Auditing data early warning method and system

Abstract

The invention belongs to the technical field of power systems, and particularly relates to an audit data early warning method and system, and the method comprises the steps: collecting the node voltage and current of a power supply network and the charging and discharging efficiency data of an energy storage system in real time through a distributed sensor network, and generating an original data flow; carrying out localized cleaning and standardized format conversion through an edge computing node; generating a dynamic risk assessment result and a scheduling scheme based on fuzzy logic reasoning and digital twinborn simulation; and analyzing and predicting deviation through meta-learning, and realizing closed-loop optimization by adopting a model distillation technology. The system comprises a distributed sensor network, an edge computing node, a multi-source data fusion module and the like. According to the method, the anomaly detection real-time performance, the risk assessment accuracy and the energy storage scheduling economy of the power system are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of circuit devices or systems for power supply or distribution; power energy storage systems, and particularly relates to the intelligent data monitoring and abnormal early warning technology of power systems, and in particular to an audit data early warning method and system based on multi-source data fusion and intelligent analysis. Background Art

[0002] With the expansion of the scale of power systems and the complexity of energy structures, the demand for data monitoring and abnormal early warning of power networks is increasing day by day. In traditional power systems, power energy storage, power distribution control, and power supply stability management highly rely on manual inspections or simple alarm mechanisms based on fixed thresholds, and it is difficult to achieve real-time and dynamic abnormal detection and risk prediction. For example, in the prior art, there is a lack of active analysis capabilities for power energy losses, equipment operating states, or abnormal fluctuations in energy storage systems, resulting in a lag in response to potential faults and affecting power supply reliability.

[0003] In addition, the data monitoring of existing power systems mostly focuses on physical layer parameters (such as voltage and current), and lacks the ability of in-depth auditing and correlation analysis of operation data. For example, how to construct a dynamic early warning model through the fusion processing of multi-source data (such as load curves, energy storage efficiency, power distribution node states) and realize the intelligent identification of abnormal behaviors in power systems is still a technical problem to be solved urgently.

[0004] Based on this, there is an urgent need for an auditing early warning method and system that combines the characteristics of power systems and data intelligent analysis technology to improve the efficiency of abnormal detection in power supply, power distribution, and energy storage scenarios, and optimize the initiative and security of power energy management. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides an audit data early warning method and system to solve the problems of the deficiencies of the prior art in the intelligent operation and maintenance of power systems.

[0006] To solve the above technical problems, the specific technical solutions of the present invention are as follows: In the first aspect, the audit data early warning method provided by the present invention includes: Step S101, real-time collect the node voltage, current, load fluctuation data of the power supply network, and the charge and discharge efficiency, capacity attenuation rate data of the power energy storage system through a distributed sensor network to generate an original data stream; Step S102, input the original data stream into an edge computing node for local cleaning and standardized format conversion to generate standardized intermediate data; Step S103: Based on the federated learning framework and knowledge graph technology, perform cross-domain data fusion on the standardized intermediate data, extract the hidden features of the distribution network topology relationship and the dynamic correlation features of the charging and discharging behavior and load fluctuation, and generate an associated feature vector. Step S104: Use self-supervised learning to perform dynamic feature engineering processing on the associated feature vector, generate multi-dimensional power system operation state features, and pre-screen abnormal patterns for the high-dimensional features through a lightweight anomaly detection model, and output a high-risk event pool. Step S105: Based on the high-risk event pool, construct a hierarchical early warning model, use a spatio-temporal fusion Transformer architecture to perform spatio-temporal feature modeling on the power supply network time series data to predict the device-level failure probability, and optimize the distribution network load distribution and energy storage scheduling strategy through a hybrid reinforcement learning framework to predict the system-level risk. Step S106: According to the device-level failure probability and system-level risk prediction results, generate a dynamic risk assessment level by combining fuzzy logic reasoning and evidence theory, and trigger real-time alarms or generate energy storage scheduling suggestions and load transfer schemes. Step S107: Collect the system response data after the early warning strategy is executed, analyze the model prediction deviation through a meta-learning framework, and dynamically adjust the feature weights and model structure to form a closed-loop optimization link. Among them, the original data flows through the edge computing node for processing and then flows to the data fusion module. The associated feature vector is input into the dynamic feature engineering module to generate a pre-screening result. The high-risk event pool drives the training and decision-making of the hierarchical early warning model, and the risk assessment result is fed back to the model optimization module in reverse.

[0007] Furthermore, for the audit data early warning method of the present invention, step S103 includes: Extract the node vulnerability features in the distribution network topology relationship through a graph embedding algorithm, and capture the non-linear correlation between the charging and discharging efficiency and load fluctuation based on the spatio-temporal attention mechanism. Align the SCADA system structured data and the smart meter unstructured data through federated learning to generate a cross-domain associated power operation knowledge graph. The hidden features and dynamic correlation features are used as inputs and flow to the dynamic feature engineering module, forming a data closed-loop for abnormal pattern pre-screening with the output of the anomaly detection model.

[0008] Furthermore, for the audit data early warning method of the present invention, step S104 includes: Extract harmonic distortion features from non-stationary time series data using wavelet transform, and perform data augmentation on abnormal samples through a generative adversarial network. Input the enhanced features into the improved Isolation Forest algorithm for fast anomaly detection, and store the screened high-risk events in the event pool; The event pool flows as input to the hierarchical early warning model, and forms a collaborative analysis of equipment-level and system-level risk assessments with the spatio-temporal feature modeling of the spatio-temporal fusion Transformer.

[0009] Furthermore, for the audit data early warning method of the present invention, the step S105 includes: Introduce an interpretability module in the primary model, locate the contribution degree of abnormal features through SHAP values, and assist in generating equipment-level alarm instructions; Embed a transfer learning module in the secondary model, and use historical power grid data to pre-train the policy network to accelerate the convergence of the load distribution strategy in the new scenario; The output of the primary model and the policy optimization result of the secondary model are jointly input into the risk assessment module to achieve the linkage of short-term fault response and medium- and long-term risk prevention and control.

[0010] Furthermore, for the audit data early warning method of the present invention, the step S106 includes: Adopt the sliding window statistical method to adjust the early warning threshold according to the real-time load status, and combine the digital twin platform to simulate the execution effect of the energy storage scheduling strategy; Balance the power supply reliability and economy through the multi-objective optimization algorithm, generate a load transfer plan and push it to the operation and maintenance terminal; The simulation result and the optimization strategy flow as input to the early warning execution module, and form an anti-tampering instruction transmission link with the blockchain traceability mechanism.

[0011] Furthermore, for the audit data early warning method of the present invention, the step S107 includes: Constrain the model weight update in the incremental learning process through the elastic weight consolidation algorithm to prevent the forgetting of historical knowledge; Use model distillation technology to transfer the knowledge of the complex hybrid model to the lightweight network to adapt to the real-time inference requirements of edge devices; The model weight update and the knowledge transfer result are reversely input into the feature engineering module to drive the iterative optimization of dynamic feature generation and anomaly detection.

[0012] Furthermore, for the audit data early warning method of the present invention, the alignment of the SCADA system structured data and the smart meter unstructured data through federated learning to generate a cross-domain associated power operation knowledge graph includes: Adopt differential privacy technology to protect the data security of each power subsystem during the data fusion process, and achieve cross-domain feature sharing through gradient encrypted transmission; The topological relationship features of the knowledge graph and the data aligned with federated learning are jointly input into the spatio-temporal attention mechanism to enhance the robustness of the correlation analysis between charge and discharge behaviors and load fluctuations.

[0013] Furthermore, for the audit data warning method of the present invention, the data enhancement of abnormal samples through the generative adversarial network includes a generator and a discriminator. The generator is used to inject adversarial perturbations simulating sensor noise into the input data, and the discriminator focuses on the classification training of high-risk samples through the Focal Loss function; Inject adversarial perturbations simulating sensor noise into the input data, and focus on the classification training of high-risk samples through the Focal Loss function; The feature vector after the adversarial training and the output of the lightweight anomaly detection model form a high-precision pre-screening result, reducing the false alarm rate of the hierarchical warning model.

[0014] Furthermore, for the audit data warning method of the present invention, embedding a transfer learning module in the secondary model and using the historical power grid data to pre-train the policy network to accelerate the convergence of the load distribution strategy in the new scenario includes: Initialize the parameters of the reinforcement learning network using the pre-trained load distribution strategy of the distribution network, and dynamically adjust the energy storage scheduling actions through the proximal policy optimization algorithm; The collaborative output of the transfer learning and the reinforcement learning drives the policy simulation of the digital twin platform, forming a closed-loop verification link from model training to actual execution.

