A new energy station fine modeling and grid-connected comprehensive detection system

Abstract

The application discloses a new energy station fine modeling and grid-connected comprehensive detection system, which comprises three layers of edge layer, area layer and cloud layer. The edge layer is provided with data-driven base classifier, knowledge-driven base classifier, fusion unit, lightweight twin model and prototype network sampler, so that knowledge-data double-drive fusion detection and low-confidence anomaly sampling are realized. The area layer is provided with super network management unit, causal weighted aggregation unit, privacy protection unit and causal distillation unit, so that parameter aggregation based on federal learning and cross-area knowledge migration are realized. The cloud layer is provided with structure causal diagram engine, causal test unit, digital twin sandbox and knowledge graph inference machine, so that global causal diagram is constructed, intervention simulation is executed and knowledge base is evolved. The system forms a closed loop of detection, tracing, correction and evolution, and has high detection accuracy, interpretability, privacy protection and self-evolution ability, so that the intelligent level of new energy station grid-connected comprehensive detection is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation and computer intelligent system technology, and in particular relates to a refined modeling and grid-connected integrated detection system for new energy power plants. Background Technology

[0002] Refined modeling and integrated grid-connection monitoring of new energy power plants (such as wind farms and photovoltaic power plants) are key technologies for ensuring the safe and stable operation of new power systems. Existing technologies for equivalent modeling of new energy power plants mainly include single-unit equivalent methods, multi-unit clustering methods, and dynamic equivalent methods. These methods collect operational data from each power generation unit within the power plant and use clustering algorithms such as K-means and support vector machines to group the units and establish an equivalent model reflecting the overall external characteristics of the power plant. Regarding grid-connection monitoring, conventional systems employ threshold-based protection logic or offline analysis based on fault recordings. Some advanced systems introduce deep learning methods such as convolutional neural networks and long short-term memory networks for fault classification and identification. For multi-power plant collaboration, existing research has adopted a centralized data aggregation architecture, uploading all power plant data to a central server for unified training and monitoring. Furthermore, some systems utilize digital twin technology to construct power plant simulation models and achieve model calibration through offline parameter identification. Overall, existing technologies have made some progress in equivalent modeling, fault detection, and data analysis.

[0003] However, existing technologies still have the following shortcomings: First, modeling and detection are disconnected. Once the equivalent model is established, it remains fixed and cannot be corrected using anomalies discovered during grid-connected testing. This leads to a decline in model accuracy over time and a lack of accurate characterization of new faults such as broadband oscillations. Second, detection methods rely on purely data-driven approaches. When historical fault samples are scarce, the accuracy drops sharply, and the output lacks interpretability, making it impossible to pinpoint the root cause of the fault. Third, there is a significant conflict between data privacy protection and cross-site knowledge sharing. Centralized architecture requires each site to upload raw data, posing a risk of commercial secret leakage. However, independent modeling cannot utilize common patterns in data from multiple sites. Fourth, the model update cycle is long. New site integration requires accumulating a large number of samples to achieve usable performance, making it unable to adapt to dynamic changes such as equipment aging and control parameter drift. Fifth, the system lacks proactive learning capabilities when rare faults occur. After manual intervention, knowledge cannot be automatically accumulated, making it difficult to form a closed loop of continuous evolution. Therefore, there is an urgent need for a grid-connected integrated detection system that can deeply integrate modeling and detection, take into account both privacy protection and collaborative evolution, and possess interpretability and self-learning capabilities. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of existing technologies, this invention provides a refined modeling and grid-connected integrated testing system for new energy power stations, which solves the problems of modeling and testing being separated, pure data-driven systems being difficult to trace causes, the contradiction between privacy and sharing, the inability of models to self-evolve, and the weak ability to learn new faults in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A refined modeling and grid-connection integrated testing system for new energy power plants includes: The edge layer, deployed at various sites, is used to collect data and perform on-site detection based on the ensemble classifier, outputting results and confidence levels; The regional layer communicates with multiple edge layers and is used for causal weighted aggregation and privacy protection of parameters of similar site models based on federated learning. The cloud layer communicates with multiple regional layers and is used to build a global causal graph, maintain a knowledge graph, and run a digital twin sandbox. The edge layer includes a data-driven base classifier, a knowledge-driven base classifier, a fusion unit, and a lightweight Siamese model; the fusion unit dynamically calculates weights and outputs fault types and fusion confidence based on the similarity between the input and the training set distribution, the matching degree with the global causal graph, and the prior probability of the fault. The regional layer includes a supernet management unit, a causal weighted aggregation unit, a privacy protection unit, and a causal distillation unit; The cloud layer includes a structural causal graph engine, a causal verification unit, a causal embedding unit, a digital twin sandbox and intervention simulator, and a knowledge graph inference engine; The edge layer also includes a prototype network sampler, which is used to extract abnormal prototypes to match the global causal graph when the fusion confidence is lower than the threshold and the diagnostic results are inconsistent. If there is no match, a node to be verified is created and an exploratory federated learning task is initiated.

[0006] Preferably, the dynamic weights in the fusion unit are calculated as follows: data-driven weights are proportional to the product of the increasing function of similarity and the increasing function of fusion confidence; knowledge-driven weights are proportional to the product of the increasing function of matching degree and the increasing function of prior probability; the final fusion probability is the product of data-driven weights and data-driven output probability plus the product of knowledge-driven weights and knowledge-driven output probability; the fusion confidence is the maximum fusion probability multiplied by the difference, which is the number 1 minus the fusion probability entropy; the fusion probability entropy refers to the Shannon entropy of the overall distribution of fusion probabilities corresponding to all fault categories, and the fusion probability entropy is calculated based on the overall fusion probability distribution of all fault categories. When calculating the fusion confidence score using the calculated Shannon entropy, the order of operations is as follows: first, calculate the fusion probability entropy completely; then, subtract this entropy value from the constant 1 to obtain the intermediate difference; finally, multiply this difference by the maximum fusion probability to obtain the fusion confidence score. Data confidence score is defined as the maximum value of the data-driven output probability, and knowledge confidence score is defined as the maximum value of the knowledge-driven output probability. The above-mentioned increasing functions are monotonically increasing functions, and the optional types include linear increasing functions, exponential increasing functions, and sigmoid increasing functions. The output of a linear increasing function increases proportionally with the input, the output of an exponential increasing function increases rapidly with the increase of the input, and the output of a sigmoid increasing function has upper and lower limits, with the intermediate range rising rapidly with the input.

