short circuit to high-resistance grounding), fault triggering location (e.g., rectifier side, inverter side, DC line midpoint, etc.), and fault duration, a large number of diverse fault samples were generated. In addition, noise that matches the characteristics of actual field measurements was actively injected into the simulation data.

[0041] This invention designs a multi-dimensional quantitative evaluation index system. This system comprehensively considers four core dimensions: fault identification accuracy, fault location accuracy, effectiveness of handling suggestions, and real-time performance of model inference. For fault identification accuracy evaluation, a quantitative evaluation system centered on classification tasks is constructed. Specifically, classic indicators such as accuracy, recall, and F1-score are used to comprehensively evaluate the model's ability to determine various faults (such as commutation failure and DC line grounding). Furthermore, a confusion matrix is ​​introduced to visualize the model's prediction results. The stability of the model's classification performance under different decision thresholds is evaluated by combining the receiver operating characteristic curve (ROC curve) and its area under the curve (AUC value). In the dimensions of fault location and handling suggestion evaluation, for location accuracy, a dynamic time warping algorithm is used to calculate the similarity score between the actual fault waveform characteristics and the standard fault template waveform. For handling suggestions, a scoring system is constructed to determine their reasonableness from multiple dimensions, including whether the operation steps comply with safety procedures, the total system recovery time, and the effect of power grid disturbance suppression after implementation. The timeliness of model inference is assessed by actually measuring the total delay from data input to outputting a complete analysis report.

[0042] To verify the advancement and reliability of the proposed method, this invention designed systematic comparative experiments and extreme scenario tests. In the comparative experiments, the intelligent analysis model proposed by this method was compared horizontally with fixed-rule-based expert systems and traditional shallow machine learning models (such as support vector machines and decision trees) on the same test set. Key indicators were compared, such as the overall misclassification rate of fault types and the average location error distance. Ablation experiments were used to analyze the contribution of each innovative component in the model (such as multimodal fusion and knowledge graph) to the final performance. For robustness testing, extreme test scenarios were constructed, including environments simulating high-intensity measurement noise and communication interference, coupled scenarios where multiple faults occur successively or simultaneously, and situations where the power grid operating topology changes.

[0043] Finally, a specific analysis was conducted on the interpretability of the intelligent model and the overall reliability of the system. Regarding interpretability, visualization tools such as SHAP value calculation and the LIME local interpretation model were used to demonstrate the ranking of the contributions of different electrical characteristics (such as DC voltage drop and harmonic distortion rate at a given moment) to the specific fault determination results. In terms of reliability quantification, uncertainty estimation techniques such as Monte Carlo Dropout were used to calculate the confidence level or variance of each prediction by the model, generating confidence intervals for the prediction results. Furthermore, a post-evaluation closed loop for the handling effect was established, correlating the simulated handling plan with the subsequent simulated recovery trajectory of the system, and quantifying the recovery effect through a predefined reward function.

[0044] It is worth noting that, on the one hand, this embodiment, through automated parsing, decoding, and DTW alignment processes, normalizes multi-source data into a standard data matrix with strict temporal synchronization and uniform format. Its technical effect is to eliminate heterogeneity and temporal ambiguity at the data level, providing a precise and reliable source of unique facts for all subsequent analysis steps. This is an irreplaceable foundation for any advanced automated analysis, transforming messy, asynchronous, and manually interpretable raw data into clean, aligned, and directly computable structured information.

[0045] It is worth noting that, on the other hand, this embodiment transforms expert experience (fault characteristics, protection logic) into parameterized knowledge that the model can understand and utilize through prompt word engineering and comparative learning, thus transforming the general large model into a domain expert model. A CNN-Transformer hybrid model extracts spatiotemporal features in parallel, and the knowledge graph and GNN are topologically linked in parallel inference. This achieves simultaneous and comprehensive consideration of high-dimensional and complex fault characteristics, with computational depth and breadth far exceeding that of manual sequential thinking. The processing results not only output diagnostic conclusions but also include SHAP interpretability analysis and uncertainty quantification (confidence intervals). This is equivalent to providing not only the answer but also the solution steps and the degree of certainty; the amount of information, objectivity, and standardization of its output are systematically provided by manual analysis. This achieves a shift from sequential logical reasoning relying on expert brains to deep computation and reasoning based on domain knowledge and parallel large models.

