Intelligent power station unattended monitoring method and system based on LoRa network

Abstract

The application discloses a smart power station unattended monitoring method and system based on a LoRa network, belongs to the technical field of smart power stations, and solves the problem that the interdependence relationship between associated nodes and the failure propagation path are not considered in the evaluation of the existing method, which leads to insufficient evaluation of the risk propagation trend and the global influence range, the method comprises the following steps: performing real-time edge feature extraction on monitoring original data, performing real-time risk evaluation on the feature extraction set based on a cascade propagation model, and performing risk propagation analysis on the distributed monitoring nodes in combination with the cascade failure influence index of the associated monitoring nodes; in the application, the power station is regarded as an organic whole through the cascade propagation model, the islanded evaluation of the distributed monitoring nodes can be avoided, the risk of a single node can be evaluated, the interdependence relationship between associated nodes and the failure propagation path can be considered, and therefore the propagation path and the influence range of the predicted failure can be accurately determined.

Description

Technical Field

[0001] This invention belongs to the field of smart power plant technology, specifically relating to a method and system for unattended monitoring of smart power plants based on LoRa networks. Background Technology

[0002] With the rapid development of the energy internet, power plants (including substations, photovoltaic power plants, and wind power plants) are core components of the power system, and their operational safety and stability are crucial to the entire power grid. However, traditional power plant monitoring and management primarily rely on manual inspections, which are not only inefficient but also have numerous drawbacks. First, manual inspections require significant manpower and have long inspection cycles, making it impossible to achieve 24 / 7, all-round real-time monitoring. Furthermore, traditional communication methods are vulnerable to damage during natural disasters (such as earthquakes and floods), leading to communication outages and preventing timely acquisition of power plant operational status, thus increasing the risk to power plant operation.

[0003] LoRa (Long Range) technology, as an emerging wireless communication technology, has advantages such as long distance, low power consumption, low cost and high reliability, making it very suitable for Internet of Things (IoT) applications. LoRa networks use spread spectrum modulation technology, which enables long-distance communication in complex electromagnetic environments, and its low power consumption allows devices to be powered by batteries, greatly reducing deployment and maintenance costs.

[0004] Chinese patent CN120127845A discloses an unattended intelligent power monitoring method and system. This method preprocesses and extracts features from collected equipment operating status data, power system data, and environmental data. Based on real-time monitoring and analysis of environmental parameters and behaviors within the substation, it detects anomalies. After ruling out substation faults caused by environmental factors, it further detects and warns of abnormalities in equipment operating status and power transmission processes. When no anomalies are detected, it continues to analyze and assess the safety risks within the substation, classifying risk levels based on the assessment results and providing early warning feedback. However, existing methods conduct isolated risk assessments based on historical data from individual nodes or devices. The assessments do not consider the interdependencies between related nodes and fault propagation paths, leading to insufficient assessment of risk propagation trends and the overall impact scope. To address these issues, we propose an unattended monitoring method and system for smart substations based on LoRa networks. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for unattended monitoring of smart power plants based on LoRa networks. This solves the problem that existing methods rely on isolated assessments based on historical data from individual nodes or devices. These assessments fail to consider the interdependencies between related nodes and fault propagation paths, resulting in insufficient assessment of risk propagation trends and the overall impact scope.

[0006] This invention is implemented as follows: a method for unattended monitoring of smart power plants based on LoRa networks, the method comprising: LoRa terminals are used to collect raw monitoring data from the distributed monitoring nodes of the power plant in real time. The raw monitoring data is preprocessed and then aggregated to the edge computing gateway via the LoRa network. The edge computing gateway acquires pre-processed raw monitoring data in real time, performs real-time edge feature extraction on the raw monitoring data, obtains real-time features associated with distributed monitoring nodes, and uses a dynamic feature filtering mechanism based on principal component analysis to filter the real-time features and output a feature extraction set. The feature extraction set is obtained, and a real-time risk assessment is performed on the feature extraction set based on a pre-built cascade propagation model to obtain the node risk value and node risk attribute that characterize the comprehensive risk of the distributed monitoring nodes. The cascading failure impact index of associated monitoring nodes is calculated based on the fault propagation path analysis algorithm. The risk propagation analysis of distributed monitoring nodes is then performed based on the cascading failure impact index of associated monitoring nodes to generate the node risk propagation path corresponding to the smart power station.

[0007] Preferably, the method for preprocessing raw monitoring data includes: Based on the real-time acquisition of sensor monitoring data and environmental monitoring data from distributed monitoring nodes using distributed sensors, the sensor monitoring data and environmental monitoring data are integrated into raw monitoring data; Load the raw monitoring data, perform outlier cleaning and unified timestamp processing on the raw monitoring data to obtain raw monitoring data with unified timestamps; The original monitoring data is subjected to categorized filtering and noise reduction based on an adaptive filter to obtain the original monitoring data after filtering and noise reduction. The original monitoring data after filtering and noise reduction is obtained, correlation analysis is performed on the original monitoring data to establish the mapping relationship between the data, and the mapping relationship of the original monitoring data is marked to obtain the original monitoring data containing the mapping relationship.

[0008] Preferably, the method for performing correlation analysis on raw monitoring data includes: Load the filtered and denoised original monitoring data, identify the distributed monitoring nodes associated with the original monitoring data, and establish a data-node mapping relationship; Calculate the Pearson correlation coefficient between the original monitoring data based on the data type, determine whether the Pearson correlation coefficient exceeds the preset Pearson correlation threshold, and if the Pearson correlation coefficient exceeds the preset Pearson correlation threshold, establish a data-data mapping relationship between the original data. Based on K-means clustering analysis, typical risk patterns and monitoring raw data are clustered to establish a data-risk mapping relationship between monitoring raw data and typical risk patterns. Based on data-node mapping, data-data mapping, and data-risk mapping, related data tables are established. Time window alignment technology is used to establish associations between the monitoring raw data and the related data tables within the same time period, resulting in monitoring raw data containing the association relationships of the related data tables.