[0015] In a second aspect, the audit data warning system provided by the present invention is applied to the audit data warning method and includes: A distributed sensor network configured to collect in real time the node voltage, current of the power supply network, and the charge and discharge efficiency data of the energy storage system; Edge computing nodes connected to the sensor network, configured to perform local cleaning and standardization processing on the original data to generate standardized intermediate data; A multi-source data fusion module connected to the edge computing nodes, configured to extract cross-domain correlation features through federated learning and knowledge graph technology; A dynamic feature engineering module connected to the multi-source data fusion module, configured to generate multi-dimensional operating state features and perform pre-screening of abnormal patterns; A hierarchical warning model construction module connected to the dynamic feature engineering module, configured to train a spatio-temporal fusion Transformer and a hybrid reinforcement learning framework; A risk assessment and decision-making module connected to the hierarchical warning model construction module, configured to generate a dynamic risk level and an energy storage scheduling strategy; A closed-loop optimization module, connected to the risk assessment and decision-making module, is configured to achieve adaptive update of the early warning model through meta-learning and model distillation techniques; Among them, the data of the sensor network flows to the edge computing node, the standardized intermediate data is input into the dynamic feature engineering module after being processed by the multi-source fusion module, the output of the hierarchical early warning model drives risk assessment and decision-making, and the decision result is fed back to the closed-loop optimization module in reverse.

[0016] Advantages of the present invention; The audit data early warning method and system provided by the present invention collect multi-source heterogeneous data of the power supply network and energy storage system in real time through a distributed sensor network, and perform local cleaning and standardization processing in combination with edge computing nodes, effectively reducing data transmission delay and improving data quality; realize cross-domain data fusion based on the federated learning framework and knowledge graph technology, extract hidden features of the distribution network topology relationship and dynamic correlation features of charge and discharge behavior and load fluctuation on the premise of protecting the data privacy of each subsystem, and enhance the semantic association analysis ability of multi-modal data; use self-supervised learning to perform dynamic feature engineering processing on the association features, extract harmonic distortion features of non-stationary time-series data through wavelet transform and combine with the generative adversarial network to enhance the abnormal sample distribution, and improve the recognition sensitivity of the lightweight anomaly detection model to hidden faults; adopt a spatio-temporal fusion Transformer architecture to model the spatio-temporal dependence of the power supply network time-series data, accurately locate the device-level fault probability, and at the same time optimize the load distribution and energy storage scheduling strategy through a hybrid reinforcement learning framework, and simulate the strategy execution effect in combination with the digital twin platform to achieve the collaborative optimization of device-level short-term response and system-level medium- and long-term risk prevention and control; build a closed-loop optimization link based on the meta-learning framework and elastic weight consolidation algorithm, dynamically adjust the feature weights and model structure, and adapt to the real-time inference requirements of edge devices in combination with model distillation technology, forming an iterative mechanism for the early warning model to adapt to the dynamic operation environment of the power system, so as to improve the real-time performance of power supply network anomaly detection, the accuracy of risk assessment and the economy of energy storage scheduling, meeting the technical requirements of intelligent monitoring and energy storage management of power systems under the H02J classification. Description of the drawings

[0017] In order to more clearly illustrate the technical solutions of the present invention, the drawings required to be used in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, other drawings can be obtained according to the drawings without creative efforts.

[0018] Figure 1 It is a flowchart of the audit data early warning method provided by the embodiment of the present invention. Detailed implementation manners

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention. The following will describe in detail the technical solutions provided by each embodiment of the present invention with reference to the drawings. To better understand the objectives of the present invention, the present invention will be further described in detail below.

[0020] Please refer to Figure 1 , in a first aspect, the present invention provides an audit data warning method, including: Step S101: Real-time collect the node voltage, current, load fluctuation data of the power supply network, as well as the charge and discharge efficiency and capacity attenuation rate data of the energy storage system through a distributed sensor network to generate an original data stream; Step S102: Input the original data stream into an edge computing node for local cleaning and standardized format conversion to generate standardized intermediate data; Step S103: Based on a hierarchical federated learning framework and knowledge graph technology, perform cross-domain data fusion on the standardized intermediate data, extract the node vulnerability features of the distribution network topology relationship through a graph embedding algorithm, and capture the non-linear correlation between the charge and discharge efficiency and load fluctuation based on a spatio-temporal attention mechanism to generate an associated feature vector; Step S104: Use self-supervised learning to perform dynamic feature engineering processing on the associated feature vector to generate multi-dimensional power system operation state features, and pre-screen abnormal patterns for high-dimensional features through a lightweight anomaly detection model to output a high-risk event pool; Step S105: Build a hierarchical warning model based on the high-risk event pool, use a spatio-temporal fusion Transformer architecture to perform spatio-temporal feature modeling on the time-series data of the power supply network to predict the device-level failure probability, and optimize the distribution network load distribution and energy storage scheduling strategy through a hybrid reinforcement learning framework to predict the system-level risk; Step S106: According to the device-level failure probability and system-level risk prediction results, generate a dynamic risk assessment level by combining fuzzy logic reasoning and evidence theory, and trigger real-time alarms or generate energy storage scheduling suggestions and load transfer plans; Step S107: Collect the system response data after the warning strategy is executed, analyze the model prediction deviation through a meta-learning framework, and dynamically adjust the feature weights and model structure to form a closed-loop optimization link; Among them, the original data flows to the data fusion module after being processed by the edge computing node, the associated feature vector is input into the dynamic feature engineering module to generate a pre-screening result, the high-risk event pool drives the training and decision-making of the hierarchical early warning model, and the risk assessment result is fed back to the model optimization module in reverse.

[0021] In step S101, the distributed sensor network is deployed at the key distribution nodes of the power supply network and the battery cluster monitoring points of the energy storage system, specifically including high-precision voltage transformers, Hall current sensors, and battery health status monitoring units. The edge computing node uses a sliding window filtering algorithm to eliminate high-frequency noise interference from the collected original voltage and current signals, and compensates for the missing time-series data caused by communication delays through linear interpolation. For the capacity attenuation rate data of the energy storage system, a state estimation algorithm based on Kalman filtering is used to correct the sensor drift error and generate an original data stream with time synchronization marks. The data encapsulation module converts multi-source heterogeneous data into a standardized JSON format, including device identifiers, normalized values, and data quality identification fields, providing a structured input for subsequent processing.

[0022] The localization cleaning process in step S102 adopts a hierarchical processing strategy: performing outlier detection and replacement based on the interquartile range for continuous data such as voltage and current; removing redundant nodes and reconstructing the adjacency matrix for the distribution network topology status data; applying a moving average smoothing process to the charge and discharge efficiency data of the energy storage system. The standardization module uniformly converts data with different dimensions to the [-1, 1] interval, where the time-series data is aligned with the sampling frequency through the dynamic time warping algorithm, and the unstructured data is converted into a graph structure representation. The edge node pushes the standardized intermediate data to the cloud message queue through the MQTT protocol and attaches a data traceability tracking label to support full-process audit and traceability.

[0023] The cross-domain data fusion in step S103 adopts a hierarchical federated learning framework to integrate the real-time monitoring data of the SCADA system and the user load curve of the smart meter while protecting data privacy. The graph embedding module uses the GraphSAGE algorithm to perform unsupervised learning on the distribution network topology and extracts low-dimensional vectors representing the electrical characteristics of nodes and the regional power supply relevance. The spatio-temporal attention mechanism calculates the dynamic correlation matrix between the charge and discharge power sequence of the energy storage system and the regional load curve to identify potential overload risk propagation paths. The knowledge graph construction module fuses the topological feature vectors, spatio-temporal correlation matrices, and device metadata to generate a power operation knowledge graph containing entity relationship weights and event causal chains, providing multi-dimensional associated input for feature engineering.

[0024] The three - level processing flow of the dynamic feature engineering in step S104: Perform multi - resolution analysis on non - stationary time - series data through wavelet packet transform, and extract the energy entropy of each frequency band as the harmonic distortion feature; Use conditional generative adversarial network to perform feature space amplification on historical abnormal samples to balance the distribution of positive and negative samples; The improved Isolation Forest model introduces information gain to optimize the feature splitting criterion, and improves the detection sensitivity to hidden fault modes. The pre - screening module dynamically divides the risk - level threshold according to the anomaly score, stores the events exceeding the adaptive threshold into the priority queue, and triggers the incremental training task of the hierarchical early - warning model.