[0007] Preferably, the weighted average formula in the causal weighted aggregation unit is: the aggregation parameter is equal to the sum of the products of the parameters of each station in the same group and their weights, divided by the sum of the weights, where the weight is the square root of the data volume multiplied by the causal consistency factor multiplied by the average confidence level. The difference between the two causal graphs is quantified by the standardized graph edit distance. The graph edit distance is uniformly set as the cost of graph modification. The total number of node, causal edge addition, deletion and flip operations required to adjust the local causal graph of a station into a typical causal graph in the group is counted. The statistical results are normalized and converted into values ​​in the range of 0 to 1. The higher the value, the greater the difference in the topological structure of the two causal graphs. The typical causal graph within a group is generated according to the following steps: First, the regional layer collects local causal graphs uploaded by all stations within the same group; second, causal features are extracted from all local causal graphs and clustering is performed, with cluster centers selected as the base graph; third, the frequency of occurrence of each causal node and causal edge in the stations within the group is counted, and accidental causal branches with a frequency below a preset threshold are removed; fourth, the causal association strength of the retained causal edges is weighted and averaged according to the frequency of occurrence, and finally a unified typical causal graph for the group is generated and stored in the regional layer database. The causal consistency factor is a decreasing function of the difference in graph edit distance between the station's local causal graph and the typical causal graph within the group. The graph edit distance is a standardized metric used to characterize the difference in topological structure between the two causal graphs. When the causal consistency factor is less than a preset isolation threshold, the corresponding station is temporarily isolated. The aforementioned decreasing function is a monotonically decreasing function, and the optional types include linear decreasing functions and exponential decay functions. The output of a linear decreasing function decreases proportionally with the input, while the output of an exponential decay function gradually slows down as the input increases.

[0008] Preferably, after creating a node to be verified, the prototype network sampler requests unlabeled segments from all edge layers whose spatiotemporal similarity to the abnormal prototype exceeds a similarity threshold. It collects similar patterns through federated clustering. When similar segments from a preset minimum number of stations are collected, the causal verification unit is triggered to perform verification. After the verification is passed, a new node and corresponding causal edge are permanently added to the global causal graph.

[0009] Preferably, the lightweight twin model is constructed using a heterogeneous graph model compressed by graph pooling. The heterogeneous graph nodes include power generation unit nodes, merge nodes, and grid connection point nodes, and the edge types include electrical topology edges and time-series correlation edges calculated based on historical operating data. The graph pooling compresses multiple power generation unit nodes into supernodes through comparative learning and pooling operations to form an equivalent model.

[0010] Preferably, the digital twin sandbox and intervention simulator perform the following operations: when the causal tracing locates the root factor module and its parameter deviation, the corresponding parameters are corrected in the twin model and the same perturbation signal is injected. If the suppression effect reaches the preset index, the corrected parameters are packaged into a model patch and distributed to all field edge layers with causal similarity greater than the preset similarity threshold through the causal weighted aggregation unit.

[0011] Preferably, when the knowledge graph inference engine updates the rules, the confidence of the rules is updated in a Bayesian manner based on the feedback from manual review. The update formula is: the new confidence equals the old confidence multiplied by the conditional probability of correct feedback divided by the old confidence multiplied by the conditional probability of correct feedback plus the sum of the old confidence multiplied by the conditional probability of incorrect feedback; and the updated rules are synchronized to the knowledge-driven base classifiers of all edge layers.

[0012] Preferably, the privacy protection unit uses differential privacy to add noise to the gradient and employs a homomorphic encryption scheme to perform a weighted average in the ciphertext; at the same time, it adopts a gradient sparsity strategy to upload only the gradient element with the largest absolute value in a preset proportion.

[0013] Preferably, it also includes a manual intervention processing unit: when the fusion confidence is lower than the first preset threshold and both the data confidence and knowledge confidence are lower than the second preset threshold, the current fault segment and multiple causal paths deduced by the system are pushed to the operation and maintenance personnel terminal. After receiving the manual feedback results, the causal path selected by the human is used as a supervision signal for incremental Bayes update. If a new diagnostic text is manually entered, the causal triples are extracted by natural language processing as candidate edges and stored in a temporary buffer.

[0014] Preferably, the communication between the edge layer, the region layer, and the cloud layer adopts an encryption protocol; the uplink data includes de-identified fault fragment features, privacy-protected versions of local causal image segments, encrypted model gradients, and low-confidence anomaly alarms and their context; the downlink data includes global causal graph update fragments, hypernet sub-model parameters, knowledge graph rule patches, and model correction instructions.

[0015] The technical effects and advantages of the refined modeling and grid-connected integrated testing system for new energy power plants of this invention are as follows: 1. This invention employs a knowledge-data dual-driven integrated classifier. The data-driven base classifier learns patterns from historical fault samples using a deep neural network, while the knowledge-driven base classifier performs symbolic reasoning based on physical mechanism rules in a knowledge graph. The fusion unit dynamically adjusts weights based on the similarity between the input data and the training set distribution, the matching degree with the global causal graph, and the prior probability of the fault, achieving adaptive fusion of data-driven and knowledge-driven approaches. When historical samples are abundant, the data-driven weights automatically increase to leverage the advantages of statistical learning; when samples are scarce or in novel fault scenarios, the knowledge-driven weights automatically increase to compensate for insufficient data. Compared to purely data-driven or purely knowledge-driven solutions, this invention significantly improves the recognition accuracy of various faults, especially rare faults, while the calculation of fusion confidence provides a reliability assessment of the detection results.

