Intelligent power supply system state monitoring and fault early warning method and system

Abstract

The invention relates to the technical field of electric power system intelligent monitoring, and discloses an intelligent power supply system state monitoring and fault early warning method and system. According to the system, power grid operation parameters are collected in real time through a heterogeneous sensor array, multi-dimensional features are extracted through wavelet transform, a fault diagnosis model is constructed based on deep learning, precise early warning is achieved in combination with a dynamic threshold optimization algorithm, an optimal disposal scheme is generated based on an expert knowledge base, and remote data transmission is achieved through dual-channel communication. Real-time monitoring, fault early warning and intelligent decision support of the state of the power supply network are realized, and the operation reliability and the operation and maintenance efficiency of the power grid are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power systems, and discloses a method and system for intelligent power supply system status monitoring and fault early warning. Background Technology

[0002] Current power supply system status monitoring mainly relies on periodic inspections and fixed threshold alarm mechanisms, which suffer from significant lag and high false alarm rates. Traditional monitoring equipment often uses single sensors to collect data, making it difficult to comprehensively reflect the power grid's operating status, and there is a lack of effective data collaboration mechanisms between monitoring units.

[0003] Existing fault diagnosis methods are mainly based on empirical threshold judgments, which cannot adapt to the complex and ever-changing power grid operating environment. When new fault modes occur in the system, traditional methods often fail to identify the fault type and location in a timely and accurate manner, leading to delays in fault handling.

[0004] Current early warning systems generally use static threshold settings, which cannot be dynamically adjusted according to the actual operating status of the power grid and environmental factors. This rigid threshold setting method easily generates a large number of invalid alarms, increasing the workload of operation and maintenance personnel.

[0005] In terms of data transmission, existing systems mostly use a single communication channel, which leads to data loss and delays. Especially when power grid faults occur, critical monitoring data cannot be transmitted to the control center in a timely manner, severely affecting the efficiency of fault handling.

[0006] Traditional fault handling methods rely primarily on human experience and lack systematic knowledge management and intelligent decision support. This approach is ill-suited to the increasingly complex operating conditions of modern power grids and can easily lead to the escalation of faults.

[0007] Therefore, we propose a method and system for intelligent power supply system status monitoring and fault early warning to solve the problems encountered above. Summary of the Invention

[0008] This invention provides a method and system for intelligent power supply system status monitoring and fault early warning. It achieves high-precision synchronous acquisition of multiple power grid parameters through a heterogeneous sensor array, constructs a multi-dimensional feature extraction and fault diagnosis model using a wavelet transform and deep learning fusion algorithm, and combines a dynamic threshold optimization mechanism to achieve accurate fault identification and early warning. The system establishes an intelligent decision support module based on an expert knowledge base and ensures reliable transmission of monitoring data through a dual-channel redundant communication network, forming a complete technology chain from data acquisition, feature analysis, fault diagnosis to early warning decision-making.

[0009] This invention innovatively combines time-frequency analysis technology with deep learning methods to design an adaptive dynamic threshold adjustment algorithm, solving the problem of false alarms and missed alarms in traditional monitoring systems under complex operating conditions. The system achieves real-time panoramic monitoring of the power supply network's operating status, intelligent identification and precise location of fault characteristics, dynamic optimization of early warning thresholds, and intelligent generation of handling plans, significantly improving the safety and reliability of the power supply system.

[0010] To achieve the above objectives, this invention provides an intelligent power supply system status monitoring and fault early warning system. The system includes a power supply network data acquisition module, a status feature extraction and analysis module, a fault diagnosis and location module, a dynamic adjustment module for early warning thresholds, an intelligent decision-making and control module, a communication and remote monitoring module, an energy efficiency analysis and optimization module, and a self-learning system upgrade module, wherein: The power supply network data acquisition module is used to synchronously acquire electrical parameters of the power supply network based on multi-sensor fusion technology and generate a real-time power data set. The state feature extraction and analysis module is used to perform feature extraction on the original power data using a wavelet transform algorithm to form a multidimensional state feature vector. The fault diagnosis and location module is used to perform pattern recognition on the operating status of the power distribution network using a deep learning model, and output fault type and location information. The early warning threshold dynamic adjustment module is used to dynamically optimize the early warning threshold parameters by combining historical operating data, and generate an adaptive threshold rule base. The intelligent decision-making and control module is used to perform decision analysis on the fault diagnosis results based on the expert knowledge base and generate a set of control strategy instructions. The communication and remote monitoring module is used to remotely transmit monitoring data through a dual-channel communication protocol and establish a cloud-based real-time monitoring interface. The energy efficiency analysis and optimization module is used to perform dynamic analysis of the energy consumption of the power distribution network using the load forecasting model, and to construct a set of energy efficiency optimization schemes. The self-learning system upgrade module is used to update the system algorithm model online using an incremental learning mechanism, forming a continuously evolving diagnostic system.

[0011] Optionally, when the power supply network data acquisition module synchronously acquires electrical parameters of the power supply network based on multi-sensor fusion technology, it includes: The IEEE 1588v2 precision clock protocol is used to achieve microsecond-level time synchronization of the real-time power data set; Configure an isolation sampling circuit to resist electromagnetic interference and ensure the accuracy of the electrical parameters acquisition under strong electric field environment; An adaptive filtering algorithm is deployed to eliminate high-frequency noise interference in the real-time power data set.

[0012] Optionally, when the state feature extraction and analysis module uses the wavelet transform algorithm to perform feature extraction on the original power data, it includes: A 6-layer decomposition tree based on the Mallat algorithm was constructed to process the time-frequency characteristics of the original power data; The frequency domain stability index obtained by singular value decomposition is fused into the multidimensional state feature vector; Configure a dynamic threshold mechanism to automatically remove abnormal feature components from the multidimensional state feature vector.

[0013] Optionally, when the fault diagnosis and location module uses a deep learning model to perform pattern recognition on the operating status of the power distribution network, it includes: A residual attention mechanism is employed to enhance the deep learning model's ability to capture transient features; Construct a topology constraint loss function to ensure that the fault location information conforms to the physical connection relationship of the power grid; Deploy the model interpretation module to visualize and output a heatmap of the decision-making basis for the fault type.

[0014] Optionally, when the early warning threshold dynamic adjustment module dynamically optimizes the early warning threshold parameters by combining historical operating data, it includes: The dynamic distribution characteristics of the historical operating data are extracted using a sliding time window statistical method. A threshold optimizer based on the Q-learning algorithm is constructed to generate the adaptive threshold rule base; Configure a robust outlier handling mechanism to prevent erroneous adjustment of the warning threshold parameter.

[0015] Optionally, when the intelligent decision-making and control module performs decision analysis on the fault diagnosis results based on the expert knowledge base, it includes: A case-based reasoning engine is used to match historical handling solutions in the expert knowledge base. A multi-objective optimization model is constructed to generate the Pareto front solution of the control strategy instruction set; Deploy an instruction verification module to verify the compatibility of the control strategy instruction set with the current system state.

[0016] Optionally, when the communication and remote monitoring module remotely transmits monitoring data via a dual-channel communication protocol, it includes: The monitoring data transmission security is ensured by using IPSec VPN and quantum encryption dual-channel. Configure an adaptive bitrate adjustment algorithm to optimize the bandwidth utilization of the cloud real-time monitoring interface; Deploy a data priority scheduler to ensure that critical alarm data is transmitted with priority in the dual-channel communication protocol.

[0017] Optionally, when the energy efficiency analysis and optimization module performs dynamic analysis of the distribution network energy consumption using a load forecasting model, it includes: A hybrid prediction model integrating LSTM and XGBoost is constructed to process the spatiotemporal characteristics of the power distribution network energy consumption; Sensitivity analysis was used to identify the key control parameters in the energy efficiency optimization scheme set; A rolling optimization mechanism is deployed to update the recommended weights of the energy efficiency optimization scheme set every 15 minutes.