[0025] The hierarchical early - warning model in step S105 adopts a dual - channel collaborative architecture: The short - term risk prediction channel deploys a spatio - temporal fusion Transformer model, captures the spatio - temporal dependence relationship of voltage and current sequences through the multi - head self - attention mechanism, and uses the interpretability module to calculate the feature contribution weight to locate the abnormal root device; The medium - and long - term risk prediction channel combines the deep deterministic policy gradient algorithm and proximal policy optimization to construct a multi - objective optimization strategy with the load balance degree and energy storage life loss as constraints. The transfer learning module uses the model parameters pre - trained on other regional power grids as the initialization benchmark, aligns the feature distribution differences through the domain adaptation algorithm, and accelerates the policy convergence efficiency in the new scenario.

[0026] The dynamic risk assessment in step S106 introduces a fuzzy Petri net model, maps the device - level fault probability and system - level risk indicators to the comprehensive risk level, and generates a multi - objective optimization strategy. The digital twin platform constructs a virtual mirror based on the real - time topology state, simulates the impact of the load transfer scheme on the power supply reliability, and outputs the optimal execution path. The blockchain smart contract adds a timestamp and digital signature to the generated scheduling instruction, verifies the instruction integrity through the decentralized node consensus mechanism, forms an anti - tampering audit tracking link, and ensures the trustworthy execution of the scheduling instruction.

[0027] The closed - loop optimization module in step S107 designs a two - level feedback mechanism: The local feedback adopts the elastic weight consolidation algorithm, calculates the Fisher information matrix of the model parameters for historical tasks, restricts the weight update amplitude in the incremental learning process, and prevents the model from forgetting historical knowledge; The global feedback analyzes the statistical characteristics of the prediction deviation of multiple batches through the meta - learning framework, and dynamically adjusts the feature selection weight and model hyperparameter configuration. The model distillation module transfers the knowledge of the complex early - warning model to the lightweight TinyLSTM network, softens the output probability distribution using the temperature scaling strategy, and adapts to the low - latency inference requirements of edge devices. The optimized model parameters are replaced with the old version in stages through the gray - release strategy, maintaining the high availability and stability of the early - warning system.

[0028] Specifically, for the audit data early - warning method described in the present invention, step S103 includes: Extract the node vulnerability features in the distribution network topology relationship through the graph embedding algorithm, and capture the non-linear correlation between the charge-discharge efficiency and the load fluctuation based on the spatio-temporal attention mechanism; Align the structured data of the SCADA system and the unstructured data of the smart meters through differential privacy technology and the Paillier homomorphic encryption algorithm for federated learning, and generate a cross-domain associated power operation knowledge graph containing device status, load distribution, and topology relationship.

[0029] The hidden features and dynamic association features flow into the dynamic feature engineering module as inputs, forming a data closed-loop for pre-screening abnormal patterns with the output of the anomaly detection model.

[0030] In the process of cross-domain data fusion, the graph embedding algorithm adopts the GraphSAGE unsupervised learning framework to hierarchically aggregate the adjacency matrix and node attribute data of the distribution network topology, extract low-dimensional vectors representing the electrical load rate, power supply radius, and connectivity of each node, and quantitatively evaluate the vulnerability indicators of each node. The spatio-temporal attention mechanism calculates the dynamic correlation coefficient matrix of the charge-discharge efficiency sequence of the energy storage system and the regional load fluctuation curve through the multi-head self-attention layer, identifies the response delay characteristics of the charge-discharge behavior to load mutations, and constructs a charge-discharge-load association weight graph. The federated learning framework introduces differential privacy technology to encrypt and align the structured monitoring data of the SCADA system and the unstructured load curves of the smart meters, and uses the homomorphic encryption algorithm to achieve cross-domain gradient sharing, generating a power operation knowledge graph containing device status, load distribution, and topology relationship. The edge weights of its entity relationships are jointly calculated by the node vulnerability vector and the spatio-temporal correlation matrix.

[0031] The knowledge graph construction module performs tensor fusion on the node vulnerability feature vector output by the graph embedding algorithm and the dynamic association matrix generated by the spatio-temporal attention mechanism, and iteratively updates the entity relationship weights through the graph convolutional network to form multi-dimensional association features including the health status of power supply equipment, regional load association patterns, and energy storage response characteristics. The fused feature data is normalized to generate an association feature vector of a unified dimension, which is input into the dynamic feature engineering module for abnormal pattern analysis. At the same time, the high-risk events output by the anomaly detection model are fed back to the knowledge graph update module to trigger the dynamic correction of the entity relationship weights, forming a closed-loop optimization link for cross-domain data fusion and anomaly detection.

[0032] Specifically, for the audit data warning method described in the present invention, step S104 includes: extracting harmonic distortion features from non-stationary time series data by using wavelet transform, and performing data enhancement on abnormal samples through a generative adversarial network; Input the enhanced features into an improved Isolation Forest algorithm that optimizes the feature splitting criterion using information gain for fast anomaly detection, and store the selected high-risk events in the event pool; The described event pool flows into the hierarchical early warning model as input, and forms a collaborative analysis of equipment-level and system-level risk assessments with the spatio-temporal feature modeling of the spatio-temporal fusion Transformer.

[0033] In the dynamic feature engineering process, the wavelet transform module uses the discrete wavelet packet decomposition algorithm to perform multi-scale decomposition on non-stationary time series data such as voltage and current, selects the Daubechies wavelet basis function to reconstruct the high-frequency subbands, and calculates the energy entropy of each subband as a harmonic distortion feature quantization index. The generator of the generative adversarial network adopts a multi-layer perceptron structure, inputs the feature vectors of historical abnormal events and the device type labels, and generates synthetic abnormal samples with diverse distributions; the discriminator introduces a conditional convolutional layer, combines the device operation status labels to verify the rationality of the generated samples, and optimizes the feature distribution alignment between the generated samples and the real abnormal samples through adversarial training. The feature set after data augmentation is input into the improved IsolationForest model, which uses information gain to optimize the feature splitting criterion, selects the feature dimension with the highest discrimination for spatial partitioning, and improves the detection sensitivity to low-probability abnormal events.

[0034] The anomaly detection model calculates the anomaly score based on the path length of the sample in the isolation tree, dynamically adjusts the threshold interval in combination with the sliding window statistical method, and marks the events with scores higher than the adaptive threshold as high-risk events. The event pool adopts a priority queue structure, stores events in a hierarchical manner according to the risk level and time urgency, and high-priority events trigger real-time alarms and are pushed to the hierarchical early warning model. The spatio-temporal fusion Transformer model receives the high-risk event features in the event pool, captures the spatio-temporal correlation patterns of the voltage and current sequences through the multi-head self-attention mechanism, and generates the equipment-level fault probability prediction results; at the same time, the reinforcement learning module dynamically adjusts the distribution network load allocation strategy based on the system-level risk features in the event pool, optimizes the charge and discharge scheduling instructions of the energy storage system, and forms a collaborative analysis link for equipment anomaly location and system strategy adjustment.

[0035] The data flow direction and model collaboration mechanism of the event pool form a closed-loop feedback: the fault location information output by the spatio-temporal fusion Transformer reversely corrects the feature extraction frequency band selection strategy of the wavelet transform; the strategy optimization results of the reinforcement learning module synchronously update the device state conditions input of the generative adversarial network, driving the data augmentation module to focus on generating abnormal features of the current system's weak links. The dynamic feature engineering module and the hierarchical early warning model achieve two-way interaction of data and strategies through the event pool, enhancing the joint perception ability of multi-dimensional operation risks of the power system.

[0036] Specifically, for the audit data early warning method described in the present invention, the step S105 includes: Introduce an interpretability module into the primary model, locate the contribution degree of abnormal features through SHAP values, and assist in generating device-level alarm instructions; Embed a transfer learning module into the secondary model, and use historical power grid data to pre-train the policy network to accelerate the convergence of the load distribution policy in the new scenario; The output of the primary model and the policy optimization result of the secondary model are jointly input into the risk assessment module to realize the linkage between short-term fault response and medium- and long-term risk prevention and control.

[0037] In the construction of the hierarchical early warning model, the interpretability module of the primary model adopts a SHAP value approximation calculation method based on kernel functions to perform feature attribution analysis on the device-level fault probability output by the spatio-temporal fusion Transformer. By calculating the marginal contribution degree of each voltage and current time-series feature to the abnormal score, a feature importance ranking list is generated, and a visual diagnostic report is generated in combination with the distribution network topology relationship. The alarm instruction generation module dynamically filters key abnormal indicators according to the feature contribution degree threshold, matches the preset device fault knowledge base, generates a structured alarm instruction containing location information and maintenance suggestions, and pushes it to the operation and maintenance terminal.