[0016] 2. This invention constructs a global structural causal graph through a cloud-based structural causal graph engine. It utilizes a variational autoencoder to infer the strength and time delay distribution of each causal edge, forming an interpretable causal chain after Granger criterion verification. The detection system outputs not only the fault category but also the causal path leading to the fault and its confidence level, allowing maintenance personnel to directly pinpoint specific submodules and parameter deviations. This mechanism transforms black-box statistical detection into white-box causal reasoning, significantly reducing fault diagnosis time and providing a clear direction for subsequent model correction and maintenance decisions.

[0017] 3. This invention forms a complete closed loop of detection, tracing, correction, evolution, and re-detection. Once causal tracing pinpoints the root factor module and its parameter deviations, the digital twin sandbox and intervention simulator correct the corresponding parameters and simulate perturbation responses within the twin model. After successful verification, a model patch is generated and distributed to all edge layers of the site with causal similarity greater than a preset threshold via a causal weighted aggregation unit. Simultaneously, the confirmed causal chain is transformed into inference rules to update the knowledge graph and synchronized to each edge knowledge-driven base classifier. This system can automatically adapt to dynamic factors such as equipment aging, control parameter drift, and environmental changes, maintaining high-precision modeling and detection performance continuously without manual intervention.

[0018] 4. This invention adopts a three-layer architecture: edge-region-cloud. The original data from each power station is stored at the edge layer; only model gradients processed with differential privacy noise enhancement and homomorphic encryption are uploaded. The region layer performs causal weighted aggregation, with aggregation parameters in encrypted form, decryptable only by the cloud layer holding the private key. Different regions exchange causal distillation samples (causal premise vectors, fault types, and confidence levels) through causal distillation units, rather than the original data or model parameters. This mechanism protects the data privacy of each new energy power station, meets data compliance requirements, and enables cross-power station and cross-regional knowledge transfer and collaborative evolution, solving the data silo problem.

[0019] 5. This invention sets up a prototype network sampler at the edge layer. When the fusion confidence is below a threshold and the diagnostic results of multiple base classifiers are inconsistent, it extracts abnormal prototypes and matches them with the global causal graph. If there is no match, a node to be verified is created and an exploratory federated learning task is initiated. Similar unlabeled segments are requested from all edge layers, and similar patterns are collected through federated clustering. After reaching a preset minimum number of sites, causal verification is automatically triggered. If the verification passes, a new node and causal edge are permanently added. This mechanism enables the system to proactively discover, verify, and learn new fault modes with very few historical samples, significantly reducing the cost of manual annotation and improving the system's response speed to unknown faults.

[0020] 6. This invention incorporates a manual intervention processing unit. When the fusion confidence level falls below a first preset threshold, and both the data confidence level and the knowledge confidence level fall below a second preset threshold, the fault fragment and multiple causal paths derived by the system are pushed to the maintenance personnel's terminal. The manual feedback results serve two purposes: firstly, as a supervisory signal to incrementally update the prior probabilities of the fusion unit using Bayesian algorithm; and secondly, to correct the confidence levels of rules in the knowledge graph. If new diagnostic text is manually input, the system extracts causal triples through natural language processing and stores them in a temporary buffer. Once the accumulated threshold is reached, verification is automatically triggered, and the information is formally added to the knowledge graph. This mechanism effectively integrates the professional knowledge of maintenance personnel into the system, forming a continuous evolutionary closed loop from "human core" to "machine knowledge."

[0021] 7. This invention maintains a supernet containing multiple sub-models in the cloud through a regional layer supernet management unit. Each sub-model corresponds to a type of causal feature pattern and is accompanied by a causal fingerprint. When a new field station joins the network, its local causal fingerprint is extracted from short-term running data. After matching it with the fingerprints of each sub-model in the supernet, the most similar sub-model is selected as the initial model, eliminating the need for retraining from scratch or waiting for a long period of data accumulation. Combined with the cross-regional knowledge transfer provided by the causal distillation unit, new field stations can achieve stable detection performance in a very short time, significantly reducing the deployment cost and time for new field station access to the system. Attached Figure Description

[0022] Figure 1 This is a system block diagram of a refined modeling and grid-connected integrated testing system for new energy power stations proposed in this invention; Figure 2 This is a flowchart of a refined modeling and grid-connected integrated testing system for new energy power stations proposed in this invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0025] refer to Figures 1-2 This invention discloses a refined modeling and grid-connected integrated detection system for new energy power stations, aiming to solve the problems of fragmented modeling and detection, lack of interpretability, contradiction between data privacy and collaboration, inability of models to self-evolve, and weak small-sample learning ability in existing technologies. The system comprises a three-layer architecture: an edge layer, a regional layer, and a cloud layer. The edge layer, deployed at each power station, includes a data-driven base classifier, a knowledge-driven base classifier, a fusion unit, a lightweight twin model, and a prototype network sampler, used for on-site detection and outputting fusion confidence scores. The regional layer includes a supergrid management unit, a causal weighted aggregation unit, a privacy protection unit, and a causal distillation unit, used for causal weighted aggregation and privacy protection of model parameters for similar power stations based on federated learning, and for cross-regional knowledge transfer through causal distillation. The cloud layer includes a structural causal graph engine, a causal verification unit, a causal embedding unit, a digital twin sandbox and intervention simulator, and a knowledge graph inference engine, used for constructing a global causal graph, maintaining the knowledge graph, and performing twin simulations and intervention verification. The edge layer, regional layer, and cloud layer work together to form a closed loop of detection, source tracing, correction, evolution, and re-detection. A knowledge-data dual-driven integrated classifier dynamically fuses data-driven and knowledge-driven weights. Causal graphs are used to trace the root causes of faults and correct the parameters of the twin model. A prototype network sampler and exploratory federated learning enable active learning with small samples. Combined with a human intervention processing unit, a human-machine co-evolution is achieved. This invention balances detection accuracy, interpretability, privacy protection, self-evolution capabilities, and rapid deployment, significantly improving the intelligent level of comprehensive grid-connected detection for new energy power plants.