[0046] It is worth noting that, furthermore, this embodiment of the system automatically matches the knowledge base based on the structured diagnostic results to generate a standardized report containing key fields (time, type, location, and recommendations). Its technical effect is to standardize and make the value delivery of analytical conclusions instantaneous, directly connecting to subsequent scheduling or management systems, achieving a seamless connection from analysis to execution.

[0047] In summary, this embodiment actively learns to filter difficult samples, updates the model through incremental learning, and updates the knowledge base through rule transformation. This allows the system's input-output mapping relationship to be continuously optimized with the input of new data. The technical effect is the construction of an adaptive system whose performance improves over time and with data accumulation, solving the fundamental problems of traditional methods being rigid and unable to adapt to new scenarios.

[0048] like Figure 3 The diagram illustrates an intelligent fault analysis system for ultra-high voltage direct current (UHVDC) transmission systems based on a domestically developed large-scale artificial intelligence model framework, according to an embodiment of the present invention. Its key feature is the adoption of a layered, decoupled agent architecture design, establishing a complete link from data access and analysis / decision-making to service output. Each module collaborates through standardized interfaces to meet the system's requirements for real-time performance, reliability, and scalability. Specifically, the system includes a data input module, a decision output module, and a data input module.

[0049] The data input module, serving as the system's perception layer, is used to acquire multi-source heterogeneous data in real time from various monitoring devices in the UHVDC transmission system. This multi-source heterogeneous data includes fault recording data, protection device event messages, and synchronous phasor measurement data. The module has a built-in data parsing, cleaning, and alignment engine that performs standardized format conversion, noise filtering, dimensional normalization, and high-precision spatiotemporal alignment processing on the multi-source heterogeneous data, outputting unified, clean, and time-consistent multimodal time-series data, laying a high-quality data foundation for upper-level intelligent analysis.

[0050] The data input module, serving as the core decision-making layer of the system, is connected to the data input module and is used to receive the multimodal time-series data. This module is based on a domestically developed large-scale artificial intelligence model that has been deeply adapted and fine-tuned using knowledge from the power industry, and further specifically includes: The fault type identification unit has its input end connected to the data input module to receive the multimodal time series data. The unit internally includes a convolutional neural network module, a Transformer module, and a random forest classifier. The convolutional neural network module extracts transient spatial features from the multimodal time series data, the Transformer module extracts long-term temporal dependencies, and the random forest classifier generates fault type determination results and their confidence levels. The fault location unit has its input end connected to the fault type identification unit and the data input module, and is used to receive the fault type determination result and the multimodal time series data. The unit stores a heterogeneous knowledge graph of power grid equipment, protection devices and topological connection relationships, and based on the knowledge graph, uses a graph neural network to infer the fault propagation path, and combines the protection action timing in the multimodal time series data to output the location result of the faulty equipment or faulty section.

[0051] This module, through the collaboration of the aforementioned units, enables accurate identification and location of typical DC faults, including commutation failure and DC line grounding.

[0052] The decision output module, serving as the system's execution and interaction layer, is connected to the intelligent analysis module. Based on the fault diagnosis results output by the intelligent analysis module and using a built-in rule base and report templates, this module automatically generates a standardized fault analysis report and handling strategy recommendations that include fault analysis, cause inference, and operation steps. The report and strategies are then pushed to maintenance personnel or the scheduling system via a visual interface or standardized data interface to support rapid closed-loop fault handling.

[0053] Through the close collaboration of the three modules mentioned above, the system constructs a complete automated analysis closed loop from data perception and intelligent diagnosis to decision output, realizing rapid, accurate, and intelligent analysis of faults in the UHVDC transmission system.

[0054] Implementation Example For ease of understanding, the following is a specific implementation example based on this embodiment.