[0009] Preferably, the method for real-time edge feature extraction of raw monitoring data includes: Load the raw monitoring data and identify the data type. Based on the preset abnormal threshold of the data type, perform real-time event detection on the raw monitoring data within the time window and extract abnormal data points from the raw monitoring data. Obtain the identification results of abnormal data points in the raw monitoring data, extract the raw monitoring data within the time window before and after the abnormal data points, and obtain the abnormal data window containing the abnormal data points. A lightweight extraction network is used to perform two-level lightweight feature extraction on the raw monitoring data within the abnormal data window. The lightweight extraction network includes a parallel CSPNet structure and an LSTM layer. The CSPNet structure and the LSTM layer are connected to a lightweight decoupling head. The lightweight extraction network uses the SiLU activation function. The CSPNet structure is used to extract features from the environmental monitoring data in the raw monitoring data, and the LSTM layer is used to extract features from the sensor monitoring data in the raw monitoring data, thus obtaining real-time features associated with distributed monitoring nodes.

[0010] Preferably, the method for filtering real-time features using a dynamic feature filtering mechanism based on principal component analysis includes: Load at least one set of real-time features, and perform Z-score standardization on the real-time features to obtain standardized features. The formula for Z-score standardization of real-time features is expressed as: in, These are standardized features and real-time features, respectively. These are the characteristic standard deviation and the characteristic mean, respectively. Obtain standardized features, construct a feature covariance matrix based on the standardized features, and assign weights to the association between the standardized features and the associated data table using the mutual information method to obtain the association weights between the standardized features and the associated data table. Assign the association weights to the feature covariance matrix to obtain the weighted covariance matrix. The weighted covariance matrix is ​​expressed as: in, For the weighted covariance matrix, This is the weighted standard characteristic matrix. Indicates the number of sample features; Load the weighted covariance matrix, perform eigenvalue decomposition on the weighted covariance matrix based on principal component analysis to obtain at least one set of principal components, calculate the variance contribution rate and cumulative contribution rate of the principal components, sort the principal components based on the cumulative contribution rate of the principal components, and output the important queue of principal components. Using the association weights of standardized features and related data tables as prior conditions for the dynamic feature screening mechanism, the importance queue of principal components is traversed, and the importance of principal components is comprehensively ranked to obtain the re-ranked importance queue of principal components. The cumulative contribution rate of principal components in the important principal component queue is identified, and the dynamic feature screening mechanism combines the correlation weight and the cumulative contribution rate to dynamically define the threshold of the cumulative contribution rate. The cumulative contribution rate threshold is calculated using the following formula: in, These are the cumulative contribution rate threshold and the basic screening threshold, respectively. The association weights of the standardized features corresponding to the principal components. These represent the cumulative contribution rate of the principal components and the mean cumulative contribution rate of the principal components, respectively. The number of principal components in the principal component importance queue. This is the threshold adjustment coefficient; Select at least one principal component from the principal component importance queue that exceeds the cumulative contribution rate threshold, integrate at least one principal component and the standardized features before principal component decomposition into a feature extraction set, and output the feature extraction set.

[0011] Preferably, the cascaded propagation model is based on a graph neural network architecture. The nodes of the graph neural network are distributed monitoring nodes, and the edges represent the physical connections, communication links, and risk propagation relationships between nodes. The node features are feature extraction sets. The cascaded propagation model also includes an input layer and an output layer. Between the input and output layers are two graph convolutional layers, a node risk assessment layer, a cascaded impact analysis layer, and a path generation layer. The graph convolutional layers are used to acquire the feature extraction sets and perform graph convolution processing on them. The processed feature extraction sets are then updated to the graph neural network structure, outputting graph convolutional information that integrates neighboring node information. The node risk assessment layer is a Bayesian network layer, incorporating an attention mechanism and a multi-task learning algorithm. Based on the attention mechanism, the risk assessment layer performs multi-dimensional weighting on the graph convolutional information integrating neighboring node information and performs attention aggregation on the multi-dimensional weighted features, outputting an attention aggregation layer. The combined node risk indicators are jointly learned by a multi-task learning algorithm based on a multi-objective learning method, outputting node risk values ​​that characterize the comprehensive risk of distributed monitoring nodes. A Bayesian network calculates the node risk attributes of the node risk indicators based on a probability density estimation function. The cascaded impact analysis layer is used to obtain the node risk values ​​and node risk attributes of the comprehensive risk of distributed monitoring nodes. A message passing mechanism is used to analyze the propagation weights of associated monitoring nodes in the graph neural network structure to obtain the propagation weights of associated monitoring nodes. The cascaded failure impact indicators of associated monitoring nodes are calculated by combining the propagation weights of associated monitoring nodes with a fault propagation path analysis algorithm. The path generation layer is a risk propagation tree structure, and the Dijkstra algorithm is introduced into the path generation layer. The path generation layer is used to perform risk propagation analysis on distributed monitoring nodes in combination with the cascaded failure impact indicators of associated monitoring nodes to generate the node risk propagation path corresponding to the smart power station.

[0012] Preferably, the method for real-time risk assessment of the feature extraction set based on a pre-built cascading propagation model includes: Obtain the feature extraction set, perform graph convolution processing on the feature extraction set, update the graph convolution processing feature extraction set to the graph neural network structure, and output graph convolution information that integrates neighboring node information; The risk assessment layer uses an attention mechanism to perform multi-dimensional weighting on graph convolutional information that integrates neighboring node information, and then performs attention aggregation on the multi-dimensional weighted features to output the node risk index after attention aggregation. Based on a multi-task learning algorithm, the node risk index after attention aggregation is jointly learned by multiple objectives, and the node risk value representing the comprehensive risk of distributed monitoring nodes is output. The node risk value of the comprehensive risk of distributed monitoring nodes is expressed as follows: Indicates the node risk value. It is the Sigmoid activation function. These are the weights and bias vectors of the distributed monitoring nodes, respectively. This is a vector representing the node risk index after attention aggregation; Bayesian networks calculate node risk attributes based on probability density estimation functions to determine node risk indices. The formula for calculating the node risk attribute of the node risk indicator is as follows: in, The node risk attribute represents the node risk indicator. The number of Gaussian components in the mixture. Indicates the first The probability density function of a Gaussian distribution. For the first A Gaussian distribution of mixed weights, The first The mean and variance of a Gaussian distribution.