[0026] Overall system composition: The present invention discloses a refined modeling and grid-connected integrated monitoring system for new energy power plants, comprising a three-layer architecture: an edge layer deployed at each new energy power plant, a regional layer deployed at a provincial or regional new energy dispatch center, and a cloud layer deployed at a group or State Grid-level new energy monitoring platform. The edge layer and the regional layer are connected via encrypted communication links, as are the regional layer and the cloud layer.

[0027] The edge layer includes: a data-driven base classifier, a knowledge-driven base classifier, a fusion unit, a lightweight twin model, a prototype network sampler, and local storage.

[0028] The regional layer includes: supernet management unit, causal weighted aggregation unit, privacy protection unit, causal distillation unit, and regional database.

[0029] The cloud layer includes: a structural causal graph engine, a causal verification unit, a causal embedding unit, a digital twin sandbox and intervention simulator, a knowledge graph inference engine, and a global database.

[0030] Basic workflow: During system operation, the edge layer continuously collects synchronous phasor data (sampling rate 100Hz), SCADA data (1Hz), meteorological data, and fault waveform data (10kHz) from each station. Data-driven and knowledge-driven base classifiers process the input data in parallel and output preliminary classification results. The fusion unit dynamically calculates weights and outputs the final fault type and fusion confidence score based on the similarity between the current input and the historical training set distribution, the matching degree with the global causal graph, and the prior probability of the fault.

[0031] When the fusion confidence score is lower than a preset threshold (typically 0.7) and the output classes of multiple base classifiers are inconsistent, a low-confidence processing procedure is triggered: the prototype network sampler extracts abnormal prototypes and matches them with the global causal graph. If the match is successful, the sample library of the corresponding causal node is enhanced; if the match is not successful, a node to be verified is created and an exploratory federated learning task is initiated.

[0032] The regional layer is responsible for collecting model gradients from similar sites. After differential privacy noise addition and homomorphic encryption, causal weighted aggregation is performed, and the aggregation results are shared among different regions through causal distillation units. The cloud layer maintains a global causal graph and uses a variational autoencoder to infer the strength and time delay of causal edges. After Granger criterion verification, the twin model parameters are corrected and intervention is simulated in a digital twin sandbox. After successful verification, the model patch is distributed to the edge layers of all sites with causal similarity greater than a threshold. The knowledge graph inference engine then transforms the confirmed causal chains into IF-THEN rules, updates the knowledge graph, and synchronizes it to each edge knowledge-driven base classifier.

[0033] Example 1: Data-driven base classifier based on CNN-LSTM and knowledge-driven base classifier for knowledge graph reasoning.

[0034] Purpose of Implementation: This embodiment aims to provide a specific implementation method for a data-driven base classifier and a knowledge-driven base classifier, demonstrate the feasibility of the system of the present invention in the grid connection detection scenario of conventional new energy power stations, and verify the effect of the fusion unit on improving the detection accuracy.

[0035] System Implementation: This embodiment adopts the complete system architecture described above, including an edge layer, a region layer, and a cloud layer. The data-driven base classifier in the edge layer uses a hybrid structure of a one-dimensional convolutional neural network and a long short-term memory (LSM) network. The input layer receives five-dimensional time-series signals (voltage, current, active power, reactive power, and frequency) with a length of 5 seconds and a sampling rate of 100Hz. The convolutional layer contains three convolutional blocks, each consisting of a convolutional layer (3x1 kernel size, stride 2, output channels 16, 32, and 64 respectively), a batch normalization layer, and a ReLU activation function, used to extract local waveform features. Subsequently, an LSM network layer (128 hidden layer dimensions, 2 layers) is connected to output time-dependent features. Finally, a fully connected layer (from 128 dimensions to 64 dimensions and then to the number of categories C) and a Softmax layer output the probability distribution of various fault types. The hidden feature vector is taken from the hidden state of the last time step of the LSM network, including the oscillation start time, energy center frequency, and fluctuation amplitude decay rate.

[0036] The knowledge-driven base classifier is implemented based on a forward inference engine, and the knowledge base is stored in the form of production rules. The rule premises involve comparison of electrical quantity thresholds, waveform morphology similarity (using dynamic time warping distance), and causal path matching (i.e., the embedding distance between the current input temporal feature and a certain path in the causal graph is less than a threshold). The inference engine uses the Rete algorithm for pattern matching and outputs the confidence distribution of each fault category.

[0037] The weights in the fusion unit are calculated as follows: data-driven weights are proportional to the product of the increasing function of the similarity between the current input and the training set distribution and the increasing function of the fusion confidence (minus one). Knowledge-driven weights are proportional to the product of the increasing function of the matching degree between the current input and the global causal graph and the increasing function of the prior probability of the fault. Each weight coefficient is optimized on the validation set through grid search, with the following specific values: similarity coefficient 0.8, fusion confidence coefficient 0.5, matching degree coefficient 1.2, and prior probability coefficient 0.3. The fusion confidence is defined as the product of the maximum fusion probability and the entropy of the fusion probability (minus one).

[0038] The lightweight twin model employs a heterogeneous graph equivalent model based on graph pooling. The heterogeneous graph nodes include power generation unit nodes (wind turbines or photovoltaic inverters), combiner nodes, and grid connection point nodes. Edge types include electrical topology edges and temporally correlated edges (determined by historical data with a Pearson correlation coefficient greater than 0.6). A differentiable graph pooling method is used to compress multiple power generation unit nodes into supernodes, forming the equivalent model.

[0039] Within the regional layer, the supernet management unit maintains eight sub-models, each corresponding to a different causal feature pattern. The causal weighted aggregation unit executes every 24 hours, with the aggregation weight determined by the product of the square root of the data volume, the causal consistency factor, and the average confidence level. The privacy protection unit employs differential privacy (privacy budget parameter set to 0.5, failure probability parameter set to 1e-5) and homomorphic encryption. The causal distillation unit executes every 7 days, exchanging causal distillation samples.