[0018] Optionally, when the self-learning system upgrade module uses an incremental learning mechanism to update the system algorithm model online, it includes: Configure an elastic weight solidification algorithm to prevent catastrophic forgetting in the incremental learning mechanism; Deploying the model performance drift detector triggers a version rollback of the continuous evolution diagnostic system. Construct a knowledge distillation pipeline to migrate features from the historical model to the new version of the system algorithm model.

[0019] To address the above problems, the present invention also provides a method for intelligent power supply system status monitoring and fault early warning, the method comprising: The power supply network electrical parameters are synchronously collected based on multi-sensor fusion technology to generate a real-time power data set. Wavelet transform algorithm is used to perform feature extraction on the original power data to form a multidimensional state feature vector; Deep learning models are used to perform pattern recognition on the operating status of power distribution networks and output fault type and location information. By combining historical operational data, the warning threshold parameters are dynamically optimized to generate an adaptive threshold rule base. Based on an expert knowledge base, decision analysis is performed on the fault diagnosis results to generate a set of control strategy instructions; The monitoring data is remotely transmitted via a dual-channel communication protocol, and a real-time cloud monitoring interface is established. Dynamic analysis of power distribution network energy consumption is performed using load forecasting models to construct a set of energy efficiency optimization schemes; An incremental learning mechanism is used to update the system algorithm model online, forming a continuously evolving diagnostic system.

[0020] This invention provides a complete solution for intelligent power supply system status monitoring and fault early warning. Its core lies in constructing a comprehensive technical system encompassing data acquisition and intelligent decision-making. The system employs heterogeneous sensor arrays and precision clock synchronization technology to achieve high-precision synchronous acquisition of multiple parameters of the power supply network. Through a fusion algorithm of wavelet packet transform and deep neural networks, a multi-scale feature extraction and fault mode recognition model is established, significantly improving the diagnostic accuracy of complex faults.

[0021] Regarding the early warning mechanism, this invention innovatively proposes a dynamic threshold optimization algorithm based on adaptive operating conditions. By analyzing historical data and environmental factors in real time, it automatically adjusts the early warning threshold parameters, effectively solving the false alarm and missed alarm problems of traditional fixed threshold systems. The system also establishes an expert knowledge base containing typical fault cases, which, combined with the real-time power grid topology, can intelligently generate optimal handling solutions.

[0022] This invention features a specially designed dual-channel redundant communication architecture, integrating power grid and wireless communication technologies to ensure reliable transmission of monitoring data and control commands. The system supports a hybrid processing mode combining edge computing and cloud collaboration, meeting both real-time requirements and possessing big data analytics capabilities. Through an incremental learning mechanism, the system can continuously optimize its algorithm model to adapt to changes in the power grid operating environment.

[0023] The resulting intelligent monitoring and early warning system achieves core functions such as panoramic perception of the power supply network status, early warning and precise fault location, and intelligent decision support for handling solutions. Compared with traditional technologies, this invention significantly improves fault identification rate, early warning accuracy, and response speed, providing reliable technical assurance for the safe operation of the smart grid.

[0024] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention significantly improves the comprehensiveness and accuracy of power supply system status monitoring. Through heterogeneous sensor fusion technology and a precise time synchronization mechanism, it achieves high-precision synchronous acquisition of multiple power grid parameters, overcoming the limitations of traditional single-sensor monitoring and providing a more complete and accurate data foundation for fault diagnosis.

[0025] 2. This invention significantly enhances the ability to identify complex faults and the accuracy of early warning. Employing an intelligent diagnostic algorithm that combines time-frequency analysis with deep learning, it can effectively identify various complex and gradual faults. Combined with a dynamic threshold optimization mechanism, it significantly reduces false alarm and false negative rates, achieving early and accurate fault warnings.

[0026] 3. This invention comprehensively optimizes fault handling efficiency and system adaptability. The intelligent decision support system based on an expert knowledge base can quickly generate optimal handling solutions. Combined with a dual-channel reliable communication network, it ensures timely instruction delivery. Simultaneously, through an incremental learning mechanism, it continuously optimizes system performance, enabling the entire monitoring and early warning system to continuously evolve. Attached Figure Description

[0027] Figure 1 This is a system architecture diagram of an intelligent power supply system status monitoring and fault early warning system provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for monitoring the status and providing early warning of faults in an intelligent power supply system according to an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0029] Furthermore, the order of steps in the following method embodiments is merely an example and not a strict limitation.

[0030] like Figure 1 The figure shown is a system architecture diagram of an intelligent power supply system status monitoring and fault early warning system provided in an embodiment of the present invention.

[0031] In this embodiment of the invention, the power supply network data acquisition module is used to synchronously acquire electrical parameters of the power supply network based on multi-sensor fusion technology, and generate a real-time power data set, including: The hardware architecture of the power supply network data acquisition module adopts a distributed deployment scheme, configuring heterogeneous sensor arrays composed of various types of sensors at key nodes in the substation. This array includes electromagnetic voltage transformers with 0.2-level accuracy, electronic current sensors with 0.5S-level accuracy, and fiber optic temperature sensors with ±0.5℃ accuracy, forming a comprehensive monitoring capability for the electrical parameters of the power grid. Each sensor unit is connected to a local data acquisition terminal via shielded twisted-pair cables. The terminal has a built-in 24-bit high-precision AD conversion chip with a sampling rate of up to 10kS / s, ensuring complete capture of the dynamic characteristics of the power signal. To eliminate clock deviations between different sensors, the system adopts a clock synchronization scheme based on the IEEE 1588v2 precision time protocol, distributing synchronization signals to each acquisition terminal through a master clock source to achieve microsecond-level alignment of the sampling time scale across the entire substation.

[0032] In terms of signal conditioning circuit design, the system employs dedicated signal conditioning modules tailored to the output characteristics of different types of sensors. Voltage signals are processed through a precision resistor divider network and an isolation amplifier, current signals are conditioned by an integrator circuit and a programmable gain amplifier, and temperature signals are converted via a thermocouple cold junction compensation circuit. All conditioned analog signals are processed by an anti-aliasing filter with a cutoff frequency set to 2kHz, conforming to the Nyquist sampling theorem. The AD conversion stage uses a Σ-Δ analog-to-digital converter, combined with oversampling technology to effectively improve the signal-to-noise ratio, achieving a conversion resolution of 24 bits and a dynamic range better than 110dB. The converted digital signals are temporarily stored in the FPGA chip's FIFO memory, awaiting further processing.

[0033] The data preprocessing stage primarily accomplishes three key tasks: First, it applies digital filtering based on moving average to the sampled values, with a window length set to 10 power frequency cycles to effectively suppress random interference; second, it performs sensor nonlinearity correction based on the least squares method to eliminate the influence of transformer angle and ratio errors; finally, it executes data encapsulation based on the IEC 61850-9-2 standard, organizing the three-phase voltage and current sampled values ​​according to a specific data structure and adding quality flag bits and timestamp information. The preprocessing process is completed in real-time on a dedicated DSP chip, with processing latency controlled within 1ms to ensure data timeliness.

[0034] The system employs a dual data verification mechanism to ensure data quality. The first verification is performed locally at the acquisition terminal, identifying abnormal data by comparing the rate of change of adjacent sampling points. The second verification is executed on the station control layer server, verifying its validity based on the three-phase balance principle and the law of power conservation. Data that passes verification is marked as valid and stored in a circular buffer awaiting upload. For invalid data, the system automatically triggers a resampling mechanism and records the anomaly information in the event log. The data buffer design uses double buffering technology to ensure that no sampled data is lost during data transmission.

[0035] The final output standardized power dataset contains the following key information: instantaneous sampled values ​​of three-phase voltage and current, calculated active / reactive power values ​​for each phase, harmonic distortion rate, frequency deviation and other derived quantities, and complete data quality identifiers. The dataset is organized according to the CID file format defined by the IEC 61850 standard and supports both SV and MMS communication services. The system also supports local data caching, which can save at least 72 hours of complete monitoring data during network interruptions and automatically retransmit it after communication is restored.

