Intelligent power grid fault diagnosis system and method based on AI self-adaption

Abstract

The invention belongs to the technical field of intelligent operation and maintenance of a power system, and particularly relates to an intelligent power grid fault diagnosis system and method based on AI self-adaption, and the system comprises a data collection module, a preprocessing module, a self-adaption AI diagnosis engine, a fault recognition and classification module, a fault positioning module, a decision support and feedback module, and a system self-learning module. Dynamic adjustment of a diagnosis strategy is realized through deep reinforcement learning, topology modeling and fault positioning are performed in combination with a graph neural network, and continuous optimization of a model is realized by adopting an incremental learning mechanism, so that the accuracy, real-time performance and adaptability of fault diagnosis of the smart grid are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and more specifically, to an AI-adaptive smart grid fault diagnosis system and method. Background Technology

[0002] With the continuous optimization of the energy structure and the intelligent development of the power system, smart grids have become an important development direction for modern power systems. Smart grids achieve real-time monitoring and dynamic management of the power grid's operating status through highly integrated communication, sensing, and control technologies. However, with the expansion of the power grid's scale and the increasing complexity of its structure, traditional fault diagnosis methods relying on manual experience and fixed rules are no longer sufficient to meet the demand for rapid and accurate fault identification. Power grid faults are diverse, including short circuits, open circuits, overloads, and equipment aging, and their occurrence is sudden and uncertain, placing higher demands on the real-time performance, adaptability, and intelligence level of fault diagnosis systems. Therefore, there is an urgent need for an intelligent fault diagnosis system that can adapt to changes in the power grid's operating status and possesses high-precision diagnostic capabilities to improve the safety and stability of power grid operation. Shortcomings of existing technologies:

[0003] 1. Traditional diagnostic methods rely on fixed rules and lack adaptability.

[0004] Currently, many power grid fault diagnosis systems still rely on preset rules and expert experience, such as setting voltage and current thresholds to determine whether a fault has occurred. This static diagnostic mechanism often fails to adjust its diagnostic strategy in a timely manner when faced with frequent changes in power grid operating conditions, large load fluctuations, or the emergence of new fault modes, resulting in a high misjudgment rate and delayed response, making it difficult to meet the needs of modern smart grids for efficient and accurate fault diagnosis.

[0005] 2. Existing AI models lack online learning capabilities, making continuous optimization difficult.

[0006] While some systems have attempted to incorporate artificial intelligence (AI) technology for fault identification, most models are trained offline, meaning they are trained on a fixed dataset and then deployed online. This makes it impossible to dynamically adjust model parameters based on changes in the power grid's operating status and the emergence of new fault characteristics. Consequently, the system's diagnostic performance deteriorates after a period of operation, making it difficult to adapt to power grid structural updates or the emergence of new fault modes.

[0007] 3. The fault location accuracy is not high, making it difficult to deal with multi-point faults in complex topology structures.

[0008] With the increasing complexity of power grid structures, the possibility of multiple faults occurring simultaneously or in a chain reaction is increasing. However, existing fault location methods mostly rely on traditional electrical quantity analysis or simple topology matching, lacking the ability to deeply model and analyze the overall power grid structure. This makes it difficult to accurately locate fault points in the case of multiple nodes and branches, affecting fault handling efficiency and power grid recovery speed.

[0009] Therefore, to address the above problems, an AI-adaptive smart grid fault diagnosis system and method are proposed. Summary of the Invention

[0010] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an AI-adaptive smart grid fault diagnosis system and method to address the problems mentioned in the background section.