[0040] In the cloud layer, the structural causal graph engine is based on graph database storage. Causal testing units employ multivariate Granger calculus (significance level 0.05). The digital twin sandbox is based on an electromagnetic transient simulation engine. The knowledge graph inference engine uses a resource description framework for storage and a semantic web rule language.

[0041] Implementation Results: This embodiment underwent a three-month field test at three wind farms and two photovoltaic power stations in a provincial power grid. A total of 287 fault events were recorded during the test, including voltage sags, frequency exceedances, subsynchronous oscillations, and harmonic exceedances. Results showed that the average accuracy of the data-driven base classifier used alone was 91.2%, the average accuracy of the knowledge-driven base classifier used alone was 83.5%, and the fusion unit improved the accuracy to 94.8%. For faults with scarce samples, such as subsynchronous oscillations, the accuracy of the fusion unit (89.2%) was 26.7 percentage points higher than that of the pure data-driven model (62.5%). The lightweight twin model kept the computation time for a single fault detection within 150 milliseconds, meeting real-time requirements. After differential privacy noise addition during the federated aggregation process, the risk of privacy leakage under reconstruction attacks approached zero.

[0042] Example 2: Data-driven base classifier based on Transformer and knowledge-driven base classifier based on Bayesian network.

[0043] Purpose of implementation: This embodiment aims to demonstrate that the technical solution of the present invention does not depend on a specific neural network structure or inference algorithm. By replacing the specific implementation of the data-driven base classifier and the knowledge-driven base classifier, the universality and scalability of the system are verified.

[0044] System Implementation: This embodiment adopts the same three-layer system architecture as Embodiment 1, only replacing the internal algorithms of the data-driven base classifier and the knowledge-driven base classifier. The data-driven base classifier uses a time-series Transformer model. The input sequence length is 1000 (corresponding to 10 seconds, 100Hz sampling), the embedding dimension is 256, and learnable positional encoding is used. The encoder consists of 6 Transformer layers, each containing a multi-head self-attention network (8 heads) and a feedforward network (2048 dimensions). The output takes the embedding corresponding to the classification label as the sequence representation, followed by the classification head. During training, label smoothing (smoothing coefficient 0.1) and the AdamW optimizer are used, with a learning rate of 1e-4 and a batch size of 64.

[0045] The knowledge-driven base classifier uses a Bayesian network, with nodes representing fault symptom variables (such as frequency change rate, voltage drop depth, harmonic distortion rate, etc., totaling 20 variables), and edges learned through domain expert knowledge combined with the K2 algorithm. During inference, the connection tree algorithm is used to calculate the posterior probability of each fault type under given observations. Other modules (fusion unit, Siamese model, units in the regional layer, and units in the cloud layer) remain consistent with Example 1, except that the fusion weight coefficients are adjusted as follows: similarity coefficient 1.0, fusion confidence coefficient 0.6, matching coefficient 1.0, and prior probability coefficient 0.4.

[0046] Implementation Results: On the same test dataset (the same 5 sites and 287 failure events as in Example 1), the Transformer model demonstrated advantages in handling long sequence dependencies. The data-driven base classifier achieved an accuracy of 93.6%, 2.4 percentage points higher than the CNN-LSTM in Example 1. The Bayesian network knowledge-driven base classifier achieved an accuracy of 85.2%, slightly higher than the rule-based inference machine in Example 1. The fusion unit achieved an accuracy of 96.1%, a further improvement of 1.3 percentage points compared to Example 1. Notably, the Transformer model has a slightly higher computational cost (approximately 210 milliseconds per inference), but still meets the 300-millisecond real-time requirement. This embodiment verifies the compatibility of the system with algorithm replacement; any deep neural network capable of outputting failure probability distributions and any expert system capable of outputting symbolic inference results can be integrated and used.

[0047] Example 3: Cross-regional causal distillation for multiple wind farm clusters.

[0048] Purpose of implementation: This embodiment aims to demonstrate the specific effect of the regional layer causal distillation unit in achieving cross-regional knowledge transfer among multiple regional aggregators, and to solve the problem of insufficient model generalization ability caused by differences in data distribution between different regions.

[0049] System Implementation: This embodiment adopts a complete system architecture, focusing on the application of causal distillation units in the regional layer. The system is deployed in a provincial power grid, including 15 wind farms, divided into three regions according to the grid structure: the northern weak grid region (5 wind farms), the eastern strong grid region (6 wind farms), and the southern coastal high wind power penetration region (4 wind farms). Each region deploys a regional layer aggregator, and the regional aggregators exchange causal distillation samples through the causal distillation unit. The causal distillation sample is in the form of (causal premise vector, fault type, confidence level), where the causal premise vector includes 10 feature dimensions such as the phase-locked loop phase deviation rate, DC bus voltage fluctuation frequency, and grid connection point power oscillation damping ratio. The distillation loss function is the KL divergence plus a causal consistency regularization term (regularization coefficient is 0.01).

[0050] In the initial stage of the system, each region ran federated aggregation independently, and sub-models were shared only within their respective regions. After three months of operation, a new type of subsynchronous oscillation (frequency of about 8Hz) was detected in the northern weak power grid area, but due to the scarcity of samples, the local sub-model identification accuracy was only 55%.

[0051] Implementation Results: After the causal distillation unit is activated, the northern region aggregator extracts a causal distillation sample set S_N (containing 23 samples) and sends it to the eastern and southern region aggregators via a homomorphic encrypted channel. The eastern region aggregator uses S_N to perform knowledge distillation on its local sub-model. After three rounds of fine-tuning (each round using distilled samples for 10 training cycles), the sub-model's accuracy in identifying the same type of 8Hz oscillations improves from 49% to 78%. The southern region aggregator performs similar processing, improving its accuracy from 52% to 74%. Simultaneously, the northern region sub-model, using seven newly added real fault samples, improves its accuracy from 55% to 86%. The entire process does not exchange any raw data; only abstract causal knowledge is transferred between the regional aggregators. This embodiment demonstrates that causal distillation can effectively solve the cross-regional generalization problem caused by uneven sample distribution while protecting data privacy.