[0036] To adapt to different application scenarios, the system provides configurable sampling strategy management. Users can flexibly set parameters such as sampling frequency and data transmission period according to the importance of the monitored object. The system also supports triggered sampling mode, automatically switching to high sampling rate mode to continuously record transient waveforms when abnormal events such as voltage drops or current surges are detected. All configuration parameters are managed through XML configuration files, supporting online modification and remote distribution, greatly improving the system's adaptability and flexibility.

[0037] In this embodiment of the invention, the state feature extraction and analysis module is used to perform feature extraction on the original power data using a wavelet transform algorithm to form a multidimensional state feature vector, including: In the state feature extraction and analysis module, the acquired raw voltage and current signals are first preprocessed. Preprocessing includes three main steps: DC component elimination, power frequency component normalization, and noise suppression. A zero-phase digital filter is used to eliminate DC offset in the signal; power frequency component synchronization and amplitude normalization are achieved through phase-locked loop technology; and finally, a wavelet threshold denoising algorithm is used to suppress measurement noise. After preprocessing, the system selects the db4 wavelet as the basis function for decomposition. The db4 wavelet has a fourth-order vanishing moment, which can better balance time-frequency resolution and is particularly suitable for power signal analysis. Based on the frequency band distribution characteristics of typical power system faults, the decomposition level is determined to be 5 levels, which can cover the important frequency band of 0-2.5kHz while ensuring computational efficiency.

[0038] The wavelet packet decomposition process employs a fully binary tree structure, where each layer decomposes the signal into approximation coefficients and detail coefficients. The system uses the Mallat fast algorithm for decomposition calculations, achieving rapid transformation through convolution and sampling operations. To maintain signal energy conservation, an orthogonal wavelet basis is selected, and the filter coefficients during the decomposition process are strictly normalized. In terms of hardware implementation, the algorithm is accelerated using an FPGA parallel processing architecture, allowing for a 5-level decomposition of a single signal channel to be completed within 1ms. A symmetrical extension approach is used during decomposition to handle boundary effects and avoid information loss at both ends of the signal. The decomposition coefficients of each node are normalized to ensure comparability of features across different scales.

[0039] For steady-state feature extraction, the system calculates the energy entropy value of the wavelet packet coefficients for each frequency band. First, the sum of squares of the coefficients at each node is calculated as the frequency band energy. Then, all frequency band energies are normalized, and finally, their information entropy value is calculated. Transient feature extraction focuses on high-frequency detail coefficients, employing a peak detection algorithm to capture abrupt changes and calculating parameters such as amplitude, rise time, and oscillation frequency. To comprehensively characterize signal features, the system also extracts statistical features of the wavelet coefficients at each level, including mean, variance, skewness, and kurtosis. All feature parameters are standardized to eliminate the influence of dimensions.

[0040] The initial feature set contained over 50 feature parameters. To avoid the "curse of dimensionality," the system employed Principal Component Analysis (PCA) for dimensionality reduction. Before PCA, the feature correlation coefficient matrix was calculated to remove highly correlated redundant features. Principal components with eigenvalues ​​greater than 1 were retained, achieving a cumulative contribution rate of over 95%. The final 20 selected features included: 5 energy entropy features, 8 transient feature parameters, 4 statistical features, and 3 time-domain features. The feature vectors formed by these features retained the main information of the original signal while significantly reducing the data dimensionality. The dimensionality-reduced feature vectors were stored as fixed-length arrays for easier processing by subsequent modules.

[0041] The system establishes a feature quality assessment mechanism, evaluating the distinguishing ability of features by calculating the ratio of intra-class distance to inter-class distance. The effectiveness of the feature set is periodically tested using newly collected fault samples. When a decrease in feature distinguishability is detected, a feature optimization process is triggered. The optimization process includes re-evaluating the weights of each feature, adjusting PCA parameters, and updating feature selection criteria. The system maintains a feature library, storing historically effective feature combinations. When significant changes occur in the power grid operation mode, it can quickly revert to previous feature sets. All feature parameters are timestamped and labeled with operating conditions, supporting feature traceability analysis.

[0042] The final generated feature vectors are output using a standardized data structure, comprising three parts: a feature value array, a feature quality flag, and metadata. The feature value array is a 20-dimensional floating-point vector, sorted by importance; the quality flag indicates the reliability level of each feature; and the metadata records the feature extraction time, signal channel information, and processing parameters. The system provides two interface methods: a real-time interface that pushes the latest feature vectors at a frequency of 50Hz, and a batch processing interface that supports querying historical feature data by time range. To ensure data security, all feature vector outputs are digitally signed to prevent data tampering. The feature extraction module also supports remote configuration updates, allowing maintenance personnel to adjust feature extraction strategies as needed.

[0043] In this embodiment of the invention, the fault diagnosis and location module is used to perform pattern recognition on the operating status of the power distribution network using a deep learning model, and output fault type and location information, including: The fault diagnosis and localization module employs an innovative hybrid neural network architecture, featuring a cascaded structure of a three-layer convolutional neural network and a two-layer long short-term memory network. The first CNN layer uses 64 3×3 convolutional kernels with the ReLU activation function, focusing on extracting local spatial patterns from feature vectors. The second CNN layer uses 128 3×3 convolutional kernels to increase the level of feature abstraction. The third CNN layer is adjusted to 256 1×1 convolutional kernels to achieve interactive integration between feature channels. Each CNN layer is followed by a batch normalization layer and a max pooling layer with a 2×2 pooling window. The CNN module output is expanded through a Flatten layer and then connected to a 128-neuron LSTM layer, whose gating mechanism effectively captures temporal dependencies. The second LSTM layer is reduced to 64 neurons, focusing on extracting key temporal features. Residual connections are used between the two LSTM layers to alleviate the gradient vanishing problem. The final layer is a fully connected layer with a Softmax activation function, outputting the probability distribution of each fault type.

[0044] This hybrid neural network employs a phased training strategy. First, it uses ImageNet pre-trained CNN weights for transfer learning, and then fine-tunes the model on a power fault dataset. The training process uses cosine annealing with hot restart for learning rate scheduling, with an initial learning rate of 0.001 and a minimum learning rate of 0.00001. The AdamW optimizer is used, with a weight decay coefficient of 0.01 to effectively prevent overfitting. The loss function is focus loss, and the imbalance of fault sample classes is addressed by adjusting parameters α=0.25 and γ=2. Training data augmentation strategies include random noise injection, random feature channel masking, and temporal jitter to improve model robustness. The model is trained in parallel on four NVIDIA Tesla V100 GPUs with a batch size of 256, and mixed-precision computing is used to accelerate the training process.

[0045] The fault probability distribution output by the model is deeply correlated with the power grid topology database. The topology database is stored using a graph database and includes the connection relationships and electrical parameters of equipment such as substations, lines, and switches. The location algorithm first determines the dominant fault type based on the fault probability, and then calculates the fault confidence level of each equipment node based on the propagation characteristics of that fault type. For short-circuit faults, the impedance method is used to calculate the fault distance; for ground faults, the zero-sequence current distribution analysis method is applied. The system establishes a fault hypothesis-verification mechanism, generating multiple possible fault location hypotheses, and then verifying them through topological connectivity and protection action logic. The final output is a structured location result containing information such as the faulty equipment ID, fault distance, and location confidence score.

[0046] During the deployment phase, the model underwent several optimizations: First, TensorRT was applied for model quantization and compression, converting the FP32 model to FP16 precision, resulting in a 1.8x speedup inference. Second, model pruning was implemented, removing low-contribution convolutional kernels and reducing the model size by 40%. Finally, a dedicated caching mechanism was designed to directly access cached results for recurring feature patterns. At the hardware level, Intel Xeon Scalable processors were used in conjunction with the OpenVINO toolkit to optimize the CPU inference pipeline, keeping single inference latency below 15ms. In case of unexpected failures, the system activated an emergency inference mode, temporarily prioritizing computational resources to ensure real-time diagnosis of critical faults.