[0011] To achieve the above objectives, the present invention provides the following technical solution: an AI-adaptive smart grid fault diagnosis system, comprising a data acquisition module, a data preprocessing module, an adaptive AI diagnostic engine module, a fault identification and classification module, a fault location module, a decision support and feedback module, and a system self-learning module. The data acquisition module is used to collect voltage, current, frequency, and power data of each node in the smart grid in real time. The data preprocessing module is used to filter, normalize, and process outliers in the collected data, and extract time-series features. The adaptive AI diagnostic engine module employs a deep reinforcement learning model with online learning capabilities. This system can dynamically adjust its diagnostic strategies based on the power grid's operating status. The fault identification and classification module identifies common fault types such as short circuits, open circuits, overloads, and equipment aging based on a multi-layer neural network. The fault location module locates fault points based on a graph neural network combined with the power grid topology. The decision support and feedback module generates fault handling suggestions and feeds them back to the operation and maintenance system. It also has a multi-level alarm mechanism. The system self-learning module adopts an incremental learning mechanism, automatically updating model parameters after each diagnostic result is compared with the manual confirmation result to improve the accuracy of subsequent diagnoses. This enables efficient, adaptive, and closed-loop diagnosis and handling of smart grid faults.

[0012] Furthermore, the adaptive AI diagnostic engine module employs a Deep Q-Network (DQN) with online learning capabilities. This network structure includes an input layer, multiple hidden layers, and an output layer. The input layer receives preprocessed power grid operation data, the hidden layers extract high-order features through activation functions, and the output layer outputs the optimal diagnostic action for the current state. By receiving real-time data on changes in power grid operation status, the Deep Q-Network dynamically adjusts the fault diagnosis model parameters, thereby achieving adaptive diagnostic capabilities under different operating conditions and improving the system's fault identification accuracy and response speed in complex environments.

[0013] Furthermore, the fault location module incorporates a graph neural network (GNN) to model the power grid topology, treating nodes and lines as nodes and edges in a graph. Topological features are extracted through graph convolution operations, and a weighted adjacency matrix is ​​constructed by combining the fault current direction and amplitude information. This enables high-precision fault path identification in complex power grid structures, making it particularly suitable for scenarios with parallel faults at multiple nodes. Simultaneously, this module supports real-time data interaction with the SCADA system, improving the timeliness and accuracy of fault location.

[0014] Furthermore, the system's self-learning module adopts an incremental learning mechanism. After the system completes a full fault diagnosis process, it compares the diagnosis results with the actual processing results reported by the maintenance personnel. If there is a discrepancy, it automatically extracts the features and label information of the sample and performs local model training only on the newly added data. This avoids repeatedly training historical data, reduces the consumption of computing resources, and enables the AI ​​diagnostic engine to continuously optimize through continuous updates of model parameters, adapting to changes in the power grid structure and the emergence of new fault modes.

[0015] Furthermore, the data preprocessing module includes a noise filtering unit based on wavelet transform and a time series feature extraction unit based on sliding window. Wavelet transform is used to remove high-frequency noise from the acquired signal and improve data quality, while the sliding window mechanism is used to extract time-domain features of signals such as voltage and current, such as mean, variance, and peak value, thereby providing high-quality input data for the subsequent fault identification and classification module and enhancing the robustness and generalization ability of the diagnostic model.

[0016] Furthermore, the decision support and feedback module has a multi-level alarm mechanism, which automatically classifies the warning level according to the fault type and severity. Level 1 alarms correspond to serious faults that need to be dealt with immediately, Level 2 alarms are potential risks that need to be paid attention to, and Level 3 alarms are minor anomalies that can be dealt with later. The system pushes alarm information to the dispatch center and relevant maintenance personnel's terminal devices through the communication interface module, and records the fault handling process and feeds it back to the system self-learning module for model parameter updates and system optimization, forming a complete fault diagnosis and handling closed loop.

[0017] An AI-adaptive smart grid fault diagnosis method includes the following steps: First, a data acquisition module collects real-time power grid operation data. Then, a data preprocessing module filters, normalizes, and extracts features from the collected data. Next, an adaptive AI diagnostic engine determines whether a fault has occurred. If a fault is detected, a fault identification and classification module is invoked to identify the fault type. Subsequently, a graph neural network combined with the power grid topology is used to locate the fault point, generate fault handling suggestions, and feed them back to the dispatching system via a communication module. Finally, the diagnostic results are compared with the actual handling results, and an incremental learning mechanism is used to update the AI ​​model parameters to achieve closed-loop control and continuous optimization of the system. The entire process has a high degree of automation, intelligence, and adaptability, significantly improving the efficiency and accuracy of smart grid fault diagnosis.