[0052] Example 4: Prototype network sampler and exploratory federated learning in a small sample scenario.

[0053] Purpose of implementation: This embodiment aims to demonstrate the self-learning capability of the edge layer prototype network sampler under conditions of extreme sample scarcity, and how the exploratory federated learning mechanism can automatically discover and confirm novel failure modes.

[0054] System Implementation: This embodiment employs a complete three-layer system architecture, focusing on the prototype network sampler at the edge layer and the structural causal graph engine and causal verification unit at the cloud layer. The prototype network sampler uses a relational network architecture to extract embedding vectors from fault segments (using a pre-trained stacked autoencoder with an embedding dimension of 128). Similarity comparison uses cosine distance, with a new prototype determination threshold of 0.3. The exploratory federated learning task is initiated by the cloud layer, with the query condition being a spatiotemporal feature cosine similarity greater than 0.85, and a preset minimum number of sites of 5. Federated clustering uses the K-means algorithm (with the number of clusters set to 3). Causal verification uses the multivariate Granger test (significance level 0.05).

[0055] This example uses a DC-side arc fault occurring in a photovoltaic power station as a test case. The fault is characterized by random fluctuations in the DC bus voltage with an amplitude of approximately 10%, and wideband harmonics (2kHz to 10kHz) in the AC-side current. Only two historical samples were available.

[0056] Implementation Results: During system operation, the prototype network sampler was triggered, identifying the fault segment as a new prototype p_new (with a minimum distance of 0.52 greater than 0.3 from the existing prototype library). p_new was matched with nodes in the global causal graph; the cosine similarity was less than 0.6 for all nodes. A node v_new was created in the cloud layer for verification and marked as "suspected arc fault." An exploratory federated learning task was then initiated, sending queries to all 128 field station edge layers. Within 24 hours, 23 similar segments from 9 different field stations were received. The cloud layer performed K-means clustering on these segments, with the main cluster containing 7 segments from 6 field stations (reaching the preset minimum number of field stations of 5). Granger causality testing showed that changes in the arc detector signal significantly affected the DC bus voltage after 0.5 milliseconds (p-value 0.008), confirming the causal edge "arc causes DC voltage fluctuations." The system permanently added the node "DC arc fault" and the causal edge to the global causal graph, setting the causal strength to 0.75. The knowledge graph inference engine automatically generated rules and synchronized them to all edge layers. When a similar fault occurred again at the photovoltaic power station, the knowledge-driven base classifier directly output "DC arc fault" with a confidence level of 0.80, and the overall confidence level of the fusion unit increased to 0.91. This embodiment demonstrates that, with only two historical samples, the system successfully discovered and confirmed the novel fault mode through active exploration and federated clustering, improving the identification accuracy from less than 30% to over 90%.

[0057] Example 5: Human intervention and closed-loop feedback.

[0058] Purpose of Implementation: This embodiment aims to demonstrate how the human intervention processing unit can integrate the professional knowledge of operation and maintenance personnel into the system evolution process, forming a closed-loop update mechanism of human-machine collaboration, and to demonstrate the application of Bayesian update and natural language processing in knowledge graph expansion.

[0059] System Implementation: This embodiment adopts a complete three-layer system architecture, including a human intervention processing unit (as an extension module of the edge layer or cloud layer, interacting with the knowledge graph inference engine and fusion unit). The human intervention interface is a web terminal, displaying fault waveforms and a list of causal paths derived by the system. The Bayesian update formula is: new confidence equals old confidence multiplied by the conditional probability of correct feedback divided by (old confidence multiplied by the conditional probability of correct feedback plus one minus the conditional probability of old confidence multiplied by the conditional probability of incorrect feedback), where the conditional probability of correct feedback is 0.95. Natural language processing uses a pre-trained BERT model (a fine-tuned version for the Chinese power sector) for named entity recognition and relation extraction. Automatic verification is triggered after three identical causal triples appear in the temporary buffer.

[0060] This embodiment uses a voltage sag fault at a wind farm as a test case. The system detected a fusion confidence level of 0.62, a data confidence level of 0.70 (predicted as "voltage sag"), and a knowledge confidence level of 0.40 (no matching rule), triggering manual intervention.

[0061] Implementation Results: Maintenance personnel received a push notification on their terminals, displaying the waveforms 3 seconds before and after the fault, along with the three most probable causal paths derived by the system. On-site inspection revealed a delay in the IGBT switching of the converter in wind turbine number 12. Path 1 (wind turbine 12 converter causing DC bus voltage fluctuations, leading to a 35kV bus voltage drop, and ultimately a voltage dip at the grid connection point) was confirmed as the correct cause, and this path was selected on the terminal. Upon receiving the feedback, the system: First, performed an incremental Bayesian update on the prior probabilities in the fusion unit, increasing the prior probability of "converter fault causing voltage dip" from 0.10 to 0.12; Second, increased the causal edge strength corresponding to path 1 from 0.55 to 0.70, and simultaneously updated the global causal graph; Third, triggered a digital twin sandbox simulation verification, confirming that the error between the corrected twin model output and the actual waveform decreased from 22% to 8%; Fourth, packaged the corrected converter model parameters into a patch and distributed it to the edge layer of the wind farm and the other three wind farms with the same type of converter.

[0062] If maintenance personnel input a new diagnostic text, "IGBT driver board capacitor aging in wind turbine 12," the system extracts a causal triple ("IGBT driver board capacitor aging," "cause," "converter switching delay") using the BERT model and stores it in a temporary buffer. When this triple appears a total of 5 times (from different sites) within the following 3 months, a Granger criterion is automatically triggered, and the information is formally added to the knowledge graph. This embodiment demonstrates that manual intervention not only solves the problem of single fault diagnosis but also drives the continuous evolution of the system's knowledge base, achieving a closed loop from "human-centric" to "machine-centric" knowledge.

[0063] Comparative Example 1: Traditional centralized data-driven grid-connected detection system.