[0013] Preferably, the method for calculating the cascading failure impact index of associated monitoring nodes based on the fault propagation path analysis algorithm includes: To obtain the node risk value and node risk attribute of the comprehensive risk of distributed monitoring nodes, a message passing mechanism is used to analyze the propagation weight of the associated monitoring nodes in the graph neural network structure, and the propagation weight of the associated monitoring nodes is obtained. The cascading failure impact index of associated monitoring nodes is calculated by combining the propagation weight of associated monitoring nodes with the fault propagation path analysis algorithm. By combining the cascading failure impact indicators of associated monitoring nodes, risk propagation analysis of distributed monitoring nodes is conducted to generate the node risk propagation path corresponding to the smart power station.

[0014] On the other hand, the present invention also provides a LoRa-based unattended monitoring system for smart power plants, the LoRa-based unattended monitoring system for smart power plants comprising: LoRa terminals are used to collect raw monitoring data from distributed monitoring nodes in power plants in real time, preprocess the raw monitoring data, and aggregate the preprocessed raw monitoring data to the edge computing gateway based on the LoRa network. The edge computing gateway is used to acquire pre-processed raw monitoring data in real time, extract edge features from the raw monitoring data in real time, obtain real-time features associated with distributed monitoring nodes, and use a dynamic feature filtering mechanism based on principal component analysis to filter the real-time features and output the feature extraction set. The risk assessment module is used to obtain a feature extraction set and perform real-time risk assessment on the feature extraction set based on a pre-built cascading propagation model to obtain the node risk value and node risk attribute that characterize the comprehensive risk of the distributed monitoring nodes. The propagation path generation module calculates the cascading failure impact index of associated monitoring nodes based on the fault propagation path analysis algorithm, and performs risk propagation analysis on distributed monitoring nodes in combination with the cascading failure impact index of associated monitoring nodes to generate the node risk propagation path corresponding to the smart power station.

[0015] Preferably, the LoRa terminal includes: The distributed sensing group collects real-time monitoring data from the sensing devices and the environment from the distributed monitoring nodes based on distributed sensors, and integrates the monitoring data from the sensing devices and the environment into raw monitoring data. The outlier handling unit is used to load the raw monitoring data, clean outliers and unify timestamps on the raw monitoring data to obtain raw monitoring data with unified timestamps. The adaptive noise reduction unit performs categorized filtering and noise reduction on the original monitoring data based on an adaptive filter to obtain the original monitoring data after filtering and noise reduction. The data association unit is used to acquire the original monitoring data after filtering and noise reduction, perform association analysis on the original monitoring data, establish mapping relationships between data, and mark the mapping relationships of the original monitoring data to obtain the original monitoring data containing the mapping relationships.

[0016] Compared with the prior art, the embodiments of this application have the following main advantages: In this embodiment of the invention, a dynamic feature screening mechanism based on principal component analysis is used to screen real-time features, and a cascaded propagation model based on graph neural networks is used for real-time risk assessment. The transformation from raw data to refined features and the initial screening of faults can be completed at the gateway side. It can dynamically screen out the features most valuable for risk assessment, improve the accuracy and flexibility of risk assessment, greatly reduce the amount of uplink data, reduce the transmission latency of LoRa network, and treat the power station as an organic whole through the cascaded propagation model, thereby avoiding the island-like assessment of distributed monitoring nodes. It can not only assess the risk of individual nodes, but also consider the interdependence between related nodes and the fault propagation path, thereby achieving accurate prediction of the fault propagation path and impact range.

[0017] In this embodiment of the invention, during the preprocessing of raw monitoring data, the original multi-source heterogeneous data can be transformed into denoised, highly correlated raw monitoring data. This effectively removes background noise while preserving useful details and features in the signal to the maximum extent. Furthermore, by performing correlation analysis on the filtered and denoised raw monitoring data, mapping relationships between data are established and labeled. This reveals the inherent connections between different data, providing a high-quality data foundation for real-time edge feature extraction.

[0018] In this embodiment of the invention, when performing correlation analysis on raw monitoring data, cluster analysis is used to establish a mapping relationship between data and typical risk patterns, directly linking data to potential risks. Furthermore, by calculating the Pearson correlation coefficient, correlations between data are identified, and data-data mapping relationships are established. This reveals the inherent connections between different data sets. Time window alignment technology allows for precise matching of monitoring data within the same time period with related data tables. This not only improves data processing speed but also ensures data consistency over time, providing a more accurate data foundation for subsequent analysis. The data after correlation analysis is clearer, more accurate, and contains richer correlation information. Finally, the cascading propagation model, combined with the correlation relationships between data, can more accurately predict the propagation path and impact range of faults, supporting fault isolation and emergency response.

[0019] In this embodiment of the invention, real-time edge feature extraction of the raw monitoring data can quickly and accurately locate the specific location and range of abnormal data. The real-time edge feature extraction method not only improves the speed of data processing but also enhances the sensitivity of anomaly detection. The design of the lightweight extraction network also makes the entire feature extraction process more efficient and energy-saving. The lightweight extraction network adopts a lightweight decoupling head and SiLU activation function, which minimizes model complexity and computation while ensuring feature extraction capabilities, ensuring real-time performance on edge devices and making it suitable for various resource-constrained distributed monitoring scenarios.

[0020] In this embodiment of the invention, a dynamic feature selection mechanism based on principal component analysis is used to select real-time features, which makes the selected feature set more correlated with risk. This allows the model to focus more on learning truly predictive patterns, thereby making more accurate risk assessments. Furthermore, irrelevant and noisy features are eliminated, reducing the risk of model overfitting and making it more stable when dealing with unseen data. The dynamic feature selection mechanism, combined with correlation weights and the dynamic definition of the cumulative contribution rate threshold, avoids the problems of information loss or insufficient dimensionality reduction that may be caused by fixed thresholds, thus achieving a dynamic optimal balance between information compression and information retention.