[0047] The system establishes a continuous learning mechanism, periodically collecting new fault samples for model iteration and updates. An elastic weighting algorithm is employed to prevent new knowledge from overwriting old knowledge, preserving important connection weights. A model performance monitoring dashboard is set up to track indicators such as accuracy and recall for each fault category in real time. When a sustained decline in the recognition performance of a certain fault type is detected, a targeted retraining process is automatically triggered. The updated model is first run in shadow mode, compared with the production model in parallel inference, and only after successful verification is it released in a canary deployment. The system maintains multiple version snapshots of the model, supporting rapid rollback to historical stable versions.

[0048] A dedicated fault diagnosis visualization interface was developed, displaying the feature maps extracted by CNN and the attention weights of LSTM in 3D. For each diagnostic result, an explanation based on SHAP values ​​is provided, highlighting the feature dimensions that have the greatest impact on the diagnostic conclusion. Location results are overlaid on the power grid geographic information system, with different colors used to indicate the probability of faults in each device. The system generates a complete report including fault waveforms, feature analysis, diagnostic basis, and remedial recommendations, supporting export in PDF and Excel formats. All diagnostic process data is stored on the blockchain to ensure the credibility of the audit trail.

[0049] In this embodiment of the invention, the early warning threshold dynamic adjustment module is used to dynamically optimize the early warning threshold parameters by combining historical operating data, and generate an adaptive threshold rule base, including: The dynamic adjustment module for early warning thresholds adopts a hierarchical adaptive model architecture, with a sliding window-based machine learning framework at its core. The system first establishes a multi-source data acquisition pipeline, extracting complete monitoring data from the historical database for the past 7 days, including electrical parameters, fault records, ambient temperature, and load rate curves. This heterogeneous data undergoes a unified preprocessing workflow: electrical parameters are subjected to outlier removal and standardization, temperature data is converted to hourly averages, and load rate data is resampled at 15-minute intervals. The preprocessed data is stored in a temporary analysis database for use by the threshold calculation engine. The system features a specially designed data quality assessment mechanism that automatically excludes periods with a data missing rate exceeding 5%, ensuring the reliability of the calculation results. Data time alignment is handled using linear interpolation to guarantee the synchronization of the time series of each parameter.

[0050] The sliding window mechanism employs a circular buffer data structure, with a fixed window size of 7 days, sliding once per minute. The system maintains three independent data windows: a main window stores electrical monitoring parameters, an environmental window stores operating condition data such as temperature and humidity, and an event window records fault and alarm history. An incremental update strategy is used during window sliding, processing only newly entering and exiting data points, significantly reducing computational overhead. To improve real-time performance, window data is stored in fragments based on feature dimensions, supporting parallel computing. The system implements a dynamic window adjustment algorithm; when a sudden change in the power grid operating mode is detected, the window is automatically shortened to 3 days, accelerating the threshold adaptation process to the new operating conditions. Window data uses a columnar storage format, optimizing query efficiency for large data volumes.

[0051] The threshold calculation employs a five-layer fuzzy logic system: the first layer is the fuzzification layer, converting input variables into membership degrees; the second layer uses an expert rule base for reasoning; the third layer aggregates the outputs of each rule; the fourth layer defuzzifies to generate precise values; and the fifth layer performs threshold rationality verification. Temperature variables are divided into three fuzzy sets: "low temperature," "normal temperature," and "high temperature," while load rates are divided into three fuzzy sets: "light load," "medium load," and "heavy load," both using trapezoidal membership functions. The rule base contains 27 expert rules, such as "IF temperature is high temperature AND load rate is heavy load THEN voltage threshold reduced by 10%." The system implements a dynamic rule weight adjustment mechanism, automatically updating weights based on historical rule accuracy. Defuzzification uses the centroid method to calculate the final threshold, which undergoes a rationality check based on statistical process control before output.

[0052] The system employs a multi-objective optimization method to coordinate the threshold relationships of various monitoring parameters. First, a parameter correlation matrix is ​​established to quantify the physical coupling relationships between various electrical quantities. Then, a constrained optimization problem is constructed, with the objective function simultaneously considering: minimizing the false alarm rate, maximizing the fault detection rate, and ensuring the smoothness of threshold changes. An improved particle swarm optimization algorithm is used to find the Pareto optimal solution in a 100-dimensional search space. For key parameters, a dual threshold strategy is implemented: a dynamically adjusted primary threshold and a fixed-boundary auxiliary threshold are set. Manual verification is triggered when the primary threshold exceeds the auxiliary threshold's range. The calculation results are distributed to each edge computing node using a consistent hashing algorithm to ensure synchronized threshold updates across the entire network.

[0053] The dynamic threshold rule base is implemented using a versioned document database, with each threshold parameter storing a complete modification history. The rule base comprises four sub-bases: a basic threshold database storing device factory settings, an adaptive threshold database storing currently valid values, a temporary threshold database for testing new algorithms, and a historical threshold database archiving all versions. Version control uses semantic numbering and supports rollback by time point. Rule updates follow a two-phase commit protocol: the effect of the new threshold is first verified in shadow mode, and only after confirmation is it committed to the production environment. The system implements role-based access control, and threshold modifications require an electronic signature from the operations manager. The rule base is automatically backed up to an off-site disaster recovery center every hour to ensure data security.

[0054] The system incorporates a comprehensive performance monitoring system to track the effectiveness of each threshold in real time. Key metrics include false alarm rate, false negative rate, threshold adjustment frequency, and computation latency. Monitoring data drives the model's self-optimization process: when the false alarm rate of a parameter exceeds 5% for 24 consecutive hours, threshold recalculation is automatically triggered; when the ambient temperature change exceeds 5°C, adaptive adjustments are initiated in advance. The system periodically performs global optimization, retraining the parameters of the fuzzy logic model and assessing whether the rule base needs expansion. All optimization operations are recorded in the audit log, supporting complete operation traceability. The system also provides visual analysis of threshold effects, displaying threshold change trends under different operating conditions through heatmaps, assisting maintenance personnel in understanding the working status of the adaptive mechanism.

[0055] In this embodiment of the invention, the intelligent decision-making and control module is used to perform decision analysis on the fault diagnosis results based on an expert knowledge base, and generate a set of control strategy instructions, including: The expert knowledge base in the intelligent decision-making and control module adopts a multi-layered distributed architecture, comprising three core components: a case library, a rule library, and a model library. The case library stores over 2000 standardized typical fault cases. Each case is structured using a framework representation, including four parts: fault feature vector, environmental operating parameters, handling plan, and evaluation of handling effect. Case features are modeled using ontology methods, establishing a unified semantic description standard to ensure the comparability of cases from different sources. The knowledge base implements a distributed storage strategy: frequently accessed hot cases are stored in the in-memory database Redis, historical cases are stored in the time-series database InfluxDB, and relationships between cases are managed through the graph database Neo4j, enabling millisecond-level complex relational queries. The system periodically extracts new cases from data sources such as SCADA systems and fault recording devices through an ETL process, and incrementally updates the knowledge base after expert review.

[0056] The case-based reasoning engine employs a four-stage cyclical architecture: retrieval, reuse, correction, and retention. The retrieval stage combines near-nearest neighbor algorithms and semantic similarity calculations to find the Top-K historical cases from the knowledge base that best match the current fault. The system innovatively designs a three-tiered retrieval strategy: the first tier uses Euclidean distance based on feature vectors for initial screening; the second tier uses secondary filtering considering operating condition similarity; and the third tier uses semantic-level fine-grained matching based on the application domain ontology. The reuse stage uses a genetic algorithm-based solution adaptation technique to adjust historical handling solutions to fit the current scenario. The correction stage introduces digital twin technology to pre-simulate handling solutions in a power grid simulation environment and verify their effectiveness. The retention stage automatically structures and stores the verified new cases and their handling experience into the knowledge base, while simultaneously updating the case weight parameters. The entire reasoning process is implemented on the distributed computing framework Spark, enabling the matching calculation of thousands of cases to be completed in seconds.