[0064] Purpose of this implementation: This comparative example aims to highlight the technical advantages of the present invention in terms of detection accuracy (especially rare faults), interpretability, model update efficiency, privacy protection, and adaptability to new sites by comparing it with traditional technical solutions.

[0065] The comparison system adopts a traditional centralized architecture without a distinction between the edge, region, and cloud layers. All site data is centrally uploaded to a single central server. The detection model is a purely data-driven CNN-LSTM classifier (with the same structure as the data-driven base classifier in Example 1, but without a knowledge-driven component and fusion unit). Model training uses centralized batch gradient descent, with manual retraining every three months. The system lacks causal attribution and Siamese correction modules, federated learning and privacy protection mechanisms, and few-shot self-learning modules. Manual intervention is limited to post-event recording and does not automatically update the model.

[0066] Experimental setup: Three years of actual operation data from a provincial power grid were used, including 15 substations and a total of 1587 fault events. Common faults (voltage sags, frequency exceedances) accounted for 85%, while rare faults (subsynchronous oscillations, DC arcs) accounted for 15%. The comparative set used data from the first two years for centralized training, while the test set used data from the third year. The system of this invention was deployed according to Example 1, with data from each substation remaining within the substation, and federated learning was used for training.

[0067] Implementation Results: The evaluation results are as follows: The comparative example achieved an accuracy rate of 94.2% in identifying common faults, while Example 1 of this invention achieved 96.8%; the comparative example achieved an accuracy rate of 61.5% in identifying rare faults, while this invention achieved 89.3%; the comparative example's average detection latency (from fault occurrence to alarm) was 320 milliseconds, while this invention achieved 180 milliseconds; the comparative example could not perform fault root cause localization, while this invention achieved a fault root cause localization accuracy rate of 87.6%; the comparative example's model update cycle was 3 months (manual), while this invention's was 1 hour (automatic); the comparative example had no privacy protection (raw data was directly uploaded), while this invention used differential privacy and homomorphic encryption; the comparative example required 30 days to achieve stable performance after a new site was connected (sufficient samples needed to be collected), while this invention achieved 2 days (through supernet matching and causal distillation).

[0068] The model performs reasonably well on common faults, but its accuracy is only 61.5% on rare faults due to the scarcity of training samples, and it provides no interpretable information. Model updates rely on manual intervention, with a cycle of up to 3 months, making it unable to adapt to aging equipment and new types of faults. Centralized data uploads pose a privacy risk. After a new site is connected, it takes 30 days to accumulate samples before reaching usable performance.

[0069] In comparison, this invention improves accuracy by 27.8 percentage points on rare faults through knowledge-data dual-drive fusion; achieves 87.6% root cause localization accuracy through causal attribution; enables hourly automatic model updates through federated learning; protects data privacy through differential privacy and homomorphic encryption; and enables new sites to reach stable performance within 2 days through supernetting and causal distillation. The overall performance is significantly better than the comparative example.

[0070] Compared to Examples 1-5 and Comparative Example 1, Examples 1-5 of this invention are all based on the three-layer architecture (edge ​​layer, regional layer, and cloud layer) proposed in this invention and adopt a knowledge-data dual-drive fusion strategy, while Comparative Example 1 is a traditional centralized pure data-driven system. Regarding detection accuracy, Example 1 improves the accuracy of common fault identification to 96.8% and rare fault identification to 89.3% through the fusion unit, while the rare fault accuracy of Comparative Example 1 is only 61.5%, mainly because it relies on historical label samples and lacks knowledge guidance. Regarding interpretability, Examples 1-5 can output the root cause causal chain of the fault through the structural causal graph engine and causal verification unit, achieving a source tracing accuracy of 87.6%, while Comparative Example 1 only outputs category labels and cannot explain the cause of the fault. Regarding model update efficiency, Examples 1-5 achieve hourly automatic updates through federated learning, while Comparative Example 1 requires manual retraining every 3 months. Regarding privacy protection, Examples 1-5 use differential privacy and homomorphic encryption, and the original data does not leave the site, while Comparative Example 1 uploads all original data centrally, posing a risk of leakage. When a new site is connected, Examples 1 to 5 use the supernet management unit to match causal fingerprints and combine them with the causal distillation unit to achieve stable performance within 2 days; Comparative Example 1 requires 30 days of sample accumulation to train a usable model.

[0071] Differentiated contributions and overall effects of each embodiment: Embodiment 1 uses a CNN-LSTM data-driven base classifier and a rule-based reasoning knowledge-driven base classifier to verify the effectiveness of dynamic weight adjustment of the fusion unit, improving accuracy by 26.7 percentage points on faults with scarce samples, such as subsynchronous oscillations. Embodiment 2 replaces the data-driven base classifier with a Transformer and the knowledge-driven base classifier with a Bayesian network, proving that the invention does not rely on a specific algorithm, while further improving the accuracy to 96.1%, demonstrating the system's versatility and scalability. Embodiment 3 focuses on the causal distillation unit, demonstrating cross-regional knowledge transfer capabilities—when samples are scarce in the northern region, exchanging causal distillation samples improves the recognition accuracy of novel oscillations in the eastern and southern regions from 49% and 52% to 78% and 74%, respectively, without exposing the original data. Example 4 utilizes a prototype network sampler and exploratory federated learning to actively collect similar fragments from six power stations in a DC arc fault scenario with only two historical samples. After Granger criterion verification, new rules are automatically generated, improving the recognition accuracy from less than 30% to over 90%, thus solving the problem of small sample learning. Example 5 introduces a human intervention processing unit, integrating feedback from maintenance personnel into the knowledge graph through Bayesian updates and natural language processing, forming a human-machine collaborative closed loop. In summary, this invention significantly outperforms Comparative Example 1, achieving breakthrough improvements in detection accuracy, interpretability, model evolution speed, privacy protection, and adaptability.