[0021] In this embodiment of the invention, when performing real-time risk assessment on the feature extraction set based on a pre-built cascaded propagation model, graph convolution processing allows the model to fuse information from neighboring nodes, treating distributed monitoring nodes as an organic whole. The risk assessment layer uses an attention mechanism to weight the graph convolution information in multiple dimensions and then aggregates the attention. This mechanism dynamically adjusts the importance of features in different dimensions, ensuring that the assessment results more accurately reflect the actual risk. Furthermore, the Bayesian network calculates the risk attributes of node risk indicators based on a probability density estimation function, providing an estimate of risk uncertainty. The method not only provides a quantitative value of the risk but also offers a confidence level, enhancing the model's interpretability. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the implementation process of the unattended monitoring method for smart power plants based on LoRa networks provided by this invention.

[0023] Figure 2 A schematic diagram of the implementation process of the preprocessing method for raw monitoring data is shown.

[0024] Figure 3 The diagram illustrates the process of performing correlation analysis on raw monitoring data.

[0025] Figure 4 The diagram illustrates the implementation process of a method for real-time edge feature extraction from raw monitoring data.

[0026] Figure 5 The diagram illustrates the implementation process of a method for filtering real-time features using a dynamic feature filtering mechanism based on principal component analysis.

[0027] Figure 6 The diagram illustrates the implementation process of a method for real-time risk assessment of feature extraction sets based on a pre-built cascading propagation model.

[0028] Figure 7 A schematic diagram of the implementation process of the method for calculating the cascading failure impact index of associated monitoring nodes based on the fault propagation path analysis algorithm is shown.

[0029] Figure 8 This is a schematic diagram of the structure of the unattended monitoring system for smart power plants based on LoRa networks provided by the present invention. Detailed Implementation

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0031] Existing risk assessment methods rely on isolated assessments based on historical data from individual nodes or devices. These assessments fail to consider the interdependencies and fault propagation paths between related nodes, leading to insufficient evaluation of risk propagation trends and the overall impact scope. To address these issues, we propose a method and system for unattended monitoring of smart power plants based on LoRa networks. In short, the method first preprocesses the raw monitoring data, extracts real-time edge features, and uses a dynamic feature filtering mechanism based on principal component analysis to filter real-time features. Then, it performs real-time risk assessment on the feature extraction set based on a pre-built cascading propagation model, obtaining node risk values ​​and node risk attributes that characterize the overall risk of distributed monitoring nodes. Finally, it analyzes the risk propagation of distributed monitoring nodes by combining cascading failure impact indicators of related monitoring nodes. In this embodiment of the invention, a dynamic feature screening mechanism based on principal component analysis is used to screen real-time features, and a cascaded propagation model based on graph neural networks is used for real-time risk assessment. The transformation from raw data to refined features and the initial screening of faults can be completed at the gateway side. It can dynamically screen out the features most valuable for risk assessment, improve the accuracy and flexibility of risk assessment, greatly reduce the amount of uplink data, reduce the transmission latency of LoRa network, and treat the power station as an organic whole through the cascaded propagation model, thereby avoiding the island-like assessment of distributed monitoring nodes. It can not only assess the risk of individual nodes, but also consider the interdependence between related nodes and the fault propagation path, thereby achieving accurate prediction of the fault propagation path and impact range.

[0032] This invention provides a method for unattended monitoring of smart power plants based on LoRa networks. Figure 1 The diagram illustrates the implementation process of an unattended monitoring method for a smart power station based on a LoRa network. Specifically, the LoRa-based unattended monitoring method for a smart power station includes: S10 uses LoRa terminal 100 to collect raw monitoring data from the distributed monitoring nodes of the power station in real time, preprocesses the raw monitoring data, and aggregates the preprocessed raw monitoring data to the edge computing gateway 200 based on the LoRa network. In this embodiment of the invention, the distributed monitoring node refers to a LoRa terminal 100 that is distributed and deployed at various key locations within the power plant area, integrating sensing, data acquisition and LoRa communication functions. The LoRa terminal 100 can wirelessly transmit the processed data to the edge computing gateway 200 in a low-power, long-distance manner via the LoRaWAN protocol.

[0033] S20, the edge computing gateway 200 acquires the preprocessed monitoring raw data in real time, performs real-time edge feature extraction on the monitoring raw data, obtains the real-time features associated with the distributed monitoring nodes, and uses a dynamic feature filtering mechanism based on principal component analysis to filter the real-time features and output the feature extraction set. S30, Obtain the feature extraction set, and perform real-time risk assessment on the feature extraction set based on the pre-built cascade propagation model to obtain the node risk value and node risk attribute that characterize the comprehensive risk of the distributed monitoring nodes. S40: Calculate the cascading failure impact index of associated monitoring nodes based on the fault propagation path analysis algorithm, and perform risk propagation analysis on distributed monitoring nodes in combination with the cascading failure impact index of associated monitoring nodes to generate the node risk propagation path corresponding to the smart power station.

[0034] In this embodiment of the invention, a dynamic feature screening mechanism based on principal component analysis is used to screen real-time features, and a cascaded propagation model based on graph neural networks is used for real-time risk assessment. The transformation from raw data to refined features and the initial screening of faults can be completed at the gateway side. It can dynamically screen out the features most valuable for risk assessment, improve the accuracy and flexibility of risk assessment, greatly reduce the amount of uplink data, reduce the transmission latency of LoRa network, and treat the power station as an organic whole through the cascaded propagation model, thereby avoiding the island-like assessment of distributed monitoring nodes. It can not only assess the risk of individual nodes, but also consider the interdependence between related nodes and the fault propagation path, thereby achieving accurate prediction of the fault propagation path and impact range.