[0057] The system establishes a multi-objective decision-making model to evaluate the merits of each candidate disposal scheme, considering five key dimensions: impact on power supply reliability, equipment safety margin, operational complexity, economic cost, and execution timeliness. Each dimension is configured with an adjustable weight coefficient, supporting strategy preference settings under different operating modes. The decision algorithm first uses the TOPSIS method to calculate the proximity of each scheme, then applies a constraint satisfaction problem solver to eliminate schemes that violate safety regulations, and finally explores the optimal operation sequence through Monte Carlo tree search. For load transfer schemes, the system integrates a real-time power flow calculation engine to predict the capacity margin of each transfer path, ensuring that no new overload problems are triggered after the operation. The optimized control instruction set is encapsulated in JSON-LD format, containing structured information such as circuit breaker operation timing tables, protection setting adjustment suggestions, and load transfer path topology, along with expected effect evaluations and fallback plans for each scheme.

[0058] The generated optimized control instruction set must undergo a rigorous security verification process before being issued for execution. The verification process consists of three levels: syntax verification checks whether the instruction format conforms to the IEC 61850 standard; logic verification verifies the temporal logic of the operation sequence to prevent conflicting operations; and physical verification simulates the entire execution process in the digital twin system, predicting changes in electrical parameters at each node. The system implements an intelligent error correction mechanism, automatically triggering a re-optimization of the process when a potentially dangerous operation is detected. The verified instruction set is converted into a standardized SCL configuration file, accompanied by a digital signature and timestamp, and transmitted to each execution terminal via a secure gateway. For particularly complex fault scenarios, the system supports generating multiple alternative solutions, with each solution's advantages and disadvantages compared for operators' final decision-making reference. All generated instruction sets are stored in a blockchain-based evidence storage system to ensure the traceability of the operation process.

[0059] The system employs a triple knowledge evolution mechanism to maintain the timeliness of the expert knowledge base: a dynamic case quality assessment system periodically re-evaluates the applicability of historical cases, marking outdated cases as obsolete; an online learning system automatically adjusts the similarity calculation weights of cases by analyzing the deviation between actual and predicted outcomes; and an active learning system identifies weaknesses in the knowledge base, intelligently generates simulated failure scenarios, and guides experts to supplement missing cases. Knowledge updates follow a strict quality control process: new cases must be cross-validated with at least three historical cases, and major rule modifications require a vote by the expert committee. The system implements knowledge traceability, allowing users to query the historical execution records and effect statistics of each recommended solution, providing decision-making confidence references for operators. The knowledge base generates a monthly health report, statistically analyzing key indicators such as case coverage and solution adoption rate, guiding the continuous optimization of the knowledge base.

[0060] To enhance the system's interpretability and human-computer collaboration efficiency, an advanced decision support interface was developed. The 3D visualization system visually presents matched similar cases, key considerations for recommended solutions, and expected outcomes. The system provides a "decision lens" function, allowing operators to adjust different target weights and observe the real-time trajectory of recommended solutions. For complex decisions, the system generates a comparison matrix, detailing the evaluation results of each alternative solution across 12 dimensions, including technical feasibility and safety risks. In emergencies, the system supports voice interaction, allowing operators to query the basis and expected consequences of specific operations using natural language. All human decision-making actions and feedback are recorded by the system for optimizing subsequent recommendation algorithms. The system also integrates virtual conferencing capabilities, supporting remote multi-expert consultations on handling major faults and enabling distributed collaborative decision-making.

[0061] In this embodiment of the invention, the communication and remote monitoring module is used to remotely transmit monitoring data through a dual-channel communication protocol and establish a cloud-based real-time monitoring interface, including: The communication and remote monitoring module employs a heterogeneous hardware architecture to achieve parallel transmission of power line carrier and 4G dual channels. The power line carrier channel is based on TI's TMS320F28379D digital signal processor, integrating a professional power line carrier modem chip, operating in the 1-30MHz frequency band, and using OFDM modulation technology to achieve a maximum communication rate of 500kbps on 400V power distribution lines. The 4G channel uses Huawei's ME909s-821 industrial-grade communication module, supporting the LTE Cat4 standard, with a downlink rate of up to 150Mbps. The system features an intelligent channel selector, implemented on a Xilinx Artix-7 FPGA, which monitors the signal quality, transmission delay, and packet loss rate of both channels in real time, dynamically allocating transmission paths according to service type. Key components employ redundant design, and the power system is configured with dual power supplies and supercapacitor backup to ensure communication continuity under extreme conditions.

[0062] In terms of protocol stack design, the system implements a layered and optimized hybrid protocol architecture. The power line carrier channel employs a streamlined protocol stack: the physical layer uses the IEEE 1901.1 standard broadband power line communication technology; the data link layer features a custom-developed TDMA mechanism MAC protocol with dynamically adjustable time slot allocation; the network layer adapts to IPv6 over PLC; and the transport layer uses a lightweight UDP protocol, coupled with the application layer's GOOSE fast message service. The 4G channel uses a standard TCP / IP protocol stack, but is optimized for the characteristics of power monitoring data: the application layer implements a power-specific extension of the MQTT protocol, supporting QoS classification; the transport layer enables TCP fast open and BBR congestion control algorithms; and the network layer implements software-defined networking-based traffic shaping to ensure the transmission quality of monitoring data. Data synchronization and state sharing between the two channels are achieved through a dedicated tunnel protocol.

[0063] The communication security system employs a "defense-in-depth" strategy, implementing multi-layered security measures. At the authentication level, each terminal device is embedded with a digital certificate using the national cryptographic algorithm SM2, establishing a two-way authentication mechanism. Data transmission encryption uses the national cryptographic algorithm SM4; the power line carrier channel uses CTR mode, and the 4G channel uses GCM mode. Each data packet is appended with a CRC-32 checksum and sequence number to prevent tampering and replay attacks. The system implements dynamic key management, automatically updating session keys every 24 hours based on the SM3 hash algorithm, with manual key reset available in emergencies. A firewall based on deep packet inspection is deployed at the network boundary to identify and block abnormal communication patterns. All security events are uploaded to the security operations and maintenance center in real time, forming a complete security audit trail. The system has passed the Level 3 security certification of the State Grid Corporation of China's "Regulations on Security Protection of Power Monitoring Systems".

[0064] The system employs an intelligent data scheduling engine to achieve optimal path allocation for business data. Critical control commands are transmitted via power line carrier channels using a publish / subscribe model with the GOOSE protocol, ensuring transmission latency is strictly controlled within 20ms. Monitoring data is processed in three categories: real-time monitoring data is transmitted via the MQTT protocol on a 4G channel, with a configurable sampling period of 10ms-1s; near-real-time data is compressed and transmitted in batches; and historical data is automatically retransmitted during network idle periods. The system implements an adaptive flow control algorithm, automatically reducing the transmission priority of non-critical data when network congestion is detected, and activating lossy data compression algorithms when necessary. All data transmissions are timestamped and quality flags are included, and the receiving end implements data reconstruction and out-of-order processing mechanisms.

[0065] A comprehensive communication monitoring system is constructed to monitor the operational status of both power line carrier channels in real time. Monitoring indicators include: spectrum utilization, signal-to-noise ratio, and bit error rate of the power line carrier channel; signal strength, network attachment status, and traffic usage of the 4G channel; and service quality indicators such as end-to-end transmission delay and packet loss rate. This data is transmitted back to the network management center via a dedicated monitoring channel with a sampling period of 5 seconds. The system implements intelligent diagnostic functions, automatically identifying common fault modes such as power line noise interference and 4G base station handover interruptions, and executing preset recovery strategies. The network management interface provides rich visualization tools, including spectrum analysis graphs, network topology diagrams, and traffic heatmaps, supporting historical data review and comparative analysis. When communication interruptions exceed a threshold, the system automatically activates an emergency communication mode, transmitting the most critical equipment status information via BeiDou short message service.