[0055] I. Overall Architecture and Data Infrastructure Construction As the system's perception front end, the data input layer is responsible for the unified access and real-time preprocessing of multi-source heterogeneous data. Its core functions include: parsing and standardizing the accessed fault recording files, protection device event messages, and synchronous phasor measurement (PMU) data; using the Dynamic Time Warping (DTW) algorithm and PMU time-stamping synchronization technology to solve the timing asynchrony problem of multi-source data and achieve precise alignment between electrical quantities and event information; and achieving millisecond-level data throughput and rapid extraction of key fault segments based on a stream processing pipeline built on Flink and Kafka, providing high-quality, time-consistent structured data streams for upper-layer analysis.

[0056] 1. Data Input Layer: Unified Access and Standardized Preprocessing This layer is responsible for collecting raw data from various devices and performing standardized preprocessing to solve the problem of multi-source heterogeneity and provide high-quality input for upper-layer analysis.

[0057] Multi-source data access: Supports access to fault recording files, protection device event messages and PMU synchronous phasor data, covering all monitoring signals of key equipment.

[0058] Data standardization processing: The core task is to parse the configuration file (.config) and the original data file (.dat) from the fault waveform data compressed package to complete the standardization transformation and time-series alignment of multi-source heterogeneous data. Specifically, this includes: Parsing and Conversion: A dedicated parsing toolchain was developed to extract metadata such as sampling rate and channel mapping from the .config file, decode the .dat file and reconstruct it into a time-series data matrix that conforms to the IEEE standard, and perform feature formatting processing such as normalization and downsampling on the original sampled values.

[0059] Timing Alignment: Establish a spatiotemporal alignment mechanism between waveform recording data and protection information. Employ the Dynamic Time Warping (DTW) algorithm or an event-driven alignment method combined with PMU absolute timescales to resolve timing asynchrony issues caused by sampling rate differences and transmission delays. This precisely aligns electrical quantity timing data with protection action event sequences on the time axis, forming a unified analysis unit.

[0060] Data smoothing and augmentation: To address the high noise and small sample size characteristics of DC fault data, a filtering algorithm based on fluctuation characteristics (such as improved wavelet threshold denoising) is designed to preserve key transient features. At the same time, data augmentation strategies such as SMOTE oversampling and GAN generative adversarial networks are combined to optimize the small sample distribution and improve the model's generalization ability.

[0061] Real-time streaming processing: Flink and Kafka are used to build a distributed data pipeline to achieve millisecond-level data throughput and dynamic load balancing, and to support the rapid extraction and uploading of critical data segments when a failure occurs.

[0062] 2. Analytical Decision-Making Layer: Integrating Reasoning and Intelligent Diagnosis refer to Figure 2 This system automates the entire fault diagnosis process based on a hierarchical reasoning architecture. Its core is fault waveform feature extraction based on wave analysis and multimodal models, combined with knowledge-driven reasoning. The system focuses on parsing the config configuration file and dat time-series data file from the fault waveform data compressed package, completing the standardization transformation and time-series alignment of multi-source heterogeneous data. Noise interference is eliminated through wave characteristic analysis (such as wavelet denoising and sliding window smoothing), and small sample distribution is optimized using data augmentation strategies (SMOTE / GAN). The preprocessed time-series data serves as the basic input for multimodal feature extraction, while establishing an association mapping with standard fault templates, providing high-quality data support for subsequent similarity calculations and fault diagnosis. Multimodal feature fusion and extraction: This is a crucial step in intelligent analysis, aiming to construct highly discriminative fault features from preprocessed data. Specifically, it defines the entire lifecycle management of fault waveform data from raw acquisition to model input, encompassing two core aspects: data acquisition specification design and multi-source heterogeneous data integration, providing a structured data foundation for subsequent analysis. A standardized access process for fault waveform files is established, clarifying the transmission protocols, file encapsulation formats, and interface specifications from the device end (e.g., DC protection devices) to the cloud platform, ensuring seamless data exchange between devices from multiple vendors. A spatiotemporal alignment mechanism between waveform data and protection information is established, using methods such as timestamp synchronization and device ID mapping to fuse dispersed protection action signals (e.g., circuit breaker status, protection alarms) with electrical quantity time-series data into a unified analysis unit.