[0038] The transfer learning module of the secondary model adopts a pre-training - fine-tuning architecture: train a deep reinforcement learning policy network on historical power grid data to learn the global optimization mode of load distribution and energy storage scheduling; align the feature distribution differences between the old and new scenarios through a domain adaptation algorithm, and adjust the hidden layer parameters of the policy network using the maximum mean discrepancy metric. In the fine-tuning stage, a curriculum learning strategy is introduced to gradually increase the training weight of the new scenario data and accelerate the convergence speed of the policy network in the target area. The optimized load distribution policy outputs a dynamic scheduling instruction set, including the adjustment parameters of the distribution line load threshold and the charge and discharge power suggestions of the energy storage system.

[0039] The device-level alarm instruction of the primary model and the system-level scheduling instruction of the secondary model are jointly input into the risk assessment module, and a multi-level risk inference engine is constructed using a fuzzy Petri net. The device fault probability is used as the input of the place node, and the system policy optimization result is used as the transition trigger condition. The comprehensive risk level index is calculated through the fuzzy rule base. The risk linkage module triggers a hierarchical response according to the index threshold: low-risk events trigger device status tracking and monitoring; medium- and high-risk events synchronously start the simulation of the load transfer scheme and the pre-adjustment of the energy storage system. After verifying the feasibility of the policy through the digital twin platform, an execution instruction is generated. The collaborative output of the two-level model is encapsulated by a blockchain smart contract, and a time stamp and device signature are added to form an audit tracking chain, realizing the data closed-loop of short-term fault handling and medium- and long-term policy optimization.

[0040] Specifically, for the audit data early warning method described in the present invention, the step S106 includes: Adopt the sliding window statistical method to adjust the early warning threshold according to the real-time load status, and combine with the digital twin platform to simulate the execution effect of the energy storage scheduling strategy; Balance the power supply reliability and economy through the multi-objective optimization algorithm, generate the load transfer plan and push it to the operation and maintenance terminal; The simulation results and optimization strategies are input into the early warning execution module as a flow, forming an anti-tampering instruction transmission link with the blockchain traceability mechanism.

[0041] In the dynamic risk assessment process, the sliding window statistical module calculates the regional load mean and variance in real time according to the preset time window length, and dynamically adjusts the device-level early warning threshold interval in combination with the power supply network topology state. The window length is adaptively adjusted based on the load fluctuation period, using a short window to capture rapid changes during peak periods and a long window to smooth noise during low periods. The digital twin platform constructs a virtual mirror of the power supply network, injects real-time topology data and energy storage system state parameters, and simulates the impact of different charge and discharge strategies on node voltage stability and line loss through a discrete event simulation engine, outputting the load balance degree change curve and the predicted value of equipment life attenuation after the strategy execution.

[0042] The multi-objective optimization module uses the NSGA-III algorithm to solve the Pareto front of power supply reliability, economy and energy storage life loss, generating the optimal solution set of the load transfer plan. The optimization model inputs the simulation results of the digital twin as constraint conditions, generates candidate strategies through genetic algorithm crossover and mutation operations, calculates the weights of each objective using the entropy weight method, and screens the plan with the highest comprehensive score. The generated load transfer instruction contains the target line switching sequence and energy storage output adjustment parameters, which are pushed to the instruction execution queue of the operation and maintenance terminal after format standardization.

[0043] The blockchain traceability mechanism calculates the hash digest of the simulation results and optimization strategies, generates a unique data fingerprint and writes it into the distributed ledger. The instruction transmission link uses an asymmetric encryption algorithm to sign the load transfer plan, and verifies the identity permissions and data integrity of the execution node through a smart contract. The early warning execution module monitors the instruction execution status in real time after the strategy is issued, and feeds back the actual load transfer effect to the digital twin platform for effect verification, forming a closed-loop verification link of simulation - execution - feedback. The dynamically adjusted early warning threshold and optimization strategy realize version tracing through the timestamp sequence of the blockchain, supporting full-life-cycle audit and traceback.

[0044] Specifically, for the audit data early warning method described in the present invention, the step S107 includes: Constrain the model weight update in the incremental learning process through the elastic weight consolidation algorithm to prevent the forgetting of historical knowledge; Use the model distillation technology to transfer the knowledge of the complex hybrid model to the lightweight network to adapt to the real-time inference requirements of edge devices; The updated model weights and the results of knowledge transfer are input into the feature engineering module in the reverse direction, driving the iterative optimization of dynamic feature generation and anomaly detection.

[0045] In the closed-loop optimization link, the Elastic Weight Consolidation (EWC) algorithm calculates the Fisher information matrix of the model parameters in historical training tasks, identifies the weight parameters that are highly important for the learned knowledge, and introduces a regularization constraint term during the incremental learning process to limit the update amplitude of the key weights. The incremental learning module dynamically adjusts the constraint strength according to the newly arrived power system operation data, adopts an adaptive learning rate strategy to balance the learning weights of new and old knowledge, and maintains the model's memory ability for historical anomaly patterns.

[0046] The model distillation technology uses the temperature scaling strategy to smooth the output probability distribution of the complex hybrid model, generating soft target labels rich in decision-making logic. The lightweight network fits the joint distribution of the soft target and the true label through a multi-layer perceptron structure, optimizes the network parameters using the knowledge distillation loss function, and reduces the model complexity. The distilled lightweight network is deployed on the edge computing node to parse the sensor data stream in real-time and output the preliminary anomaly detection results, reducing the communication delay of cloud model inference.

[0047] The feature engineering module receives the updated model parameters from the Elastic Weight Consolidation algorithm and the anomaly detection results output by the distilled network, and dynamically adjusts the frequency band selection strategy of wavelet transform and the sample generation weights of the generative adversarial network. The feature selection module optimizes the combination method of multi-dimensional features according to the weight importance ranking feedback by the model, improving the accuracy of anomaly pattern recognition. The updated feature generation strategy is re-input into the incremental learning module for model fine-tuning, forming a two-way collaborative mechanism of feature optimization and model update, and enhancing the adaptive ability of the early warning system to the dynamic operation environment of the power system.

[0048] Specifically, for the audit data early warning method described in the present invention, the federated learning alignment of the structured data of the SCADA system and the unstructured data of the smart meter to generate a cross-domain associated power operation knowledge graph includes: During the data fusion process, differential privacy technology is used to protect the data security of each power subsystem, and the gradient is encrypted and transmitted through the Paillier homomorphic encryption algorithm to achieve cross-domain feature sharing; The topological relationship features of the knowledge graph and the data aligned by federated learning are jointly input into the spatio-temporal attention mechanism to enhance the robustness of the correlation analysis of charge and discharge behaviors and load fluctuations.

[0049] In the implementation of the federated learning framework, each power subsystem, as a participant, locally trains a feature extraction model, and uses the Laplace mechanism to add random noise that meets the differential privacy conditions to the model gradient to protect the privacy of the original data. The central aggregation server performs weighted averaging on the encrypted gradients of each participant through a secure multi-party computation protocol to generate global model update parameters. The Paillier homomorphic encryption algorithm is used in the gradient transmission process, enabling direct mathematical operations on the encrypted gradient ciphertext uploaded by the participants during the aggregation stage, thus avoiding the risk of data leakage in the decryption link.

[0050] The knowledge graph construction module receives the cross-domain data aligned by federated learning, aggregates the neighborhood features of the distribution network topology nodes through a graph attention network, and generates a topology relationship vector representing the regional power supply association strength. The spatio-temporal attention mechanism performs tensor splicing on the topology relationship vector and the charge-discharge efficiency features extracted by federated learning, and uses a multi-head self-attention layer to calculate the dynamic correlation matrix of the charge-discharge power sequence and the load fluctuation curve at different time scales. After the attention weight matrix is normalized by Softmax, a spatio-temporal association graph is generated, revealing the causal link between the energy storage response delay and the load mutation event, and enhancing the robustness of abnormal association pattern mining.

[0051] The collaborative effect of the federated learning framework and the knowledge graph forms a two-way data flow: the global model parameter update of federated learning drives the dynamic optimization of the topology relationship vector of the knowledge graph; the spatio-temporal association features output by the knowledge graph inversely guide the design of the local training objective function of the federated learning participants, enhancing the focusing ability of the feature extraction module on key association patterns. The fusion processing of the encrypted gradient and the topology feature is realized through an edge-cloud collaborative architecture. The edge node performs local model inference and privacy protection processing, and the cloud center completes cross-domain feature fusion and association analysis, constructing a secure and trustworthy distributed audit and warning architecture.

[0052] Specifically, for the audit data warning method described in the present invention, the data augmentation of abnormal samples through the generative adversarial network includes a generator and a discriminator. The generator is used to inject adversarial perturbations that simulate sensor noise into the input data, and the discriminator focuses on the classification training of high-risk samples through the Focal Loss function; Inject adversarial perturbations that simulate sensor noise into the input data, and focus on the classification training of high-risk samples through the Focal Loss function; The feature vector after the adversarial training and the output of the lightweight anomaly detection model form a high-precision pre-screening result, reducing the false alarm rate of the hierarchical warning model.