[0035] This invention provides a method for preprocessing raw monitoring data. Figure 2 This diagram illustrates the implementation flow of a method for preprocessing raw monitoring data. The method specifically includes: S101 is based on distributed sensors to collect real-time monitoring data of distributed monitoring nodes and environmental monitoring data, and integrates the monitoring data of distributed monitoring nodes and environmental monitoring data into raw monitoring data. It should be noted that sensing devices include, but are not limited to, current and voltage sensors, temperature sensors, humidity sensors, pressure sensors, and vibration sensors. The data monitored by these sensing devices includes, but is not limited to, current and voltage monitoring data, temperature monitoring data, pressure monitoring data, vibration monitoring data, insulation monitoring data, harmonic monitoring data, real-time power plant monitoring video data, light intensity monitoring data, and air quality monitoring data.

[0036] S102, Load the raw monitoring data, perform outlier cleaning and unified timestamp processing on the raw monitoring data to obtain raw monitoring data with unified timestamps. Outlier cleaning can remove errors or unreasonable values ​​in the data, avoiding these outliers from misleading subsequent analysis and processing. Unifying data from different sources to the same timestamp solves the problem of inconsistent data times. S103, perform type-based filtering and noise reduction processing on the original monitoring data based on the adaptive filter to obtain the original monitoring data after filtering and noise reduction. In this embodiment, the adaptive filter can be an LMS filter, a Kalman filter or an adaptive noise canceller. S104: Obtain the original monitoring data after filtering and noise reduction, perform correlation analysis on the original monitoring data, establish mapping relationships between data, and mark the mapping relationships of the original monitoring data to obtain the original monitoring data containing the mapping relationships.

[0037] In this embodiment of the invention, during the preprocessing of raw monitoring data, the original multi-source heterogeneous data can be transformed into denoised, highly correlated raw monitoring data. This effectively removes background noise while preserving useful details and features in the signal to the maximum extent. Furthermore, by performing correlation analysis on the filtered and denoised raw monitoring data, mapping relationships between data are established and labeled. This reveals the inherent connections between different data, providing a high-quality data foundation for real-time edge feature extraction.

[0038] This invention provides a method for performing correlation analysis on raw monitoring data. Figure 3 This diagram illustrates the implementation process of a method for correlation analysis of raw monitoring data. The method specifically includes: S1041 loads the filtered and denoised raw monitoring data, identifies the distributed monitoring nodes associated with the raw data, and establishes a data-node mapping relationship. By establishing this mapping relationship between data and distributed monitoring nodes, the source and location of each data point can be clearly identified. This allows operations and maintenance personnel to more intuitively understand the physical context of the data and improves its interpretability. S1042, calculate the Pearson correlation coefficient between the raw monitoring data based on the data type, determine whether the Pearson correlation coefficient exceeds a preset Pearson correlation threshold, and if the Pearson correlation coefficient exceeds the preset Pearson correlation threshold, establish a data-data mapping relationship between the raw data. Specifically, by calculating the Pearson correlation coefficient, the correlation between data is identified, and a data-data mapping relationship is established. This can reveal the inherent connections between different data. S1043, Based on K-means clustering analysis, cluster analysis is performed on typical risk patterns and monitoring raw data to establish a data-risk mapping relationship between monitoring raw data and typical risk patterns; S1044 establishes a related data table based on data-node mapping, data-data mapping, and data-risk mapping. It uses time window alignment technology to establish associations between the monitoring raw data and the related data table within the same time period, resulting in monitoring raw data containing the association relationships of the related data table. The discrete mapping relationships are integrated into a systematic related data table. By using time window alignment technology, it ensures that the association analysis is a dynamic and real-time process performed on the same time base, making the related data table a structured and queryable power plant monitoring knowledge base. Any analysis module can easily call these association relationships.

[0039] In this embodiment of the invention, when performing correlation analysis on raw monitoring data, cluster analysis is used to establish a mapping relationship between data and typical risk patterns, directly linking data to potential risks. Furthermore, by calculating the Pearson correlation coefficient, correlations between data are identified, and data-data mapping relationships are established. This reveals the inherent connections between different data sets. Time window alignment technology allows for precise matching of monitoring data within the same time period with related data tables. This not only improves data processing speed but also ensures data consistency over time, providing a more accurate data foundation for subsequent analysis. The data after correlation analysis is clearer, more accurate, and contains richer correlation information. Finally, the cascading propagation model, combined with the correlation relationships between data, can more accurately predict the propagation path and impact range of faults, supporting fault isolation and emergency response.

[0040] This invention provides a method for real-time edge feature extraction from raw monitoring data. Figure 4 This diagram illustrates the implementation flow of a method for real-time edge feature extraction from raw monitoring data. The method specifically includes: S201, load the monitoring raw data and identify the data type, perform real-time event detection on the monitoring raw data within the time window based on the preset abnormal threshold of the data type, and extract abnormal data points in the monitoring raw data. S202, obtain the abnormal data point identification results in the monitoring raw data, extract the monitoring raw data within the time window of the time period before and after the abnormal data point, and obtain the abnormal data window containing the abnormal data point. S203 utilizes a lightweight extraction network to perform two-level lightweight feature extraction on the raw monitoring data within the anomaly data window. The lightweight extraction network comprises a parallel CSPNet structure and an LSTM layer, both connected to a lightweight decoupled header. The network employs the SiLU activation function. The CSPNet structure extracts features from environmental monitoring data within the raw monitoring data, while the LSTM layer extracts features from sensor monitoring data, resulting in real-time features associated with distributed monitoring nodes. CSPNet, as a highly efficient variant of CNN, excels at extracting local features and patterns from image or spatial data, making it ideal for identifying environmental anomalies such as smoke, water accumulation, and illegal intrusion from camera footage or spatial sensor arrays. Meanwhile, temperature, vibration, and current generated by devices are typical time-series data, their value lying in their dynamic changes over time. LSTM is an expert in processing time-series data, effectively capturing long-term dependencies and thus identifying early signs of device failure.