[0063] Feature extraction based on wave analysis: A joint time-frequency domain analysis method is used to deeply explore the dynamic characteristics of fault signals. The heterogeneity of fault waveform files is addressed by parsing the configuration file (.config) and the original data file (.dat) within the compressed package to complete the standardized reconstruction and feature engineering preprocessing of the time-series data.

[0064] First, a dedicated parsing toolchain is developed to support the extraction of metadata such as sampling rate and channel mapping from .config files, as well as the decoding and floating-point restoration of .dat binary files, generating a time-series data matrix that conforms to IEEE standards.

[0065] Time-frequency analysis techniques employ wavelet transform for multi-resolution decomposition to locate transient disturbance frequency bands; or utilize Hilbert-Huang transform (HHT) to extract instantaneous frequency and amplitude characteristics of the signal through empirical mode decomposition (EMD), effectively characterizing the oscillation modes and energy evolution laws of the fault transient process. Specifically, to address the heterogeneity issue of fault recording files, the standardized reconstruction and feature engineering preprocessing of time-series data are completed by parsing the configuration file (.config) and original data file (.dat) within the compressed package.

[0066] Temporal depth modeling: Combining LSTM networks to capture long-term temporal trends, and utilizing the self-attention mechanism of the Transformer model to calculate global temporal dependencies, thereby extracting joint time-frequency domain depth features of fault signals.

[0067] Cross-modal feature fusion mechanism: solves the problem of fusing multi-source data such as waveform recording data and protection information.

[0068] Attention-weighted fusion: Utilizes a multi-head attention mechanism to dynamically allocate the weights of electrical quantity characteristics and protection event characteristics, thereby enhancing information during critical periods.

[0069] Graph network spatiotemporal correlation modeling: Construct a topology graph of power grid equipment, and explicitly inject spatial correlation information between equipment through graph convolutional networks (GCN) and graph attention mechanism (GAT) to model the spatiotemporal dependencies of fault propagation.

[0070] Intelligent fault type determination: Construct a multimodal feature fusion network for accurate classification.

[0071] Model collaborative decision-making: Deploying an integrated architecture of CNN-Transformer hybrid networks and classifiers such as random forests. CNN is responsible for extracting spatial features of transient signals, Transformer captures temporal dependencies, and finally generates diagnostic results through weighted voting.

[0072] Interpretability and Uncertainty Quantification: SHAP values ​​are used to calculate the importance weight of each feature to the judgment result, providing interpretability analysis. Simultaneously, methods such as Monte Carlo Dropout (MCDropout) are used to estimate the uncertainty of predictions made by large AI models, generating confidence intervals to provide a basis for manual review.

[0073] Accurate fault location and handling generation: Knowledge Graph Causal Reasoning: Construct a heterogeneous knowledge graph of equipment, protection, and topology, embedding attributes such as equipment parameters and protection settings, and using graph neural networks (GNN) to reason about fault propagation paths, combined with protection action timing constraints to narrow down the fault area.

[0074] Optimize decision generation: Model Markov decision process (MDP) based on reinforcement learning, or establish an optimized artificial intelligence model with the goal of minimizing recovery time and number of operations, and generate the optimal disposal sequence that meets the N-1 safety constraint through Monte Carlo tree search, genetic algorithm, etc.

[0075] 3. Output and Service Layer: Result Delivery and Human-Machine Collaboration Transform the analysis results into actionable decisions and services.

[0076] Standardized report generation: Automatically generates standardized incident reports in XML format, including key fields such as fault time, type, location, scope of impact, and handling recommendations, and supports exporting to PDF / Word and other formats.

[0077] Human-machine collaboration interface: Provides a visual interactive interface, supports functions such as fault process playback, comparison of different handling strategies, and model parameter optimization, and allows operation and maintenance personnel to modify system suggested solutions and redistribute them.

[0078] II. Model Iteration and Continuous Learning Establish a data-driven continuous learning mechanism to ensure the adaptive evolution of the system's intelligence.