[0053] In the implementation of the adversarial generative network, the generator module adopts a conditional generative adversarial architecture, takes the input device type label and the normal state feature vector, and generates synthetic data injected with simulated sensor noise. The adversarial perturbation is modeled based on the historical sensor error distribution, generates a noise sequence with spatio-temporal correlation through a Gaussian mixture model, and superimposes it on the original voltage and current signals to simulate signal distortion in the real acquisition environment. The discriminator module adopts a deep convolutional network structure, combines spectral normalization technology to stabilize the training process, and distinguishes the distribution differences between real data and synthetic data through multi-layer feature extraction.

[0054] The Focal Loss function introduces a dynamic adjustment mechanism in the classifier training, dynamically calculates the weight factor according to the prediction confidence of abnormal samples, and increases the loss contribution ratio of high-risk samples. The classifier adopts a two-tower structure network, processes the feature vectors generated by adversarial training and the original feature vectors respectively, and aligns the feature space distribution through contrastive learning. During the training process, a hard sample mining strategy is adopted to iteratively resample the misclassified high-risk samples to improve the discrimination of the classification boundary.

[0055] The lightweight anomaly detection model adopts knowledge distillation technology to transfer the decision logic of the deep classifier after adversarial training to the TinyLSTM lightweight network. During the distillation process, the softened classification probability distribution is adjusted through the temperature parameter, and the KL divergence loss function is used to align the output distribution characteristics of the teacher-student models. The anomaly probability score output by the lightweight model is weighted and fused with the adversarial training feature vector to generate a comprehensive anomaly index, and the instantaneous false alarms are eliminated through moving window mean filtering before inputting to the hierarchical early warning model.

[0056] The collaborative training of the adversarial generative network and the lightweight model forms a dynamic optimization closed-loop: the false alarm detection result of the lightweight model is fed back to the generator module to trigger the iterative optimization of the adversarial noise pattern; the distribution characteristics of the high-risk samples output by the classifier synchronously update the weight adjustment strategy of the Focal Loss to enhance the adaptability to new anomaly patterns. The feature fusion module associates and stores the optimized comprehensive anomaly index with the device operation state label to construct a dynamically updated anomaly pattern knowledge base to support the continuous identification of new and old fault types by the early warning model.

[0057] Specifically, for the audit data early warning method described in the present invention, embedding a transfer learning module in the secondary model and using the historical power grid data to pre-train the policy network to accelerate the convergence of the load distribution policy in the new scenario includes: Initializing the reinforcement learning network parameters with the pre-trained distribution network load distribution policy, and dynamically adjusting the energy storage scheduling actions through the proximal policy optimization algorithm; The collaborative output of the transfer learning and the reinforcement learning drives the policy simulation of the digital twin platform to form a closed-loop verification link from model training to actual execution.

[0058] In the implementation of the transfer learning module, in the pre-training stage, the load distribution data of the historical distribution network and the device operation logs are used to train a deep neural network to learn the load balancing distribution patterns under different topological structures. The policy network parameters of the reinforcement learning network are initialized with the weights of the pre-trained model. The domain adaptation algorithm is used to align the differences in the node voltage and line impedance feature distributions between the old and new scenarios, and the maximum mean discrepancy loss function is used to adjust the feature mapping relationship of the hidden layer. The proximal policy optimization algorithm introduces a trust region constraint in the action selection process to limit the policy update step size to maintain training stability, and the generalized advantage function is used to evaluate the long-term benefits of the energy storage charge and discharge actions, dynamically adjusting the charge and discharge power thresholds and response priorities.

[0059] The digital twin platform constructs a virtual mirror of the distribution network, imports real-time topological data and the state-of-charge parameters of the energy storage system, and simulates the node voltage fluctuations and line loss changes under different scheduling strategies through a discrete event simulation engine. The simulation results are compared with the scheduling instructions output by the reinforcement learning for feasibility, and a candidate policy set that meets the power supply reliability constraints is screened. The curriculum learning module dynamically adjusts the training data distribution according to the policy execution difficulty, preferentially learning the historical scenario data with a high degree of match with the current system state, and accelerating the convergence speed of the policy network in the target area.

[0060] The synergistic effect of transfer learning and reinforcement learning is realized through a closed-loop verification link: the load distribution knowledge of the pre-trained model drives the generation of the initial strategy of the digital twin platform; the optimized strategy after simulation verification iteratively updates the network parameters through the proximal policy optimization algorithm; the output of the updated policy network synchronously corrects the parameters of the domain adaptation module of the pre-trained model. The blockchain smart contract adds device digital signatures and timestamps to the verified scheduling instructions to generate an anti-tampering audit trail record, forming a trusted control link from virtual simulation to physical execution.

[0061] In a second aspect, the audit data warning system provided by the present invention is applied to an audit data warning method, and includes: A distributed sensor network configured to collect node voltage, current of the power supply network, and charge and discharge efficiency data of the energy storage system in real time; An edge computing node connected to the sensor network and configured to perform local cleaning and standardization processing on the original data to generate standardized intermediate data; A multi-source data fusion module connected to the edge computing node and configured to extract cross-domain association features through federated learning and knowledge graph technology; A dynamic feature engineering module connected to the multi-source data fusion module and configured to generate multi-dimensional operating state features and perform pre-screening of abnormal patterns; The hierarchical early warning model construction module, connected to the dynamic feature engineering module, is configured to train a spatio-temporal fusion Transformer and a hybrid reinforcement learning framework; The risk assessment and decision-making module, connected to the hierarchical early warning model construction module, is configured to generate a dynamic risk level and an energy storage scheduling strategy; The closed-loop optimization module, connected to the risk assessment and decision-making module, is configured to achieve adaptive update of the early warning model through meta-learning and model distillation techniques; Among them, the data of the sensor network flows to the edge computing node, the standardized intermediate data is input into the dynamic feature engineering module after being processed by the multi-source fusion module, the output of the hierarchical early warning model drives risk assessment and decision-making, and the decision result is fed back to the closed-loop optimization module in reverse.

[0062] The distributed sensor network is deployed at the monitoring points of distribution transformers, line sectional switches and energy storage system battery clusters in the power supply network, and includes Hall current sensors, voltage transformers and wireless temperature sensors. The sensor nodes use the low-power wide-area network protocol for data transmission, and perform sliding window filtering on the high-sampling-rate current signals to eliminate high-frequency electromagnetic interference noise. The edge computing node is built-in with an outlier detection algorithm, establishes a voltage fluctuation confidence interval based on the historical data distribution, and performs linear interpolation compensation on the raw data exceeding the threshold. The data encapsulation module converts multi-source heterogeneous data into a standardized JSON format, adds device geographical tags and time synchronization identifiers, and transmits them to the edge node message queue through the MQTT protocol.

[0063] The local cleaning module of the edge computing node uses a moving average algorithm to smooth the load fluctuation data, and performs missing value filling based on Gaussian process on the charge and discharge efficiency data of the energy storage system. The standardization processing module performs dynamic range compression on the voltage and current time series data, maps the values to the [-1, 1] interval; converts the unstructured topology data into an adjacency matrix and a node attribute table. The time series alignment module uses the dynamic time warping algorithm to unify the sampling frequencies of multi-source data and generates an intermediate data set with a unified time index. The data version management module attaches timestamps and hash check codes to the intermediate data, supporting incremental updates and version backtracking.

[0064] The federated learning framework of the multi-source data fusion module adopts a hierarchical architecture, establishes an encrypted communication channel between the SCADA system and the smart meter data source, and realizes cross-domain gradient aggregation through homomorphic encryption technology. The knowledge graph construction module applies the GraphSAGE algorithm to perform unsupervised learning on the distribution network topology, and generates low-dimensional vectors representing the power supply capacity of nodes and the regional relevance. The spatio-temporal attention mechanism calculates the dynamic correlation matrix between the charge and discharge power sequence of the energy storage system and the regional load curve, and identifies the lag effect of load mutations on the energy storage response. The fused cross-domain feature vectors are appended with device status labels and input into the dynamic feature engineering module for abnormal pattern mining.

[0065] The wavelet transform unit of the dynamic feature engineering module uses the DB4 wavelet basis to perform multi-scale decomposition on the voltage harmonic data, and extracts the energy entropy of each sub-band as the harmonic distortion feature. The generator module of the generative adversarial network generates adversarial samples based on the device type label, and the discriminator uses a spectrally normalized convolutional network to verify the rationality of the samples. The improved IsolationForest model introduces a feature importance weighting mechanism, optimizes the tree structure splitting criterion, and improves the detection sensitivity to hidden faults. The pre-screening results are stored in the priority event pool, triggering the real-time training task of the hierarchical warning model.