[0066] To ensure communication reliability under extreme conditions, the system is designed with a multi-layered disaster recovery and backup scheme. At the channel level, in addition to the primary power line carrier and 4G channels, an RS485 wired interface is reserved as a last resort channel. For data caching, the terminal equipment is equipped with 8GB of industrial-grade eMMC memory, capable of caching 7 days of complete monitoring data. The communication protocol implements a breakpoint resume mechanism, automatically synchronizing data status after network recovery. The system supports a "black start" mode, maintaining critical communication functions through backup power in the event of a complete power outage. Regular communication drills are conducted to simulate various fault scenarios and verify the system's self-recovery capabilities. All key components adopt an A / B dual-system design, supporting hot switching without interrupting services. The communication module's MTBF design index exceeds 100,000 hours, meeting the high reliability requirements of the power system for communication equipment.

[0067] In this embodiment of the invention, the energy efficiency analysis and optimization module is used to dynamically analyze the energy consumption of the power distribution network using a load forecasting model, and to construct a set of energy efficiency optimization schemes, including: The energy efficiency analysis and optimization module employs a multi-layered LSTM neural network architecture to construct a load forecasting model. The model input layer is designed as a multimodal data interface, receiving three key inputs: historical load data, meteorological data, and calendar features. During data preprocessing, a sliding window technique is used to generate training samples, with a window length of 30 days and a sliding step size of 1 hour. The input data is standardized: load data is normalized using Z-score, meteorological data using Min-Max normalization, and calendar features using one-hot encoding. To address the issue of missing data, the system implements an attention-based interpolation algorithm, which significantly outperforms traditional linear interpolation methods. The preprocessed data stream is divided into training, validation, and test sets to ensure the objectivity of model evaluation.

[0068] The core prediction model employs a stacked LSTM structure, containing three hidden layers and one fully connected output layer. An attention mechanism is introduced to dynamically adjust the weights at different time steps, focusing on the impact of recent data and special dates. The training process uses an Adam optimizer with gradient pruning, an initial learning rate of 0.001, and a ReduceLROnPlateau strategy for dynamic adjustment. The Huber loss function is used to balance MSE and MAE, improving the model's robustness to outliers. Dropout and L2 regularization are implemented to prevent overfitting. The model is trained on an NVIDIA Tesla V100 GPU with a batch size of 128, and an early stopping mechanism controls the number of training epochs. The final model achieves an average prediction accuracy of 95.2% on the test set, with peak load prediction error controlled within ±3%.

[0069] The system establishes a transformer load-line loss analysis model, constructs transformer loss characteristic curves based on the IEEE C57.91 standard, and considers different loss components such as copper loss and iron loss. The nonlinear relationship between load rate and line loss rate is quantified through Pearson correlation coefficient and mutual information analysis. The energy efficiency optimization engine adopts a constraint satisfaction problem-solving framework, aiming at optimal overall system energy efficiency, and considers various constraints: node voltage limits, branch capacity limitations, and equipment operation frequency limits. The optimization algorithm output includes three types of strategies: capacitor switching schemes, distributed power supply scheduling plans, and transformer economic operation suggestions. Each strategy is accompanied by a detailed benefit assessment, including indicators such as expected energy savings, loss reduction ratio, and equipment life extension estimates.

[0070] The generated optimization schemes must be verified through a digital twin system before execution. The verification process employs real-time simulation technology, building a full power grid model on the OPAL-RT simulation platform to simulate system state changes after scheme implementation. Verification indicators include 10 key technical parameters such as voltage qualification rate, equipment load rate, and harmonic distortion rate. Verified schemes are converted into standardized knowledge entries, containing structured information such as applicable scenario descriptions, detailed scheme content, expected effects, and precautions. The knowledge base is updated using a version control mechanism, generating a new knowledge version with each update and retaining historical versions for retrospective querying. The system implements a knowledge association network, automatically establishing semantic connections between different knowledge entries and supporting multi-dimensional knowledge retrieval. The knowledge quality assessment module periodically checks the actual application effect of each entry, marking knowledge entries with declining effectiveness as requiring optimization.

[0071] The energy efficiency optimization system employs a closed-loop learning mechanism to continuously improve prediction and optimization capabilities. Deviations between actual operational data and prediction results are automatically collected and used for incremental model training. The system implements two learning modes: online learning for real-time fine-tuning of model parameters, and batch learning for monthly retraining of the entire model. A knowledge base maintains a "case-effect" mapping table, statistically analyzing the historical execution effects of various solutions for weight allocation during strategy generation. The system also constructs a virtual optimization environment to explore novel optimization strategies through reinforcement learning algorithms, expanding the knowledge base's coverage. Detailed experimental logs are recorded throughout the learning process, including metadata such as data version, model parameters, and training metrics, ensuring traceability. The system generates monthly performance reports, analyzing KPIs such as prediction accuracy, solution adoption rate, and actual energy-saving effects to guide subsequent optimization directions.

[0072] To assist operators in understanding and utilizing optimization results, the system features an advanced visual analysis interface. A 3D situational awareness view displays the overall network energy efficiency, using color gradients to indicate the energy efficiency levels of different regions. The optimization scheme comparison function supports side-by-side display of the technical and economic indicators of multiple alternative schemes. The decision-making simulation tool can model the medium- and long-term impacts of different choices, including predicted curves for energy consumption changes and equipment aging rates. The system provides a natural language query interface, allowing operators to ask questions via voice or text, and the system automatically retrieves relevant knowledge and generates answers. Important operations are equipped with a dual confirmation mechanism; critical equipment switching requires manual review before execution. All decision-making processes generate complete audit logs, recording information such as operators, decision basis, and execution results, meeting the compliance requirements of power system operation and management.

[0073] In this embodiment of the invention, the self-learning system upgrade module is used to update the system algorithm model online using an incremental learning mechanism to form a continuously evolving diagnostic system, including: The self-learning system upgrade module adopts a distributed microservice architecture, constructing a complete closed-loop system for model iteration. The data acquisition service subscribes to the output of the fault diagnosis module in real time via a Kafka message queue, processing up to 5000 fault sample data points per second. The system employs a three-level data caching structure: the first level uses Redis to store the raw samples collected in real time, retaining data from the last 72 hours; the second level uses Elasticsearch to build an indexed sample library, supporting complex condition queries; and the third level archives data to the HDFS distributed file system, forming a permanent knowledge reserve. Each fault sample contains 42 dimensions of structured data, including basic electrical characteristics, as well as metadata such as equipment operating conditions, environmental parameters, and handling effects. The system implements an automated data annotation process, combining an expert rule engine and clustering algorithms to classify the quality of new samples, with only A-level quality samples entering the training set. The data version control system fully records the composition of the dataset used for each incremental learning iteration, ensuring the reproducibility of experiments.

[0074] The core algorithm employs an improved elastic weight solidification method to prevent model performance degradation on three levels: First, it calculates the importance scores of parameters in each layer using the Fisher information matrix, establishing a parameter protection priority queue; second, it designs dynamic regularization coefficients to adaptively adjust constraint strength based on parameter sensitivity; and finally, it introduces a knowledge distillation mechanism, using a teacher-student model structure to retain historical knowledge. The training process uses mixed-precision computation, employing FP32 precision for critical parts to maintain stability and FP16 for general computations to improve efficiency. The optimizer is AdamW, combined with a periodic learning rate hot-restart strategy, achieving a balance between exploration and convergence. The system implements gradient pruning and weight normalization techniques to avoid gradient explosion in incremental learning. A specially designed parameter isolation region manages network weights corresponding to different types of faults in separate areas, automatically expanding the network structure without affecting existing regions when a new fault category is added.