[0079] Closed-loop management of samples: Intelligent labeling and collection: Semi-supervised clustering and contrastive learning strategies are used to perform intelligent initial labeling on unlabeled samples; through active learning strategies, cases with low confidence of large artificial intelligence models are automatically selected to trigger manual labeling and generate high-value samples.

[0080] Dynamic sample library: A dynamic sample library is built based on Redis and other technologies, supporting multi-dimensional indexing and real-time data import.

[0081] Online model iteration: Performance monitoring: Build a real-time monitoring system covering multiple dimensions such as accuracy, recall, and inference latency (such as Prometheus and Grafana).

[0082] Incremental learning and fine-tuning: An incremental learning framework is adopted to locally update the large artificial intelligence model only with new data, and knowledge distillation is combined to retain historical knowledge and prevent catastrophic forgetting.

[0083] The knowledge base is updated dynamically: Validated and effective handling cases are transformed into structured rules using natural language processing (NLP) technology and dynamically injected into the knowledge graph and rule engine.

[0084] The significance of new associations in knowledge graphs is verified by methods such as Granger causality test, thereby continuously improving the reliability of knowledge.

[0085] III. Technology Verification and Evaluation The effectiveness and engineering applicability of the scheme are demonstrated by establishing a comprehensive experimental and evaluation system.

[0086] Experimental environment construction: Deploy a heterogeneous computing cluster of CPU and GPU, integrate deep learning frameworks such as TensorFlow / PyTorch and power system simulation tools (such as OpenDSS), configure stream processing middleware, and build a simulation dataset covering typical faults and noise interference.

[0087] Multi-level performance evaluation: Fault identification and evaluation: The model classification performance is comprehensively quantified using metrics such as accuracy, recall, F1-score, confusion matrix, and ROC curve. At the same time, the confusion matrix is ​​introduced to analyze the discriminative power of easily confused fault modes, thus comprehensively evaluating the model classification performance.

[0088] To assess fault identification accuracy, the proportion of correctly classified fault samples is calculated by comparing the model's predictions with the true labels. The formula for calculating fault identification accuracy is: ; In this system, TP (True Positive) represents the number of samples correctly predicted as positive; TN (True Negative) represents the number of samples correctly predicted as negative; FP (False Positive) represents the number of samples incorrectly predicted as positive (actually negative); and FN (False Negative) represents the number of samples incorrectly predicted as negative (actually positive). This metric reflects the overall classification reliability of the large-scale AI model, but in scenarios with imbalanced class distribution, it needs to be combined with other metrics for comprehensive evaluation.

[0089] Fault recall measures the ability of a large AI model to cover positive fault samples. The formula for calculating fault recall is: ; TP indicates a true positive and TN indicates a true negative. A high recall rate indicates that the model can effectively reduce the risk of missed detections.

[0090] The F1-score for each fault type is the harmonic mean of precision and recall, resolving the discrepancy between the two when there is class imbalance. ; Where Recall represents the fault identification recall rate, and Precision represents the fault identification precision rate. Precision is determined by the following formula: .

[0091] Confusion matrix analysis is used to construct a two-dimensional confusion matrix for fault categories. By analyzing the distribution relationships of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), the ability of the large-scale AI model to distinguish easily confused fault pairs is visualized. Furthermore, by combining ROC curves and AUC values, the classification performance stability of the large-scale AI model under different thresholds is evaluated.

[0092] Location and handling assessment: The similarity between the fault waveform and the template is calculated using the Dynamic Time Warping (DTW) algorithm to assess the location accuracy; the effectiveness of the handling plan is comprehensively evaluated from the dimensions of operational compliance, recovery timeliness, and disturbance suppression effect.

[0093] Real-time performance assessment: Monitor the inference latency of large AI models to ensure that the response time requirement is met within seconds.

[0094] Comparison and robustness verification: Comparative experiments were conducted with traditional expert systems and shallow models; and the robustness of large artificial intelligence models was tested under extreme scenarios such as high noise and multiple fault coupling.

[0095] Interpretability and reliability analysis: Visualize the contribution of key features through SHAP value heatmaps, LIME and other methods; provide confidence intervals through uncertainty quantification; establish a closed loop of diagnosis → treatment → verification, and use reinforcement learning reward function to quantify the actual treatment effect to continuously enhance the credibility of the system.