[0018] The technical effects and advantages of this invention are as follows:

[0019] Compared with existing technologies, this invention proposes an AI-adaptive smart grid fault diagnosis system and method. By constructing a complete technical architecture including modules such as data acquisition, preprocessing, an adaptive AI diagnostic engine, fault identification and classification, fault location, decision support and feedback, and system self-learning, it achieves real-time perception of the smart grid's operating status and rapid fault identification and processing. The system employs a deep reinforcement learning model, which can dynamically adjust diagnostic strategies according to changes in the grid operating environment, improving the adaptability and accuracy of diagnosis. Simultaneously, it introduces graph neural network technology, combining grid topology to model and analyze fault paths, effectively improving the location accuracy of multi-point faults under complex structures. Furthermore, the system integrates an incremental learning mechanism, enabling the AI ​​model to perform local optimization only on new samples without repeatedly training on historical data, thereby achieving continuous model evolution and performance improvement. Through efficient collaboration between modules, the entire system forms a closed-loop process from data acquisition, feature extraction, intelligent diagnosis, result feedback to model optimization, significantly enhancing the intelligence level, real-time response capability, and long-term adaptability of power grid fault diagnosis. Attached Figure Description

[0020] Figure 1 This is a system framework diagram of the present invention.

[0021] Figure 2 This is a flowchart of the process of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Embodiment 1: Real-time Adaptive Diagnostic Process Based on Deep Reinforcement Learning

[0023] Applicable scenarios: Scenarios where the power grid has complex operating conditions, frequent load changes, and diverse fault types, requiring real-time dynamic adjustment of diagnostic strategies.

[0024] Process description:

[0025] This implementation process utilizes a deep reinforcement learning (DQN) model to achieve real-time adaptive diagnosis of power grid faults. This model can automatically adjust the diagnostic strategy based on changes in the power grid's operating status, thereby improving diagnostic accuracy and response speed.

[0026] Step 1: Collect power grid operation data.

[0027] The system first collects operating parameters such as voltage, current, frequency, and power at various nodes in the power grid in real time through a data acquisition module. This data is acquired by sensors installed at substations, line nodes, and smart meters, and transmitted to the data processing module via a communication network. The data acquisition frequency is once per second to ensure the real-time performance of the diagnostic system.

[0028] Step 2: Preprocess the collected data.

[0029] The collected raw data often contains noise and outliers, thus requiring preprocessing. The preprocessing module uses wavelet transform to remove high-frequency noise from the signal, and then normalizes the data to ensure that the data from different nodes have uniform dimensions. Next, the sliding window technique is used to extract time series features, such as voltage fluctuation rate, mean current, and power change trend, as input features for subsequent diagnostic models.

[0030] Step 3: Construct a deep Q-network model.

[0031] The system constructs a Deep Q-Network (DQN), which consists of an input layer, multiple hidden layers, and an output layer. The input layer receives preprocessed power grid operation characteristic data, the hidden layers extract high-order features through activation functions, and the output layer outputs the optimal diagnostic action for the current state (e.g., normal operation, short circuit, overload, etc.). The model uses an experience replay mechanism to store historical state-action-reward data to avoid overfitting the model to the current data.

[0032] Step 4: Set up the state space and action space.

[0033] The state space consists of the power grid's operating status, including the changing trends of parameters such as voltage, current, and power. The action space contains possible diagnostic results, such as "normal," "short circuit," "open circuit," "overload," and "equipment aging." The reward function is set based on the diagnostic results and actual operation and maintenance feedback; a positive reward is given if the diagnosis is correct, and a negative reward is given if it is incorrect.

[0034] Step 5: Online training and strategy updates.

[0035] The system employs an online learning approach, receiving new power grid operation data every second and updating its model parameters accordingly. During each diagnostic process, the model selects the optimal action (i.e., the diagnostic result) based on the current state and adjusts its parameters based on actual feedback. This dynamic adjustment mechanism enables the system to adapt to changes in the power grid's operating state, improving its long-term diagnostic performance.