[0066] The spatio-temporal fusion Transformer of the hierarchical warning model construction module uses a multi-head self-attention layer to capture the spatio-temporal dependence of the voltage sequence, and outputs a device-level fault heat map. The reinforcement learning framework combines the proximal policy optimization algorithm to generate a scheduling strategy with the load balance degree and the energy storage life loss as the constraint conditions. The transfer learning module uses the model parameters pre-trained on the historical power grid as the initialization benchmark, aligns the feature distributions of the new and old scenarios through domain adversarial training, and shortens the policy convergence period.

[0067] The fuzzy inference engine of the risk assessment and decision-making module maps the device failure probability and the system policy score to a comprehensive risk index, triggering the hierarchical response mechanism. The digital twin platform constructs a virtual mirror of the power supply network, injects real-time topology parameters to simulate the impact of the load transfer strategy on the line loss, and outputs the optimal execution path. The blockchain smart contract calculates the hash digest of the scheduling instruction and performs multiple signatures to generate an anti-tampering audit trail chain.

[0068] The meta-learning framework of the closed-loop optimization module analyzes the multi-batch prediction deviation features, and dynamically adjusts the feature selection weights and the model hyperparameter configuration. The elastic weight consolidation algorithm calculates the Fisher information matrix of the model parameters for historical tasks, and constrains the weight update amplitude during the incremental learning process. The model distillation module uses the temperature scaling strategy to transfer the knowledge of the spatio-temporal fusion Transformer to the lightweight TinyLSTM network, adapting to the real-time inference requirements of edge nodes. The optimized model parameters are deployed in stages through the gray release mechanism to maintain the continuity of the system service.

[0069] The data flow between modules forms a closed-loop control link: the raw data collected by the sensor network flows to the multi-source fusion module after being processed by the edge nodes; the cross-domain features extracted by federated learning drive dynamic feature engineering and anomaly detection; the output of the hierarchical early warning model guides risk assessment decisions; the execution instructions encapsulated by the blockchain are fed back to the digital twin platform for effect verification; the closed-loop optimization module iteratively updates the model parameters and feature generation strategies according to the verification results. The collaboration between federated learning and knowledge graph realizes feature fusion under privacy protection, the linkage between adversarial training and model distillation balances detection accuracy and inference efficiency, and the combination of transfer learning and reinforcement learning accelerates the strategy optimization process. The overall architecture meets the technical requirements of intelligent monitoring and energy storage scheduling in the power system.

[0070] The specific implementation of the present invention in the technical field of circuit devices for power supply or distribution and energy storage systems is as follows: The distributed sensor network is deployed at the transformer, sectional switch node of the distribution network, and the battery cluster monitoring points of the energy storage system, including Hall current sensors, voltage transformers, and temperature sensors. The sensor nodes transmit data through the low-power wide-area network protocol, and the sampling frequency is set to 1 kHz. The sliding window filtering algorithm is used to eliminate high-frequency noise from voltage and current signals, and the window length is adaptively adjusted according to the load fluctuation period. The edge computing node is built-in with an interquartile range outlier detection module, which performs linear interpolation compensation and normalization processing on the raw data, generates an intermediate data stream in JSON format, attaches a timestamp and device geographical tags, and pushes it to the cloud message queue through the MQTT protocol.

[0071] The multi-source data fusion module adopts a hierarchical federated learning framework, establishes an encrypted communication channel between the SCADA system and the smart meter data source, realizes cross-domain gradient sharing through the Paillier homomorphic encryption algorithm, and combines the graph embedding algorithm to extract the vulnerability feature vector of the distribution network topology nodes, with the dimension set to 128 dimensions. The knowledge graph construction module applies the GraphSAGE algorithm to aggregate the attributes of neighboring nodes, generates a topological relationship matrix representing the power supply association strength of the region, and performs tensor fusion with the charge-discharge-load dynamic association matrix calculated by the spatio-temporal attention mechanism to construct a power operation knowledge graph.

[0072] In the dynamic feature engineering module, the wavelet transform unit uses the DB4 wavelet basis to decompose the voltage harmonic data by 5 layers, and extracts the energy entropy of each sub-band as the harmonic distortion feature; the generator of the generative adversarial network generates adversarial samples with simulated sensor noise based on the device type label, and the discriminator uses a spectrally normalized convolutional network to verify the rationality of the samples. The improved Isolation Forest model introduces information gain to optimize the feature splitting criterion, sets the number of trees to 100, and filters out the events with abnormal scores higher than the dynamic threshold and stores them in the priority event pool.

[0073] In the hierarchical early warning model construction module, the spatio-temporal fusion Transformer architecture uses 4-head self-attention layers, the input sequence length is 60 time steps, and the device-level fault heat map is output; the hybrid reinforcement learning framework takes the load balance threshold of ±5% and the energy storage life attenuation rate of ≤0.1% / cycle as constraints, and generates the charge and discharge scheduling strategy through the proximal policy optimization algorithm. The transfer learning module pre-trains the policy network parameters using historical power grid data, adopts domain adversarial training to align the feature distribution differences, and the learning rate in the fine-tuning stage is set to 1e-4.

[0074] In the risk assessment and decision-making module, the fuzzy Petri net model maps the device failure probability to a comprehensive risk index in the interval [0,1], combines the digital twin platform to simulate the impact of the load transfer strategy on the line loss, the simulation step size is set to 1 second, and the optimal execution path is output. The blockchain smart contract performs SHA-256 hash calculation on the scheduling instruction, adopts a multi-signature mechanism to verify the instruction integrity, and generates an anti-tampering audit trace chain.

[0075] The closed-loop optimization module calculates the Fisher information matrix of the model parameters through the elastic weight consolidation algorithm, constrains the weight update amplitude in the incremental learning process, and the regularization coefficient is set to 0.5; the model distillation module uses the temperature parameter T = 2 to smooth the output distribution of the spatio-temporal fusion Transformer, transfers the knowledge to the lightweight TinyLSTM network, and compresses the number of layers to 3 layers to adapt to the real-time inference requirements of the edge computing node. The data flow between modules is realized by the Kafka message queue for asynchronous transmission to ensure the high concurrency processing ability of the system.

[0076] The explanations of the technical feature terms of the present invention are as follows: Distributed sensor network: A set of sensors deployed at key nodes of the power supply network (such as distribution transformers, line sectional switches) and monitoring points of the energy storage system, including voltage transformers, Hall current sensors, temperature sensors, etc. Real-time collection of voltage, current, load fluctuation data and energy storage system charge and discharge efficiency, capacity attenuation rate and other parameters through low-power wide-area network protocols (such as LoRaWAN) to generate the original data stream.

[0077] Function: Provide a high-precision and real-time data source for subsequent analysis, covering multi-dimensional information of the physical layer and operating state of the power system.

[0078] Edge computing node: A local computing unit deployed at the data acquisition end, responsible for the preliminary processing of the original data, including data cleaning (such as eliminating noise by sliding window filtering), outlier detection (based on the interquartile range algorithm), missing value compensation (linear interpolation or Kalman filtering), and normalization (normalizing the data to the interval [-1,1]).

[0079] Function: Reduce the cloud transmission pressure, improve data quality, and provide structured input for subsequent analysis (such as standardized intermediate data in JSON format).

[0080] Federated learning framework: A distributed machine learning technology that allows multiple participating parties (such as SCADA systems, smart meters) to collaboratively train models without sharing raw data. In the present invention, a hierarchical federated learning architecture is adopted, and cross-domain data alignment and feature sharing are achieved through differential privacy technology (adding Laplace noise) and Paillier homomorphic encryption algorithm (encrypting gradient transmission).

[0081] Function: While protecting the data privacy of each subsystem, integrate the structured monitoring data of the SCADA system and the unstructured load curves of smart meters, and extract cross-domain correlation features.

[0082] Knowledge graph technology: A data representation method based on graph structure, used to construct a power operation knowledge graph. Extract the vulnerability features of distribution network topology nodes (such as electrical load rate, power supply radius) through graph embedding algorithms (such as GraphSAGE), and use spatio-temporal attention mechanisms to capture the dynamic associations between charge-discharge efficiency and load fluctuations.

[0083] Function: Form a multi-dimensional semantic network including device status, load association patterns, and energy storage response characteristics, and enhance the cross-modal data association analysis ability.

[0084] Self-supervised learning: A machine learning method that does not require manual labeling of labels, and generates supervision signals through the structure of the data itself. In the present invention, it is used for dynamic feature engineering processing, for example, extracting harmonic distortion features from non-stationary time series data through wavelet transform, or generating abnormal samples through generative adversarial networks (GANs) to enhance the data distribution.