[0096] IV. Specialized Adaptation for Domestic Large Models To achieve the leap from a general model to an expert model in the power field, a deep transformation of the domestically developed large-scale artificial intelligence model was carried out.

[0097] Architecture and knowledge adaptation: Prompt word engineering: Design dedicated prompt templates that include DC fault phenomena, equipment parameters, and protection logic to guide the large-scale artificial intelligence model to generate standardized conclusions.

[0098] Domain knowledge injection: Structured knowledge such as power system standards and fault case libraries are embedded as contextual input.

[0099] Multimodal adaptation layer development: Extend the large-scale AI model encoder to enable it to jointly process multimodal data such as time-series signals, discrete events, and topological graphs.

[0100] Scenario fine-tuning strategy: Contrastive learning: Construct positive and negative sample pairs and enhance the model's ability to distinguish DC-specific fault modes by comparing the loss function.

[0101] Domain Adaptive Transfer: Initialization is performed using a pre-trained AC fault analysis model to accelerate convergence in DC scenarios.

[0102] Active learning iteration: Low-confidence predictions trigger manual annotation, forming a prediction-feedback-iteration closed loop.

[0103] Lightweight and Real-Time Inference: Model compression: Knowledge distillation is used to transfer knowledge from large AI models to lightweight models, and quantized perceptual training (QAT) is combined to reduce computational precision.

[0104] Hardware acceleration: TensorRT is used for operator fusion and optimization, deployed on GPU or ARM edge devices, and combined with strategies such as dynamic batch processing and feature caching to ultimately achieve second-level real-time inference that meets the requirements of power early warning.

[0105] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0106] The above embodiments mainly describe the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A method for intelligent fault analysis in ultra-high voltage direct current transmission systems, characterized in that, Including the following steps: S1. Acquire multi-source heterogeneous data of the UHVDC transmission system, including fault recording data, protection device event messages and synchronous phasor measurement data; perform standardization conversion and spatiotemporal alignment processing on the multi-source heterogeneous data to obtain preprocessed multimodal time series data. S2. Input the spatiotemporally aligned multimodal time-series data into a large-scale artificial intelligence model adapted with knowledge from the power sector, and perform multimodal feature extraction, including spatial features of transient signals, temporal dependency features, and topological correlation features, to generate a fused feature vector. S3. Input the fused feature vector into the fault diagnosis model and output the fault diagnosis result; The fault diagnosis results include at least fault type determination, fault location, and diagnosis confidence level; S4. Based on the fault diagnosis results, generate a fault analysis report and suggestions for handling strategies.

2. The intelligent fault analysis method for ultra-high voltage direct current transmission systems according to claim 1, characterized in that, Step S2 specifically includes: The large-scale artificial intelligence model uses a hybrid neural network composed of convolutional neural network modules and Transformer modules; The preprocessed multimodal time series data is input into a pre-constructed convolutional neural network module, and the transient spatial features characterizing the local morphology and spatial correlation of the signal in the fault recording data are extracted through the convolution operation of the convolutional neural network module. The multimodal time series data is input into a pre-built Transformer module, and long-term time-series dependency features characterizing the evolution of electrical quantities before and after a fault are extracted through the self-attention mechanism of the Transformer module. The transient spatial features are fused with the long-term temporal dependency features to generate the fused feature vector.

3. The intelligent fault analysis method for ultra-high voltage direct current transmission systems according to claim 2, characterized in that, Step S3 specifically includes: locating the fault location in the output fault diagnosis result, which is achieved through the following steps: Pre-construct a heterogeneous knowledge graph of equipment, protection devices, and electrical connections; The fusion feature vector generated in step S2 is mapped to the state information of nodes in the heterogeneous knowledge graph; By using graph neural networks to perform message propagation and node state updates in the heterogeneous knowledge graph, the propagation path of electrical disturbances can be inferred. By combining the action timing information contained in the protection device event messages in the multi-source heterogeneous data, the faulty component or the smallest fault section is determined on the propagation path, and the fault location is output.