[0036] Step Six: Generate diagnostic results and provide feedback.

[0037] After the model completes a full diagnosis, the system sends the diagnostic results to the dispatch center and the terminal devices of maintenance personnel via the communication interface module. Simultaneously, the system records the diagnostic process and results and feeds them back to the system's self-learning module for model parameter updates and system optimization.

[0038] Step 7: Incremental learning to optimize the model.

[0039] The system's self-learning module employs an incremental learning mechanism. After each diagnosis, the diagnostic result is compared with the manually confirmed result. If a discrepancy exists, the features and label information of that sample are extracted, and the model is trained only on that new data. This avoids repeatedly training on historical data, reduces computational resource consumption, and improves the model's adaptability.

[0040] Summary: This implementation process achieves adaptive diagnosis of power grid faults by constructing a deep Q-network model. It can dynamically adjust diagnostic strategies to adapt to complex operating environments, improving diagnostic efficiency and accuracy. Example 2: Multi-point fault location process based on graph neural networks.

[0041] Applicable scenarios: Scenarios with complex power grid structures, multiple parallel branches and nodes, and where precise fault location is required.

[0042] Process description:

[0043] This implementation process utilizes graph neural networks (GNNs) combined with power grid topology to achieve accurate location of multi-point faults. By abstracting power grid nodes and lines as nodes and edges in a graph, a weighted adjacency matrix is ​​constructed to identify fault paths in complex structures.

[0044] Step 1: Construct the power grid topology diagram.

[0045] The system first abstracts the nodes and lines in the power grid into a graph structure. Each node represents a substation, transformer, or load node, and each line represents the electrical connection between nodes. The relationships between nodes and lines in the topology graph are represented by an adjacency matrix, where the elements represent the connection strength between nodes.

[0046] Step 2: Collect fault current direction and amplitude data.

[0047] When a power grid fault occurs, the system acquires information on the direction and magnitude of the fault current at each node through a data acquisition module. This data is obtained via current transformers and directional relays and used as input for subsequent graph neural networks.

[0048] Step 3: Construct a weighted adjacency matrix.

[0049] In graph neural networks, each element of the adjacency matrix represents the connection weight between two nodes. The system sets weights based on factors such as line impedance, distance, and historical failure rate to construct a weighted adjacency matrix, which is used to describe the topological characteristics of the power grid.

[0050] Step 4: Design the graph neural network model.

[0051] The system constructs a Graph Convolutional Network (GCN), which extracts feature information of nodes and edges through graph convolution operations. The input layer receives power grid topology and fault current data, the hidden layer extracts topological features through graph convolution operations, and the output layer outputs the fault probability distribution of each node.

[0052] Step 5: Perform graph convolution operation.

[0053] The core of graph convolution is to aggregate information about a node and its neighbors. Each node's feature vector is weighted and averaged with the feature vectors of its neighbors to form a new feature representation. This process is repeated across multiple layers of the network, enabling the model to capture relationships between nodes at greater distances.

[0054] Step 6: Identify the fault path.

[0055] After the model outputs the failure probability of each node, the system sorts the probabilities to determine the most likely failure point. For multi-point failure scenarios, the system uses a path search algorithm to identify the associated paths between multiple failure points, thereby achieving accurate localization of complex failure structures.

[0056] Step 7: Verify integration with the SCADA system.

[0057] The system compares the location results with historical alarm data from the SCADA system to verify the accuracy of the location. If the location results match the actual situation, the case is recorded for model optimization; if they do not match, more data is retrieved to retrain the model and improve the location accuracy.

[0058] Step 8: Feedback to optimize model parameters.

[0059] The system compares each location result with the manual confirmation result. If there is a discrepancy, it extracts the features and label information of the sample and updates the model parameters through an incremental learning mechanism, so that the graph neural network has the ability to continuously optimize and adapt to changes in power grid structure and new fault modes.