[0085] Function: Improve the recognition sensitivity of the anomaly detection model to hidden faults and solve the problem of sample imbalance.

[0086] Spatio-temporal fusion Transformer architecture: A deep learning model that combines time series and spatial dependencies. Adopt a multi-head self-attention mechanism to model the spatio-temporal associations of voltage and current sequences (such as the impact of device-level voltage fluctuations on adjacent nodes), and use an interpretability module (such as SHAP values) to locate the contribution degree of abnormal features.

[0087] Function: Accurately predict the device-level fault probability (such as transformer overload risk) and generate a visual diagnostic report.

[0088] Hybrid Reinforcement Learning Framework: Combines the Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO) algorithms. Constrained by load balancing degree and energy storage life loss, it optimizes the load distribution and energy storage scheduling strategies of the distribution network. The policy network is pre-trained through the transfer learning module to accelerate policy convergence in new scenarios.

[0089] Function: Achieves system-level risk prevention and control (such as avoiding line overload) and medium- and long-term economic optimization (such as extending the life of energy storage).

[0090] Model Distillation Technology: Transfers the knowledge of complex models (such as spatio-temporal fusion Transformer) to lightweight networks (such as TinyLSTM). Adopts the temperature scaling strategy to soften the output probability distribution, and aligns the decision-making logic of the teacher and student models through the KL divergence loss function.

[0091] Function: Compresses the model size, adapts to the low-latency inference requirements of edge devices, and maintains detection accuracy at the same time.

[0092] Meta-Learning Framework: A "learning how to learn" mechanism that dynamically adjusts feature weights and model hyperparameters by analyzing the statistical characteristics of multi-batch prediction deviations. Combines the Elastic Weight Consolidation (EWC) algorithm to prevent the forgetting of historical knowledge in incremental learning.

[0093] Function: Achieves the adaptive iteration of the early warning model and adapts to the dynamic operating environment of the power system.

[0094] Digital Twin Platform: Builds a virtual mirror of the power supply network based on real-time topology data, and simulates the impact of different scheduling strategies (such as load transfer schemes) on line loss and voltage stability through a discrete event simulation engine.

[0095] Function: Verifies the feasibility of the strategy, outputs the optimal execution path, and combines blockchain technology to achieve tamper-proof instruction traceability.

[0096] Federated Learning Framework: Achieves the secure fusion of cross-domain data, and integrates multi-source heterogeneous data on the premise of protecting the data privacy of each subsystem (such as SCADA systems and smart meters).

[0097] Implementation Method: Hierarchical Federated Architecture: Each participating party (power subsystem) locally trains the model, and the central server aggregates the encrypted gradients to generate the global model.

[0098] Privacy Protection Technology: Differential Privacy: Adds Laplace noise to the model gradient to prevent the leakage of original data.

[0099] Paillier Homomorphic Encryption: Encrypts the gradients for transmission to ensure the security of the cross-domain sharing process.

[0100] Application: Align the structured monitoring data of the SCADA system with the unstructured load curves of smart meters to generate cross-domain correlation features, supporting subsequent knowledge graph construction.

[0101] Knowledge graph technology: Construct a power operation knowledge graph to represent the associations of device status, topological relationships, and dynamic behaviors.

[0102] Implementation method: Graph Embedding Algorithm (GraphSAGE): Perform unsupervised learning on the distribution network topology to extract node vulnerability features (such as electrical load rate, power supply radius).

[0103] Spatio-temporal attention mechanism: Calculate the dynamic correlation matrix of the charge-discharge efficiency sequence and the load fluctuation curve to identify the risk propagation path.

[0104] Application: Generate a multi-dimensional semantic network containing device health status and regional load association patterns, providing an association basis for anomaly detection and risk assessment.

[0105] Self-supervised learning: Automatically extract features from unlabeled data to solve the problem of scarce abnormal samples.

[0106] Implementation method: Wavelet transform: Perform multi-resolution decomposition on non-stationary time series data (such as voltage harmonics), and extract the energy entropy of each frequency band as the harmonic distortion feature.

[0107] Generative Adversarial Network (GAN): Generator: Simulate sensor noise to synthesize diverse abnormal samples.

[0108] Discriminator: Focus on the classification training of high-risk samples through the Focal Loss function.

[0109] Application: Enhance the data distribution of the anomaly detection model and improve the recognition sensitivity to concealed faults (such as partial discharge).

[0110] Spatio-temporal fusion Transformer architecture: Model the spatio-temporal dependence of the power supply network time series data to accurately locate device-level faults.

[0111] Implementation method: Multi-head self-attention mechanism: Capture the spatio-temporal associations of voltage and current sequences (such as the impact of overload at a certain node on adjacent lines).

[0112] Interpretability module (SHAP value): Calculate the contribution degree of each feature to the anomaly score and generate a visual diagnostic report.

[0113] Application: Output a device-level fault heat map (such as the probability of transformer insulation aging), assisting maintenance personnel to quickly locate the root cause of anomalies.

[0114] Hybrid Reinforcement Learning Framework: Optimize the load distribution and energy storage scheduling strategies in the distribution network to balance power supply reliability and economy.

[0115] Implementation Method: Deep Deterministic Policy Gradient (DDPG): Generate a continuous action space policy with the load balance degree and energy storage life loss as constraint conditions.

[0116] Proximal Policy Optimization (PPO): Dynamically adjust the charge and discharge power thresholds through trust region constraints to ensure policy stability.

[0117] Transfer Learning Module: Use historical power grid data to pre-train the policy network and accelerate policy convergence in new scenarios.

[0118] Application: Generate the optimal load transfer scheme (such as switching to the standby line) and energy storage scheduling instructions (such as adjusting the charge and discharge power) to avoid system-level risks (such as cascading failures).

[0119] Model Distillation Technology: Compress the scale of complex models to adapt to edge computing resources.

[0120] Implementation Method: Temperature Scaling: Soften the output probability distribution of the spatio-temporal fusion Transformer to generate soft target labels.

[0121] KL Divergence Loss Function: Align the decision-making logics of the teacher-student models (complex model and lightweight TinyLSTM).

[0122] Application: Deploy the device-level fault prediction model to edge nodes to achieve low-latency real-time inference (such as millisecond-level anomaly detection).

[0123] Meta-Learning Framework: Dynamically optimize model parameters and feature weights to adapt to the dynamic operating environment of the power system.

[0124] Implementation Method: Elastic Weight Consolidation (EWC) Algorithm: Calculate the Fisher information matrix of model parameters for historical tasks, constrain the weight update amplitude in incremental learning, and prevent knowledge forgetting.

[0125] Deviation Analysis Module: Statistically analyze the prediction error features of multiple batches and dynamically adjust the feature selection weights (such as the priority of harmonic distortion features).

[0126] Application: Achieve the adaptive iteration of the early warning model (such as periodic updates) and improve the generalization ability for new fault modes (such as new energy grid connection disturbances).

[0127] Digital Twin Platform: Simulate the policy execution effect and verify the feasibility of scheduling instructions.

[0128] Implementation Method: Discrete Event Simulation Engine: Construct a virtual image based on real-time topology data to simulate the impact of load transfer on line loss and voltage stability.

[0129] Blockchain Traceability Mechanism: Calculate the hash digest of the simulation results and optimization strategies to generate a tamper-proof audit trail chain.

[0130] Application: Output the optimal execution path (such as switching loads in stages), and achieve the trusted execution of instructions through smart contracts.

[0131] The present invention realizes the efficient acquisition and preprocessing of data through a distributed sensor network and edge computing, fuses multi-source heterogeneous data by using federated learning and knowledge graph technology, improves the accuracy of anomaly detection by combining self-supervised learning and spatio-temporal fusion Transformer, optimizes the system-level scheduling strategy through hybrid reinforcement learning, and achieves model lightweight and adaptive iteration by means of model distillation and meta-learning. The synergistic effect of each technical feature finally achieves an overall improvement in the real-time performance of power system anomaly detection, the accuracy of risk assessment, and the economy of energy storage scheduling.

[0132] The technical solution of the present invention effectively solves the problem of lag in anomaly detection caused by the traditional power system relying on static thresholds through dynamic feature engineering and incremental learning mechanisms. Wavelet transform is used to extract harmonic distortion features from non-stationary time series data, combined with a generative adversarial network to enhance the diversity distribution of anomaly samples and optimize the sensitivity of the lightweight anomaly detection model; self-supervised learning dynamically adjusts feature weights, combined with the meta-learning framework to analyze prediction deviations, forming a feature generation strategy adapted to the changes in the operating state of the power system, and improving the ability to identify hidden faults under complex working conditions.