[0072] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0073] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A refined modeling and grid-connection integrated testing system for new energy power stations, characterized in that, include: The edge layer, deployed at various sites, is used to collect data and perform on-site detection based on the ensemble classifier, outputting results and confidence levels; The regional layer communicates with multiple edge layers and is used for causal weighted aggregation and privacy protection of parameters of similar site models based on federated learning. The cloud layer communicates with multiple regional layers and is used to build a global causal graph, maintain a knowledge graph, and run a digital twin sandbox. The edge layer includes a data-driven base classifier, a knowledge-driven base classifier, a fusion unit, and a lightweight Siamese model; the fusion unit dynamically calculates weights and outputs fault types and fusion confidence based on the similarity between the input and the training set distribution, the matching degree with the global causal graph, and the prior probability of the fault. The regional layer includes a supernet management unit, a causal weighted aggregation unit, a privacy protection unit, and a causal distillation unit; The cloud layer includes a structural causal graph engine, a causal verification unit, a causal embedding unit, a digital twin sandbox and intervention simulator, and a knowledge graph inference engine; The edge layer also includes a prototype network sampler, which is used to extract abnormal prototypes to match the global causal graph when the fusion confidence is lower than the threshold and the diagnostic results are inconsistent. If there is no match, a node to be verified is created and an exploratory federated learning task is initiated. The dynamic weights in the fusion unit are calculated as follows: data-driven weights are proportional to the product of the similarity's increasing function and the fusion confidence's increasing function; knowledge-driven weights are proportional to the product of the matching degree's increasing function and the prior probability's increasing function; the final fusion probability is the product of the data-driven weights and the data-driven output probability plus the product of the knowledge-driven weights and the knowledge-driven output probability; the fusion confidence is the maximum fusion probability multiplied by the difference, which is the number 1 minus the fusion probability entropy. Fusion probability entropy refers to the Shannon entropy of the overall distribution of fusion probabilities corresponding to all fault categories; data confidence is defined as the maximum value of data-driven output probability, and knowledge confidence is defined as the maximum value of knowledge-driven output probability. The above increasing functions include linear increasing functions, exponential increasing functions, or Sigmoid increasing functions. The weighted average formula in the causal weighted aggregation unit is as follows: the aggregation parameter is equal to the sum of the products of the parameters of each station in the same group and their weights, divided by the sum of the weights. The weight is the square root of the local training sample data volume of the station multiplied by the causal consistency factor multiplied by the average confidence level. The causal consistency factor is a decreasing function of the difference in graph edit distance between the local causal graph of the station and the typical causal graph in the group. The graph edit distance is a standardized metric used to characterize the difference in the topological structure of the two causal graphs. When the causal consistency factor is less than the preset isolation threshold, the corresponding station is temporarily isolated. The above decreasing function includes a linear decreasing function or an exponential decay function.

2. The refined modeling and grid-connection integrated testing system for new energy power stations as described in claim 1, characterized in that, After creating a node to be verified, the prototype network sampler requests unlabeled segments from all edge layers whose spatiotemporal similarity to the abnormal prototype exceeds a similarity threshold. It collects similar patterns through federated clustering. When similar segments from a preset minimum number of stations are collected, the causal verification unit is triggered to perform the verification. After the verification is passed, a new node and corresponding causal edge are permanently added to the global causal graph.

3. The refined modeling and grid-connection integrated testing system for new energy power stations as described in claim 1, characterized in that, The lightweight twin model is constructed using a heterogeneous graph model compressed by graph pooling. The heterogeneous graph model includes power generation unit nodes, confluence nodes, and grid connection point nodes. The edge types include electrical topology edges and time-series correlation edges calculated based on historical operating data. Graph pooling compresses multiple power generation unit nodes into supernodes through comparative learning and pooling operations to form an equivalent model.

4. The refined modeling and grid-connection integrated testing system for new energy power stations as described in claim 1, characterized in that, The digital twin sandbox and intervention simulator perform the following operations: when the root factor module and its parameter deviation are located by causal tracing, the corresponding parameters are corrected in the twin model and the same perturbation signal is injected. If the suppression effect reaches the preset index, the corrected parameters are packaged into a model patch and distributed to all field edge layers with causal similarity greater than the preset similarity threshold through the causal weighted aggregation unit.

5. The refined modeling and grid-connection integrated testing system for new energy power stations as described in claim 1, characterized in that, When the knowledge graph inference engine updates the rules, it updates the confidence of the rules in a Bayesian manner based on the feedback from manual review. The update formula is: the new confidence equals the old confidence multiplied by the conditional probability of correct feedback divided by the old confidence multiplied by the conditional probability of correct feedback plus the sum of the old confidence multiplied by the conditional probability of incorrect feedback; and the updated rules are synchronized to the knowledge-driven base classifiers of all edge layers.

6. The refined modeling and grid-connection integrated testing system for new energy power stations as described in claim 1, characterized in that, The privacy protection unit uses differential privacy to add noise to the gradient and employs a homomorphic encryption scheme to perform a weighted average in the ciphertext; at the same time, it adopts a gradient sparsity strategy to upload only the gradient element with the largest absolute value in a preset proportion.

7. The refined modeling and grid-connection integrated testing system for new energy power stations as described in claim 1, characterized in that, It also includes a human intervention processing unit: when the fusion confidence is lower than the first preset threshold and both the data confidence and knowledge confidence are lower than the second preset threshold, the current fault segment and multiple causal paths deduced by the system are pushed to the operation and maintenance personnel terminal. After receiving the human feedback results, the human-selected causal path is used as a supervision signal for incremental Bayes update. If a new diagnostic text is manually entered, causal triples are extracted through natural language processing as candidate edges and stored in a temporary buffer.

8. The refined modeling and grid connection integrated testing system for new energy power stations as described in claim 1, characterized in that, The communication between the edge layer, the region layer and the cloud layer adopts an encryption protocol; the uplink data includes de-identified fault fragment features, privacy-protected versions of local causal image segments, encrypted model gradients, and low-confidence anomaly alarms and their context; the downlink data includes global causal graph update fragments, hypernet sub-model parameters, knowledge graph rule patches and model correction instructions.

Images