[0060] In summary, this implementation process, through graph neural networks combined with the power grid topology, enables precise location of multi-point faults. It is suitable for smart grid systems with complex structures and numerous nodes, significantly improving the efficiency and accuracy of fault location.

[0061] Example 3: Model Continuous Optimization Process Based on Incremental Learning

[0062] Applicable scenarios: Scenarios where the power grid structure changes frequently and new fault modes emerge continuously, requiring the system to have continuous learning capabilities.

[0063] Process description:

[0064] This implementation process uses an incremental learning mechanism to enable the system to perform local training on new data without retraining the entire model, thereby achieving continuous model optimization and adapting to changes in power grid structure and new fault modes.

[0065] Step 1: Collect diagnostic and feedback data.

[0066] After each fault diagnosis, the system compares the diagnostic results with the feedback from manual maintenance personnel. The diagnostic results are output by the AI ​​model, while the feedback results are confirmed by maintenance personnel. The comparison results will be used as label data for model training.

[0067] Step 2: Screening for biased samples.

[0068] The system compares the diagnostic results with the feedback results one by one. If they match, the model is considered to have made a correct judgment. If they do not match, the sample is marked as a deviation sample, and its features and label information are extracted for subsequent model training.

[0069] Step 3: Construct the incremental learning dataset.

[0070] The system constructs an incremental learning dataset from all biased samples. This dataset contains only newly generated samples and excludes historical training data, thus avoiding redundant training and improving training efficiency.

[0071] Step 4: Load existing model parameters.

[0072] The system loads the parameters of the current AI diagnostic model, including the weights and biases of the neural network. These parameters serve as the initial values ​​for incremental training, ensuring that the model is fine-tuned based on its original state.

[0073] Step 5: Perform the incremental training process.

[0074] The system uses an incremental learning algorithm for local training on new samples. During training, only the model parameters related to the new samples are updated, while other parameters remain unchanged. The training objective is to minimize the prediction error for new samples while maintaining the performance of the original model.

[0075] Step 6: Verify model performance.

[0076] After incremental training is complete, the system uses a subset of historical data and new samples to validate the model. Validation metrics include accuracy, recall, and F1 score. If model performance improves, the model is updated; if performance deteriorates, the system reverts to the original model.

[0077] Step 7: Update the model and deploy it online.

[0078] Once verification is successful, the system will deploy the new model to the production environment, replacing the original model. The deployment process employs a hot update mechanism to ensure that the diagnostic system continues to run during the update process.

[0079] Step 8: Record update logs and manage versions.

[0080] The system records information such as the time of each model update, the content of the update, the number of training samples, and performance changes, and manages version updates. Operations personnel can view the model evolution process through a visual interface, facilitating later maintenance and traceability.

[0081] In summary, this implementation process achieves continuous model optimization through an incremental learning mechanism, enabling it to quickly adapt to changes in power grid structure and new fault modes without retraining all data, thereby improving the system's intelligence and adaptability.

[0082] Finally, the following points should be noted: First, in the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection", and "linkage" should be interpreted broadly, and can be mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc. are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change.

[0083] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0084] In conclusion, the above description is only a preferred embodiment of the present invention and is 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. An AI-adaptive smart grid fault diagnosis system, characterized in that, The system comprises a data acquisition module, a data preprocessing module, an adaptive AI diagnostic engine module, a fault identification and classification module, a fault location module, a decision support and feedback module, and a system self-learning module. The data acquisition module collects real-time voltage, current, frequency, and power data from various nodes in the smart grid. The data preprocessing module filters, normalizes, and processes outliers from the collected data, and extracts time-series features. The adaptive AI diagnostic engine module employs a deep reinforcement learning model with online learning capabilities, dynamically adjusting diagnostic strategies based on the grid's operating status. The fault identification and classification module identifies common fault types such as short circuits, open circuits, overloads, and equipment aging based on multi-layer neural networks. The fault location module locates fault points based on graph neural networks combined with the grid topology. The decision support and feedback module generates fault handling suggestions and feeds them back to the operation and maintenance system, while also featuring a multi-level alarm mechanism. The system self-learning module uses an incremental learning mechanism, automatically updating model parameters after each diagnostic result is compared with manually confirmed results to improve subsequent diagnostic accuracy, thereby achieving efficient, adaptive, and closed-loop diagnosis and handling of smart grid faults.