4. The intelligent fault analysis method for ultra-high voltage direct current transmission systems according to claim 1, characterized in that, Step S3 specifically includes: the adaptation method of the large-scale artificial intelligence model adapted with knowledge from the power field includes: Construct a structured prompt word template, which includes descriptive fields of typical fault phenomena of UHVDC systems, numerical fields of key equipment parameters, and logical fields of protection action logic; Obtain training sample pairs consisting of historical failure case data and their corresponding standard expert analysis reports; Using historical fault case data from the training sample pairs as model input, and using diagnostic text generated by the structured prompt word template that matches the standard expert analysis report as training target, supervised fine-tuning training is performed on the basic artificial intelligence large model.

5. The intelligent fault analysis method for ultra-high voltage direct current transmission systems according to claim 1, characterized in that, Step S1 specifically includes: performing standardization transformation and spatiotemporal alignment processing on the multi-source heterogeneous data, which is achieved through the following steps: Extract the continuous time-series electrical quantity curves of the fault recording data from the multi-source heterogeneous data, as well as the discrete action event sequences in the event messages of the protection device; The optimal time warping path between the continuous time sequence curve of the electrical quantity and the discrete action event sequence is calculated using a dynamic time warping algorithm. Based on the optimal time curvature path, the continuous time sequence curve of the electrical quantity or the discrete action event sequence is calibrated on the time axis to align their timestamps.

6. The intelligent fault analysis method for ultra-high voltage direct current transmission systems according to claim 1, characterized in that, The method also includes a closed-loop iterative step: The method also includes a continuous iterative optimization step for the model: Implement an active learning strategy to automatically trigger manual annotation for fault cases where the confidence level of the AI ​​large model prediction is lower than a set threshold, and add the newly annotated data to the training set for incremental learning. Validated and effective handling cases are transformed into structured rules and dynamically updated into the knowledge graph used for strategy generation.

7. The intelligent fault analysis method for ultra-high voltage direct current transmission systems according to claim 1, characterized in that, The method also includes a real-time inference optimization step: The knowledge of the large artificial intelligence model is transferred to a lightweight reasoning model using knowledge distillation technology. The lightweight inference model is optimized by operator fusion and precision quantization using the TensorRT framework to improve diagnostic response speed.

8. A fault intelligent analysis system for ultra-high voltage direct current transmission systems, characterized in that, include: The data input module is used to acquire multi-source heterogeneous data from the UHVDC transmission system, including fault recording data, protection device event messages, and synchronous phasor measurement data; and to perform standardization conversion and spatiotemporal alignment processing on the multi-source heterogeneous data to obtain preprocessed multimodal time series data. An intelligent analysis module, connected to the data input module, is used to receive the multimodal time series data; The intelligent analysis module includes a large artificial intelligence model adapted with knowledge of the power field, used to perform multimodal feature extraction on the multimodal time series data to generate a fused feature vector, and includes a fault diagnosis model used to output fault diagnosis results based on the fused feature vector; The fault diagnosis results include at least fault type determination, fault location, and diagnosis confidence level; The decision output module, connected to the intelligent analysis module, is used to generate a fault analysis report and handling strategy recommendations based on the fault diagnosis results.

9. The intelligent fault analysis system for ultra-high voltage direct current transmission systems according to claim 8, characterized in that: The intelligent analysis module includes a fault type identification unit and a fault location unit; The input end of the fault type identification unit is connected to the data input module to receive the preprocessed multimodal time series data. It is equipped with a convolutional neural network, a Transformer module and a random forest classifier. The unit extracts spatial features from the multimodal time series data through the convolutional neural network, extracts the temporal dependencies through the Transformer module, and uses the random forest classifier to generate fault type determination results and their confidence levels. The input terminal of the fault location unit is connected to the fault type identification unit and the data input module, and is used to receive the fault type determination result and the multimodal time series data; This unit stores a heterogeneous knowledge graph of power grid equipment, protection devices, and topological connections. Based on the knowledge graph, it uses a graph neural network to infer the fault propagation path and combines the protection action timing in the multimodal time series data to output the location result of the faulty equipment or faulty section.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images