[0041] In this embodiment of the invention, real-time edge feature extraction of the raw monitoring data can quickly and accurately locate the specific location and range of abnormal data. The real-time edge feature extraction method not only improves the speed of data processing but also enhances the sensitivity of anomaly detection. The design of the lightweight extraction network also makes the entire feature extraction process more efficient and energy-saving. The lightweight extraction network adopts a lightweight decoupling head and SiLU activation function, which minimizes model complexity and computation while ensuring feature extraction capabilities, ensuring real-time performance on edge devices and making it suitable for various resource-constrained distributed monitoring scenarios.

[0042] This invention provides a method for filtering real-time features using a dynamic feature filtering mechanism based on principal component analysis. Figure 5 The diagram illustrates the implementation process of a method for filtering real-time features using a dynamic feature filtering mechanism based on principal component analysis. This method specifically includes: S301, Load at least one set of real-time features, and perform Z-score standardization on the real-time features to obtain standardized features. The formula for Z-score standardization of real-time features is expressed as: in, These are standardized features and real-time features, respectively. These are the characteristic standard deviation and the characteristic mean, respectively. S302, Obtain standardized features, construct a feature covariance matrix based on the standardized features, and assign weights to the association relationships between the standardized features and the associated data tables using the mutual information method to obtain the association weights between the standardized features and the associated data tables. These association weights are then assigned to the feature covariance matrix to obtain a weighted covariance matrix. The "association weights" calculated using the mutual information method quantify the business relevance of each feature to the final risk objective. Assigning these weights to the covariance matrix means that the covariance relationships between features are no longer equal; features with a stronger correlation to risk have a greater influence in principal component analysis. This ensures that the subsequently generated principal components will naturally tilt towards features with "high business value." The weighted covariance matrix is ​​expressed as: in, For the weighted covariance matrix, This is the weighted standard characteristic matrix. Indicates the number of sample features; S303, load the weighted covariance matrix, perform eigenvalue decomposition on the weighted covariance matrix based on principal component analysis to obtain at least one set of principal components, calculate the variance contribution rate and cumulative contribution rate of the principal components, sort the principal components based on the cumulative contribution rate of the principal components, and output the important queue of principal components. S304 uses the association weights of standardized features and related data tables as the prior conditions for the dynamic feature screening mechanism, traverses the principal component importance queue, and performs a comprehensive ranking of the importance of principal components to obtain the re-ranked principal component importance queue. S305 identifies the cumulative contribution rate of principal components in the principal component importance queue. Its dynamic feature filtering mechanism combines correlation weights and cumulative contribution rates to dynamically define the cumulative contribution rate threshold, making the filtering process more sensitive and targeted rather than rigid. It intelligently determines how much information to retain based on the specific data conditions at each analysis time, thus finding the optimal balance between dimensionality reduction and retaining key information. The cumulative contribution rate threshold is calculated using the following formula: in, These are the cumulative contribution rate threshold and the basic screening threshold, respectively. The association weights of the standardized features corresponding to the principal components. These represent the cumulative contribution rate of the principal components and the mean cumulative contribution rate of the principal components, respectively. The number of principal components in the principal component importance queue. This is the threshold adjustment coefficient; S306: Select at least one principal component from the principal component importance queue that exceeds the cumulative contribution rate threshold, integrate at least one principal component and the standardized features before principal component decomposition into a feature extraction set, and output the feature extraction set.

[0043] In this embodiment of the invention, a dynamic feature selection mechanism based on principal component analysis is used to select real-time features, which makes the selected feature set more correlated with risk. This allows the model to focus more on learning truly predictive patterns, thereby making more accurate risk assessments. Furthermore, irrelevant and noisy features are eliminated, reducing the risk of model overfitting and making it more stable when dealing with unseen data. The dynamic feature selection mechanism, combined with correlation weights and the dynamic definition of the cumulative contribution rate threshold, avoids the problems of information loss or insufficient dimensionality reduction that may be caused by fixed thresholds, thus achieving a dynamic optimal balance between information compression and information retention.

[0044] In this embodiment of the invention, the cascaded propagation model is based on a graph neural network architecture. The nodes of the graph neural network are distributed monitoring nodes, and the edges represent the physical connections, communication links, and risk propagation relationships between nodes. The node features are feature extraction sets. The cascaded propagation model also includes an input layer and an output layer. Between the input and output layers are two graph convolutional layers, a node risk assessment layer, a cascaded impact analysis layer, and a path generation layer. The graph convolutional layers are used to acquire the feature extraction sets, perform graph convolution processing on the feature extraction sets, update the graph neural network structure with the processed feature extraction sets, and output graph convolutional information that integrates neighboring node information. The node risk assessment layer is a Bayesian network layer, and it incorporates an attention mechanism and a multi-task learning algorithm. Based on the attention mechanism, the risk assessment layer performs multi-dimensional weighting on the graph convolutional information that integrates neighboring node information, and performs attention aggregation on the multi-dimensional weighted features to output attention information. The node risk index after force aggregation is used to perform multi-objective joint learning on the node risk index after attention aggregation based on a multi-task learning algorithm, outputting a node risk value that represents the comprehensive risk of distributed monitoring nodes. A Bayesian network calculates the node risk attribute of the node risk index based on a probability density estimation function. The cascading impact analysis layer is used to obtain the node risk value and node risk attribute of the comprehensive risk of distributed monitoring nodes. A message passing mechanism is used to analyze the propagation weights of associated monitoring nodes in the graph neural network structure to obtain the propagation weights of associated monitoring nodes. The cascading failure impact index of associated monitoring nodes is calculated by combining the propagation weights of associated monitoring nodes with a fault propagation path analysis algorithm. The path generation layer is a risk propagation tree structure, and the Dijkstra algorithm is introduced into the path generation layer. The path generation layer is used to perform risk propagation analysis on distributed monitoring nodes in conjunction with the cascading failure impact index of associated monitoring nodes, generating the node risk propagation path corresponding to the smart power station. In this embodiment of the invention, the cascading propagation model introduces a fault propagation path analysis algorithm and a message passing mechanism, giving it dynamic behavior. This allows for the simulation of risks or faults spreading along pre-defined physical or logical paths (edges) within a network, calculating the "propagation weights" between nodes, and ultimately generating a visualized risk propagation path. Simultaneously, the model can reveal the chain reaction and amplification effects of risks. It can assess how a minor fault can trigger a "snowballing" cascading failure through the system network, thus enabling early warning of systemic risks.