2. The AI-adaptive smart grid fault diagnosis system according to claim 1, characterized in that, The adaptive AI diagnostic engine module adopts a deep Q-network (DQN) with online learning capabilities. The network structure includes an input layer, multiple hidden layers, and an output layer. The input layer receives preprocessed power grid operation data, the hidden layers extract high-order features through activation functions, and the output layer outputs the optimal diagnostic action for the current state. The deep Q-network dynamically adjusts the fault diagnosis model parameters by receiving real-time data on changes in the power grid operation state, thereby achieving adaptive diagnostic capabilities under different operating conditions and improving the system's fault identification accuracy and response speed in complex environments.

3. The AI-adaptive smart grid fault diagnosis system according to claim 1, characterized in that, The fault location module uses a graph neural network (GNN) to model the power grid topology, treating nodes and lines as nodes and edges in a graph. It extracts topological features through graph convolution operations and constructs a weighted adjacency matrix by combining fault current direction and amplitude information. This enables high-precision fault path identification in complex power grid structures, and is particularly suitable for scenarios with parallel faults at multiple nodes. At the same time, the module supports real-time data interaction with the SCADA system, improving the timeliness and accuracy of fault location.

4. The AI-adaptive smart grid fault diagnosis system according to claim 1, characterized in that, The system's self-learning module adopts an incremental learning mechanism. After the system completes a full fault diagnosis process, it compares the diagnosis results with the actual processing results reported by the maintenance personnel. If there is a discrepancy, it automatically extracts the features and label information of the sample and only performs local model training on the newly added data to avoid repeatedly training historical data and reduce the consumption of computing resources. At the same time, through continuous updates of model parameters, the AI ​​diagnostic engine has continuous optimization capabilities to adapt to changes in the power grid structure and the emergence of new fault modes.

5. The AI-adaptive smart grid fault diagnosis system according to claim 1, characterized in that, The data preprocessing module includes a noise filtering unit based on wavelet transform and a time series feature extraction unit based on sliding window. Wavelet transform is used to remove high-frequency noise from the acquired signal and improve data quality, while the sliding window mechanism is used to extract time-domain features of signals such as voltage and current, such as mean, variance, and peak value, thereby providing high-quality input data for the subsequent fault identification and classification module and enhancing the robustness and generalization ability of the diagnostic model.

6. The AI-adaptive smart grid fault diagnosis system according to claim 1, characterized in that, The decision support and feedback module has a multi-level alarm mechanism, which automatically classifies the warning level according to the fault type and severity. Level 1 alarms correspond to serious faults that need to be dealt with immediately, Level 2 alarms are potential risks that need to be paid attention to, and Level 3 alarms are minor anomalies that can be dealt with later. The system pushes alarm information to the dispatch center and relevant maintenance personnel's terminal devices through the communication interface module, and records the fault handling process and feeds it back to the system self-learning module for model parameter updates and system optimization, forming a complete fault diagnosis and handling closed loop.

7. A smart grid fault diagnosis method based on AI adaptive technology, characterized in that, The process includes the following steps: First, the data acquisition module collects real-time power grid operation data. Then, the data preprocessing module filters, normalizes, and extracts features from the collected data. Next, the adaptive AI diagnostic engine determines whether a fault has occurred. If a fault is detected, the fault identification and classification module is invoked to identify the fault type. Subsequently, a graph neural network combined with the power grid topology is used to locate the fault point, generate fault handling suggestions, and feed them back to the dispatching system through the communication module. Finally, the diagnostic results are compared with the actual handling results, and the AI ​​model parameters are updated using an incremental learning mechanism to achieve closed-loop control and continuous optimization of the system. The entire process has a high degree of automation, intelligence, and adaptability, significantly improving the efficiency and accuracy of smart grid fault diagnosis.

Images