[0075] Updating the model requires passing four levels of verification before deployment: unit testing verifies that the identification accuracy of each fault category is no lower than historical levels; integration testing checks the overall performance of the model in complex scenarios; regression testing ensures that new knowledge does not undermine existing capabilities; and stress testing evaluates the model's robustness under extreme conditions. The verification platform uses a combination of real data and digital twins to build a verification set containing over 100,000 test cases. The system implements an automated testing pipeline, triggering the execution of complete test suites via Jenkins and generating a quality report containing 58 indicators. Safety assessments pay particular attention to model interpretability, using SHAP values ​​to analyze the contribution of each feature and ensure that decision-making aligns with the physical laws of the power system. To address model bias issues, a fairness testing framework is established to check the consistency of identification across different equipment types and voltage levels. All test results are stored in a blockchain-based notarization system, providing an immutable audit trail.

[0076] The system implements a zero-downtime hot deployment mechanism and smoothly switches model versions based on Kubernetes' blue-green deployment strategy. The new model initially runs in shadow mode, processing requests in parallel with the production model but without producing results, continuously monitoring its performance metrics. After a 24-hour observation period, 5%, 20%, and 50% of traffic are gradually switched to the new model, eventually completing a full switch. The version management system maintains a complete model family, with each version containing metadata such as training data, hyperparameters, and performance metrics, supporting quick rollback to any historical version. The model service encapsulation uses the TensorFlow Serving framework, supporting parallel loading of multiple models and automatic version routing. The performance monitoring system tracks runtime metrics such as model inference latency and memory usage in real time, automatically triggering alerts when anomalies occur. The system automatically archives old model versions monthly, releasing storage resources while retaining snapshots of critical versions.

[0077] The system employs a triple feedback loop to drive continuous optimization: a short-term loop analyzes the latest fault samples every 15 minutes and fine-tunes model parameters; a medium-term loop evaluates model performance trends daily and adjusts feature weights; and a long-term loop performs architecture-level optimization monthly, expanding network capacity as needed. Knowledge evolution utilizes a Darwinian selection mechanism, periodically evaluating the effectiveness of each knowledge unit, eliminating outdated models, and reinforcing high-frequency scenarios. The system implements active learning capabilities, identifying samples with high prediction uncertainty and proactively requesting expert annotation. An innovative federated learning architecture is introduced, aggregating local model updates from multiple substations to form a global knowledge base without sharing original data. All optimization processes record complete experimental logs, including metadata such as data version, hyperparameters, and evaluation metrics, supporting complete experimental reproducibility and effect attribution analysis.

[0078] A comprehensive model operation and maintenance monitoring system is built, tracking over 20 key indicators in real time, including accuracy and recall rates for each fault category, model inference latency, and resource utilization. Monitoring data is stored in a time-series database, supporting millisecond-level queries within 30 days. A visualization analysis platform provides multi-dimensional insights: a model performance dashboard displays the overall health status, a feature importance heatmap reveals decision-making basis, and error case analysis tools assist in locating model defects. The system implements automated root cause analysis, automatically associating recent model changes, data distribution changes, and other potential factors when performance degradation is detected. Operation and maintenance personnel can query system status via natural language, and the system automatically generates analysis reports. All monitoring data regularly generates operation and maintenance reports to guide continuous system improvement.

[0079] Reference Figure 2The diagram shown is a flowchart illustrating a method for monitoring the status and providing early warning of faults in an intelligent power supply system according to an embodiment of the present invention. In this embodiment, the method includes: The power supply network electrical parameters are synchronously collected based on multi-sensor fusion technology to generate a real-time power data set. Wavelet transform algorithm is used to perform feature extraction on the original power data to form a multidimensional state feature vector; Deep learning models are used to perform pattern recognition on the operating status of power distribution networks and output fault type and location information. By combining historical operational data, the warning threshold parameters are dynamically optimized to generate an adaptive threshold rule base. Based on an expert knowledge base, decision analysis is performed on the fault diagnosis results to generate a set of control strategy instructions; The monitoring data is remotely transmitted via a dual-channel communication protocol, and a real-time cloud monitoring interface is established. Dynamic analysis of power distribution network energy consumption is performed using load forecasting models to construct a set of energy efficiency optimization schemes; An incremental learning mechanism is used to update the system algorithm model online, forming a continuously evolving diagnostic system.

[0080] This invention enables comprehensive monitoring of the power supply network through a heterogeneous sensor array, deploying a data acquisition network composed of high-precision voltage transformers, electronic current sensors, and distributed temperature sensors. The IEEE 1588 precision time protocol is used to achieve microsecond-level synchronous sampling among multiple nodes, with the sampling frequency configurable from 1kHz to 10kHz according to monitoring requirements. The acquired analog signals are converted by a 24-bit Σ-Δ ADC and then encapsulated in SV messages according to the IEC 61850-9-2 standard to form a timestamped standardized dataset. The system applies the db4 wavelet basis function to perform 5-level wavelet packet decomposition on the electrical signals, calculating the energy entropy of each frequency band as steady-state features, extracting d4-d5 level detail coefficients to represent transient features, and finally fusing them to form a 20-dimensional feature vector containing time-domain and frequency-domain features. The feature extraction process is implemented on a dedicated DSP chip to ensure real-time requirements.

[0081] The constructed hybrid neural network model comprises three convolutional layers and two LSTM layers. First, spatial patterns of the feature matrix are extracted using convolutional kernels. Then, LSTM units are used to analyze temporal correlations. Finally, a softmax classifier outputs the fault probability distribution. The model training employs a focal loss function to address the sample imbalance problem, achieving a recognition accuracy of 95.2% on the test set. The early warning system establishes a dynamic threshold adjustment mechanism. Based on sliding window analysis of historical data, and considering environmental factors such as temperature and load rate, a fuzzy logic algorithm is used to calculate the early warning thresholds for each monitored quantity in real time. When a new fault mode is detected, the system automatically initiates a threshold relaxation mechanism, gradually learning the optimal early warning boundary. All threshold parameters are updated to the rule base hourly, and their integrity is ensured through digital signatures.

[0082] The expert knowledge base stores over 2000 typical fault cases, employing case-based reasoning technology to match the current scenario and evaluate the impact of various handling solutions on power supply reliability, equipment safety margin, and five other dimensions. The decision engine generates an optimized control instruction set including circuit breaker action sequences and load transfer paths, and verifies the feasibility of the solutions through a digital twin system. The communication system features a dual-channel architecture of power line carrier and 4G, ensuring critical commands are delivered within 20ms via the dedicated power network, while monitoring data is uploaded in real-time through a cloud interface. The transmission process implements SM4 encryption and CRC checksum, meeting the Level 3 requirements of the "Regulations for Security Protection of Power Monitoring Systems".

[0083] The load forecasting model employs an attention-enhanced LSTM network, taking historical load data, meteorological data, and calendar features as input, and outputting load curves for the next 24 hours. The energy efficiency analysis engine establishes a transformer load-line loss correlation model, automatically generating capacitor switching schemes and distributed power dispatch plans, expected to reduce line losses by 5%-8%. The incremental learning system collects new fault samples weekly, updating model parameters using the EWC algorithm while retaining important weights to prevent forgetting. The updated model is validated using a test suite containing over 100,000 cases, and finally seamlessly transitioned to the production environment using a blue-green deployment strategy.

[0084] Each module interacts with the other via an Enterprise Service Bus (ESB), employing a unified OPC UA interface specification. The system achieves a complete closed loop from data acquisition, feature analysis, fault diagnosis to decision-making, with average end-to-end latency controlled within 500ms. The operations and maintenance monitoring platform tracks over 20 key indicators in real time, including fault identification rate, early warning accuracy, and communication latency, automatically triggering alarms when anomalies occur. A monthly system health report is generated, analyzing performance trends of each module and guiding optimization efforts.