[0045] This invention provides a method for real-time risk assessment of feature extraction sets based on a pre-built cascade propagation model. Figure 6 The diagram illustrates the implementation flow of a method for real-time risk assessment of feature extraction sets based on a pre-built cascading propagation model. Specifically, this method includes: S401, obtain the feature extraction set, perform graph convolution processing on the feature extraction set, update the graph convolution processing feature extraction set to the graph neural network structure, and output graph convolution information that integrates neighboring node information; S402, the risk assessment layer uses an attention mechanism to perform multi-dimensional weighting on graph convolutional information that integrates neighboring node information, and performs attention aggregation on the multi-dimensional weighted features to output the node risk index after attention aggregation. The attention mechanism can be understood as the model assigning appropriate weights or importance scores to different neighboring nodes and different features.

[0046] S403 utilizes a multi-task learning algorithm to jointly learn the node risk indicators after attention aggregation, outputting a node risk value that represents the comprehensive risk of distributed monitoring nodes. Multi-task learning allows the model to share underlying features and simultaneously learn to predict multiple objectives. This means the model does not simply output a general risk score, but can simultaneously output risk indicators with multiple dimensions such as "failure probability," "severity index," and "expected impact range." The node risk value of the comprehensive risk of distributed monitoring nodes is expressed as follows: Indicates the node risk value. It is the Sigmoid activation function. These are the weights and bias vectors of the distributed monitoring nodes, respectively. This is a vector representing the node risk index after attention aggregation; S404, a node risk attribute of Bayesian networks that calculates node risk indicators based on probability density estimation functions; The formula for calculating the node risk attribute of the node risk indicator is as follows: in, The node risk attribute represents the node risk indicator. The number of Gaussian components in the mixture. Indicates the first The probability density function of a Gaussian distribution. For the first A Gaussian distribution of mixed weights, The first The mean and variance of a Gaussian distribution.

[0047] In this embodiment of the invention, when performing real-time risk assessment on the feature extraction set based on a pre-built cascaded propagation model, graph convolution processing allows the model to fuse information from neighboring nodes, treating distributed monitoring nodes as an organic whole. The risk assessment layer uses an attention mechanism to weight the graph convolution information in multiple dimensions and then aggregates the attention. This mechanism dynamically adjusts the importance of features in different dimensions, ensuring that the assessment results more accurately reflect the actual risk. Furthermore, the Bayesian network calculates the risk attributes of node risk indicators based on a probability density estimation function, providing an estimate of risk uncertainty. The method not only provides a quantitative value of the risk but also offers a confidence level, enhancing the model's interpretability.

[0048] This invention provides a method for calculating the cascading failure impact index of associated monitoring nodes based on a fault propagation path analysis algorithm. Figure 7 The diagram illustrates the implementation flow of a method for calculating the cascading failure impact index of associated monitoring nodes based on a fault propagation path analysis algorithm. The method specifically includes: S501, obtain the node risk value and node risk attribute of the comprehensive risk of the distributed monitoring node, and use the message passing mechanism to analyze the propagation weight of the associated monitoring node in the graph neural network structure to obtain the propagation weight of the associated monitoring node. S502, combine the propagation weight of associated monitoring nodes and the fault propagation path analysis algorithm to calculate the cascading failure impact index of associated monitoring nodes; S503 combines the cascading failure impact indicators of associated monitoring nodes to perform risk propagation analysis on distributed monitoring nodes and generate the node risk propagation path corresponding to the smart power station.

[0049] In this embodiment of the invention, the cascading failure impact index is a quantitative and comparable indicator. It measures the maximum expected loss range that may result from the failure of a node, and the generated node risk propagation path can serve as a "system vulnerability map," enabling managers to go beyond single-point data and gain a deep understanding of the power plant's risk distribution from a system architecture perspective. Based on the cascading failure impact index, maintenance personnel can implement effective fault isolation measures to prevent further fault propagation. This not only reduces the impact of faults on the overall system but also improves the system's resilience.

[0050] On the other hand, embodiments of the present invention also provide an unattended monitoring system for smart power plants based on LoRa networks. Figure 8 A schematic diagram of a LoRa-based unattended monitoring system for a smart power plant is shown. This LoRa-based unattended monitoring system specifically includes: LoRa terminal 100 is used to collect raw monitoring data from the distributed monitoring nodes of the power plant in real time, preprocess the raw monitoring data, and aggregate the preprocessed raw monitoring data to the edge computing gateway 200 based on the LoRa network. The LoRa terminal 100 includes: The distributed sensing group 110 collects real-time monitoring data from the sensing devices and environmental monitoring data of the distributed monitoring nodes based on distributed sensors, and integrates the monitoring data from the sensing devices and environmental monitoring data into raw monitoring data. The outlier processing unit 120 is used to load the original monitoring data, perform outlier cleaning and unified timestamp processing on the original monitoring data, and obtain the original monitoring data with unified timestamps. The adaptive noise reduction unit 130 performs type-based filtering and noise reduction on the original monitoring data based on an adaptive filter to obtain the original monitoring data after filtering and noise reduction. The data association unit 140 is used to acquire the original monitoring data after filtering and noise reduction, perform association analysis on the original monitoring data, establish mapping relationships between data, and mark the mapping relationships of the original monitoring data to obtain the original monitoring data containing the mapping relationships.