[0133] Aiming at the problems of insufficient utilization of multi-source data and privacy protection, the multi-source data fusion module integrates heterogeneous data from SCADA systems, smart meters, and energy storage BMS through a federated learning framework, and uses differential privacy and homomorphic encryption technologies to achieve cross-domain gradient sharing. The knowledge graph technology fuses the dynamic correlation matrix generated by the vulnerability characteristics of topological nodes and the spatio-temporal attention mechanism to construct a multi-dimensional knowledge graph including device health status, load association patterns, and energy storage response characteristics, enhancing the semantic correlation analysis ability of cross-modal data and solving the problem of one-sidedness in traditional single data source analysis.

[0134] To optimize the system-level risk prevention and control efficiency, the hierarchical early warning model adopts a spatio-temporal fusion Transformer architecture to capture the spatio-temporal dependencies of voltage and current sequences, accurately locate the root causes of device-level anomalies; the reinforcement learning framework combines transfer learning pre-training strategies to dynamically optimize the load distribution and energy storage scheduling actions of the distribution network, and simulates the execution effects of the strategies through a digital twin platform. The closed-loop optimization module constrains the parameter updates in incremental learning through the elastic weight consolidation algorithm, combines model distillation technology to compress the scale of complex models, adapts to edge computing resources, realizes the efficient iteration of the early warning model and real-time inference at the edge, and finally improves the anomaly response speed of the power supply network and the economy of energy storage scheduling.

Claims

1. An audit data warning method, characterized in that include: Step S101: collecting node voltage, current, and load fluctuation data of the power supply network, as well as charge and discharge efficiency and capacity attenuation rate data of the electric energy storage system, to generate an original data stream; Step S102: Input the original data stream into the edge computing node for local cleaning and standardized format conversion to generate standardized intermediate data; Step S103: Based on the federated learning framework and knowledge graph technology, cross-domain data fusion is performed on the standardized intermediate data to extract the hidden features of the distribution network topology relationship and the dynamic correlation features of the charging and discharging behavior and load fluctuation, and generate a correlation feature vector; Step S104: Dynamic feature engineering is performed on the associated feature vector using self-supervised learning to generate multi-dimensional power system operation status features, and abnormal pattern pre-screening is performed on the high-dimensional features using a lightweight anomaly detection model to output a high-risk event pool; Step S105: constructing a hierarchical early warning model based on the high-risk event pool, using a spatiotemporal fusion Transformer architecture to model the spatiotemporal characteristics of the power supply network time series data to predict device-level failure probabilities, and optimizing the distribution network load distribution and energy storage scheduling strategies through a hybrid reinforcement learning framework to predict system-level risks; Step S106: generating a dynamic risk assessment level based on the device-level failure probability and system-level risk prediction results, combining fuzzy logic reasoning and evidence theory, and triggering a real-time alarm or generating energy storage scheduling suggestions and load transfer plans; Step S107: Collect system response data after the early warning strategy is executed, analyze the model prediction deviation through the meta-learning framework, and dynamically adjust the feature weights and model structure to form a closed-loop optimization link; Among them, the original data flows through the edge computing node and then flows to the data fusion module, the associated feature vector is input into the dynamic feature engineering module to generate pre-screening results, the high-risk event pool drives the training and decision-making of the graded warning model, and the risk assessment results are fed back to the model optimization module.

2. The audit data warning method according to claim 1, characterized in that The step S103 includes: A graph embedding algorithm is used to extract node vulnerability characteristics in the topology of the distribution network, and a spatiotemporal attention mechanism is used to capture the nonlinear relationship between charging and discharging efficiency and load fluctuations. Align the structured data of the SCADA system with the unstructured data of smart meters through federated learning to generate a cross-domain associated power operation knowledge graph; The hidden features and dynamic correlation features are fed into the dynamic feature engineering module as input, and form a data closed loop for pre-screening abnormal patterns with the output of the anomaly detection model.

3. The audit data warning method according to claim 2, wherein, The step S104 includes: extracting harmonic distortion features from non-stationary time series data using wavelet transform, and performing data enhancement on abnormal samples using a generative adversarial network; The enhanced features are input into the improved Isolation Forest algorithm for rapid anomaly detection, and the selected high-risk events are stored in the event pool; The event pool serves as an input to the hierarchical warning model, and forms a collaborative analysis of device-level and system-level risk assessment with the spatiotemporal feature modeling of the spatiotemporal fusion Transformer.

4. The auditing data early warning method according to claim 3, wherein The step S105 includes: Introduce an interpretability module into the primary model, locate the contribution degree of abnormal features through SHAP values, and assist in generating device-level alarm instructions; Embed a transfer learning module in the secondary model, and use historical power grid data to pre-train the policy network to accelerate the convergence of the load distribution strategy in new scenarios; The output of the primary model and the policy optimization result of the secondary model are jointly input into the risk assessment module to realize the linkage between short-term fault response and medium- and long-term risk prevention and control.

5. The audit data warning method according to claim 4, wherein The step S106 includes: Adopt the sliding window statistical method to adjust the warning threshold according to the real-time load status, and combine the digital twin platform to simulate the execution effect of the energy storage scheduling strategy; Balance power supply reliability and economy through a multi-objective optimization algorithm, generate a load transfer plan and push it to the operation and maintenance terminal; The simulation results and optimization strategies flow into the warning execution module as inputs, forming an anti-tampering instruction transmission link with the blockchain traceability mechanism.

6. The audit data warning method according to claim 5, wherein The step S107 includes: Constrain the model weight update in the incremental learning process through the elastic weight consolidation algorithm to prevent the forgetting of historical knowledge; Use model distillation technology to transfer the knowledge of complex hybrid models to lightweight networks to adapt to the real-time inference requirements of edge devices; The model weight update and knowledge transfer results are reversely input into the feature engineering module to drive the iterative optimization of dynamic feature generation and anomaly detection.

7. The audit data warning method according to claim 6, characterized in that The alignment of the structured data of the SCADA system and the unstructured data of the smart meter through federated learning to generate a cross-domain associated power operation knowledge graph includes: Adopt differential privacy technology to protect the data security of each power subsystem during the data fusion process, and realize cross-domain feature sharing through gradient encrypted transmission; The topological relationship features of the knowledge graph and the data aligned by federated learning are jointly input into the spatio-temporal attention mechanism to enhance the robustness of the correlation analysis between charge and discharge behaviors and load fluctuations.

8. The audit data warning method according to claim 7, characterized in that The data augmentation of abnormal samples through the generative adversarial network includes a generator and a discriminator, where the generator is used to inject adversarial perturbations simulating sensor noise into the input data, and the discriminator focuses on the classification training of high-risk samples through the Focal Loss function; Inject adversarial perturbations simulating sensor noise into the input data, and focus on the classification training of high-risk samples through the Focal Loss function; The feature vectors after adversarial training and the output of the lightweight anomaly detection model form a high-precision pre-screening result, reducing the false alarm rate of the hierarchical warning model.

9. The audit data warning method according to claim 8, wherein The embedding of the transfer learning module in the secondary model, using historical power grid data to pre-train the policy network to accelerate the convergence of the load distribution strategy in new scenarios includes: Use the pre-trained load distribution strategy of the distribution network to initialize the parameters of the reinforcement learning network, and dynamically adjust the energy storage scheduling actions through the proximal policy optimization algorithm; The collaborative output of transfer learning and reinforcement learning drives the policy simulation of the digital twin platform, forming a closed-loop verification link from model training to actual execution.

10. An audit data warning system, applied to the audit data warning method as described in any one of claims 1 to 9, characterized in that, Includes: A distributed sensor network configured to collect data on the node voltage, current of the power supply network, and the charge and discharge efficiency of the energy storage system in real time; An edge computing node, connected to the sensor network, configured to perform local cleaning and standardization processing on the original data to generate standardized intermediate data; A multi-source data fusion module, connected to the edge computing node, configured to extract cross-domain association features through federated learning and knowledge graph technology; A dynamic feature engineering module, connected to the multi-source data fusion module, configured to generate multi-dimensional operating state features and perform pre-screening of abnormal patterns; A hierarchical early warning model construction module, connected to the dynamic feature engineering module, configured to train a spatio-temporal fusion Transformer and a hybrid reinforcement learning framework; A risk assessment and decision-making module, connected to the hierarchical early warning model construction module, configured to generate a dynamic risk level and an energy storage scheduling strategy; A closed-loop optimization module, connected to the risk assessment and decision-making module, configured to achieve adaptive update of the early warning model through meta-learning and model distillation technology; Wherein, the data of the sensor network flows to the edge computing node, the standardized intermediate data is input into the dynamic feature engineering module after being processed by the multi-source fusion module, the output of the hierarchical early warning model drives risk assessment and decision-making, and the decision result is fed back to the closed-loop optimization module in reverse.

Images