[0085] Comparison chart of intelligent power supply system status detection and fault early warning system with traditional solutions Comparison Dimensions Traditional solution This invention Advantages Data collection Single sensor, asynchronous sampling Heterogeneous sensor array, IEEE 1588 time synchronization Improve data accuracy and eliminate time domain bias Feature extraction Fourier transform, fixed frequency band analysis Wavelet packet decomposition (db4-db8), joint time-frequency analysis More accurate capture of transient / steady-state features Fault Diagnosis Threshold judgment + simple machine learning Hybrid CNN-LSTM-Attention Mechanism Model The recognition accuracy rate has been improved to 95.2%. Early warning mechanism Fixed threshold, high false alarm rate Dynamic threshold optimization (Gaussian mixture model + working condition compensation) False alarm rate reduced by more than 80% Decision support Human experience judgment Expert knowledge base + digital twin verification The speed of generating disposal plans has increased tenfold. Communication reliability Single-channel transmission, prone to packet loss Dual-channel (GOOSE + 4G), SM9 encryption Critical instruction transmission latency <20ms System self-evolution Static model, requires manual upgrade. Incremental learning (EWC++ algorithm) The adaptation time for new faults has been shortened to 24 hours. Energy efficiency optimization Fixed strategy, no real-time adjustments LSTM load forecasting + dynamic scheduling Line loss rate reduced by 5%-8% This invention constructs a complete intelligent monitoring and early warning system through technologies such as high-precision synchronous acquisition by heterogeneous sensors, feature extraction through wavelet-deep learning fusion, dynamic threshold optimization algorithms, and dual-channel secure communication. Compared with traditional solutions, it achieves breakthrough improvements in fault identification accuracy, false alarm rate, and response speed. It also possesses self-learning and continuous optimization capabilities, significantly improving the reliability and operational efficiency of the power grid, and providing safer and more accurate technical support for smart grid construction.

[0086] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0087] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart power supply system status monitoring and fault early warning system, characterized in that, The system includes the following modules: The power supply network data acquisition module is used to synchronously acquire electrical parameters of the power supply network based on multi-sensor fusion technology and generate a real-time power data set. The state feature extraction and analysis module is used to perform feature extraction on the raw power data using the wavelet transform algorithm to form a multidimensional state feature vector. The fault diagnosis and location module is used to perform pattern recognition on the operating status of the power distribution network using a deep learning model, and output fault type and location information. The early warning threshold dynamic adjustment module is used to dynamically optimize the early warning threshold parameters by combining historical operation data and generate an adaptive threshold rule base. The intelligent decision-making and control module is used to perform decision analysis on fault diagnosis results based on an expert knowledge base and generate a set of control strategy instructions. The communication and remote monitoring module is used to remotely transmit monitoring data through a dual-channel communication protocol and establish a real-time monitoring interface in the cloud. The energy efficiency analysis and optimization module is used to perform dynamic analysis of the energy consumption of the power distribution network using the load forecasting model, and to build a set of energy efficiency optimization schemes. The self-learning system upgrade module is used to update the system algorithm model online using an incremental learning mechanism, forming a continuously evolving diagnostic system.

2. The system according to claim 1, characterized in that, When the power supply network data acquisition module synchronously acquires electrical parameters of the power supply network based on multi-sensor fusion technology, it includes: The IEEE 1588v2 precision clock protocol is used to achieve microsecond-level time synchronization of the real-time power data set; Configure an isolation sampling circuit to resist electromagnetic interference and ensure the accuracy of electrical parameter acquisition in a strong electric field environment; An adaptive filtering algorithm is deployed to eliminate high-frequency noise interference in the real-time power data set.

3. The system according to claim 1, characterized in that, When the state feature extraction and analysis module uses the wavelet transform algorithm to perform feature extraction on the original power data, it includes: A 6-layer decomposition tree based on the Mallat algorithm was constructed to process the time-frequency characteristics of the original power data; The frequency domain stability index obtained by singular value decomposition is fused into the multidimensional state feature vector; Configure a dynamic threshold mechanism to automatically remove abnormal feature components from the multidimensional state feature vector.

4. The system according to claim 1, characterized in that, When the fault diagnosis and location module uses a deep learning model to perform pattern recognition on the operating status of the power distribution network, it includes: A residual attention mechanism is employed to enhance the deep learning model's ability to capture transient features; Construct a topology constraint loss function to ensure that the fault location information conforms to the physical connection relationship of the power grid; Deploy the model interpretation module to visualize and output a heatmap of the decision-making basis for the fault type.

5. The system according to claim 1, characterized in that, When the early warning threshold dynamic adjustment module dynamically optimizes the early warning threshold parameters by combining historical operating data, it includes: The dynamic distribution characteristics of the historical operating data are extracted using a sliding time window statistical method. A threshold optimizer based on the Q-learning algorithm is constructed to generate the adaptive threshold rule base; Configure a robust outlier handling mechanism to prevent erroneous adjustment of the warning threshold parameter.

6. The system according to claim 1, characterized in that, When the intelligent decision-making and control module performs decision analysis on the fault diagnosis results based on the expert knowledge base, it includes: A case-based reasoning engine is used to match historical handling solutions in the expert knowledge base. A multi-objective optimization model is constructed to generate the Pareto front solution of the control strategy instruction set; Deploy an instruction verification module to verify the compatibility of the control strategy instruction set with the current system state.

7. The system according to claim 1, characterized in that, When the communication and remote monitoring module remotely transmits monitoring data via a dual-channel communication protocol, it includes: The monitoring data transmission security is ensured by using IPSec VPN and quantum encryption dual-channel. Configure an adaptive bitrate adjustment algorithm to optimize the bandwidth utilization of the cloud real-time monitoring interface; Deploy a data priority scheduler to ensure that critical alarm data is transmitted with priority in the dual-channel communication protocol.

8. The system according to claim 1, characterized in that, When the energy efficiency analysis and optimization module uses the load forecasting model to perform dynamic analysis of the energy consumption of the distribution network, it includes: A hybrid prediction model integrating LSTM and XGBoost is constructed to process the spatiotemporal characteristics of the power distribution network energy consumption; Sensitivity analysis was used to identify the key control parameters in the energy efficiency optimization scheme set; A rolling optimization mechanism is deployed to update the recommended weights of the energy efficiency optimization scheme set every 15 minutes.

9. The system according to claim 1, characterized in that, When the self-learning system upgrade module uses an incremental learning mechanism to update the system algorithm model online, it includes: Configure an elastic weight solidification algorithm to prevent catastrophic forgetting in the incremental learning mechanism; Deploying the model performance drift detector triggers a version rollback of the continuous evolution diagnostic system. Construct a knowledge distillation pipeline to migrate features from the historical model to the new version of the system algorithm model.

10. A method for status monitoring and fault early warning of an intelligent power supply system, characterized in that, The method includes: S1: Based on multi-sensor fusion technology, the electrical parameters of the power supply network are synchronously collected to generate a real-time power data set; S2: Wavelet transform algorithm is used to extract features from the original power data to form a multidimensional state feature vector; S3: Use deep learning models to perform pattern recognition on the operating status of the power distribution network and output fault type and location information; S4: Dynamically optimize the warning threshold parameters by combining historical operation data to generate an adaptive threshold rule base; S5: Based on the expert knowledge base, perform decision analysis on the fault diagnosis results and generate a set of control strategy instructions; S6: Remotely transmits monitoring data via dual-channel communication protocol and establishes a cloud-based real-time monitoring interface; S7: Utilize load forecasting models to perform dynamic analysis of energy consumption in power distribution networks and construct a set of energy efficiency optimization schemes; S8: An incremental learning mechanism is used to update the system algorithm model online, forming a continuously evolving diagnostic system.

Images