[0051] The edge computing gateway 200 is used to acquire pre-processed monitoring raw data in real time, extract edge features from the monitoring raw data in real time, obtain real-time features associated with distributed monitoring nodes, and use a dynamic feature filtering mechanism based on principal component analysis to filter the real-time features and output the feature extraction set. The risk assessment module 300 is used to acquire a feature extraction set and perform real-time risk assessment on the feature extraction set based on a pre-built cascade propagation model to obtain the node risk value and node risk attribute that characterize the comprehensive risk of the distributed monitoring nodes. The propagation path generation module 400 calculates the cascading failure impact index of associated monitoring nodes based on the fault propagation path analysis algorithm, and performs risk propagation analysis on distributed monitoring nodes in combination with the cascading failure impact index of associated monitoring nodes to generate the node risk propagation path corresponding to the smart power station.

[0052] In summary, this invention provides a method and system for unattended monitoring of smart power plants based on LoRa networks. In the embodiments of this invention, a dynamic feature screening mechanism based on principal component analysis is used to screen real-time features, and a cascaded propagation model based on graph neural networks is used for real-time risk assessment. The transformation from raw data to refined features and the initial screening of faults can be completed at the gateway side. It can dynamically select the features most valuable for risk assessment, improve the accuracy and flexibility of risk assessment, greatly reduce the amount of uplink data, reduce the transmission latency of LoRa networks, and treat the power plant as an organic whole through the cascaded propagation model, thereby avoiding the island-like assessment of distributed monitoring nodes. It can not only assess the risk of individual nodes, but also consider the interdependence and fault propagation paths between related nodes, thereby achieving accurate prediction of the fault propagation path and impact range.

[0053] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0054] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A method for unattended monitoring of smart power plants based on LoRa networks, characterized in that, The method includes: LoRa terminals are used to collect raw monitoring data from the distributed monitoring nodes of the power plant in real time. The raw monitoring data is preprocessed and then aggregated to the edge computing gateway via the LoRa network. The edge computing gateway acquires pre-processed raw monitoring data in real time, performs real-time edge feature extraction on the raw monitoring data, obtains real-time features associated with distributed monitoring nodes, and uses a dynamic feature filtering mechanism based on principal component analysis to filter the real-time features and output a feature extraction set. The feature extraction set is obtained, and a real-time risk assessment is performed on the feature extraction set based on a pre-built cascade propagation model to obtain the node risk value and node risk attribute that characterize the comprehensive risk of the distributed monitoring nodes. The cascading failure impact index of associated monitoring nodes is calculated based on the fault propagation path analysis algorithm. The risk propagation analysis of distributed monitoring nodes is then performed based on the cascading failure impact index of associated monitoring nodes to generate the node risk propagation path corresponding to the smart power station. The method for calculating the cascading failure impact index of associated monitoring nodes based on the fault propagation path analysis algorithm includes: To obtain the node risk value and node risk attribute of the comprehensive risk of distributed monitoring nodes, a message passing mechanism is used to analyze the propagation weight of the associated monitoring nodes in the graph neural network structure, and the propagation weight of the associated monitoring nodes is obtained. The cascading failure impact index of associated monitoring nodes is calculated by combining the propagation weight of associated monitoring nodes with the fault propagation path analysis algorithm.

2. The unattended monitoring method for smart power plants based on LoRa networks as described in claim 1, characterized in that: The method for preprocessing raw monitoring data includes: Based on the real-time acquisition of sensor monitoring data and environmental monitoring data from distributed monitoring nodes using distributed sensors, the sensor monitoring data and environmental monitoring data are integrated into raw monitoring data; Load the raw monitoring data, perform outlier cleaning and unified timestamp processing on the raw monitoring data to obtain raw monitoring data with unified timestamps; The original monitoring data is subjected to categorized filtering and noise reduction based on an adaptive filter to obtain the original monitoring data after filtering and noise reduction. The original monitoring data after filtering and noise reduction is obtained, correlation analysis is performed on the original monitoring data to establish the mapping relationship between the data, and the mapping relationship of the original monitoring data is marked to obtain the original monitoring data containing the mapping relationship.

3. The unattended monitoring method for smart power plants based on LoRa networks as described in claim 2, characterized in that: The method for performing correlation analysis on raw monitoring data includes: Load the filtered and denoised original monitoring data, identify the distributed monitoring nodes associated with the original monitoring data, and establish a data-node mapping relationship; Calculate the Pearson correlation coefficient between the original monitoring data based on the data type, determine whether the Pearson correlation coefficient exceeds the preset Pearson correlation threshold, and if the Pearson correlation coefficient exceeds the preset Pearson correlation threshold, establish a data-data mapping relationship between the original data. Based on K-means clustering analysis, typical risk patterns and monitoring raw data are clustered to establish a data-risk mapping relationship between monitoring raw data and typical risk patterns. Based on data-node mapping, data-data mapping, and data-risk mapping, related data tables are established. Time window alignment technology is used to establish associations between the monitoring raw data and the related data tables within the same time period, resulting in monitoring raw data containing the association relationships of the related data tables.

4. The unattended monitoring method for smart power plants based on LoRa networks as described in claim 3, characterized in that: The method for real-time edge feature extraction of raw monitoring data includes: Load the raw monitoring data and identify the data type. Based on the preset abnormal threshold of the data type, perform real-time event detection on the raw monitoring data within the time window and extract abnormal data points from the raw monitoring data. Obtain the identification results of abnormal data points in the raw monitoring data, extract the raw monitoring data within the time window before and after the abnormal data points, and obtain the abnormal data window containing the abnormal data points. A lightweight extraction network is used to perform two-level lightweight feature extraction on the raw monitoring data within the abnormal data window.

5. The unattended monitoring method for smart power plants based on LoRa networks as described in claim 4, characterized in that: The method for filtering real-time features using a dynamic feature filtering mechanism based on principal component analysis includes: Load at least one set of real-time features, and perform Z-score standardization on the real-time features to obtain standardized features; Obtain standardized features, construct a feature covariance matrix based on the standardized features, and assign weights to the association between the standardized features and the associated data table using the mutual information method to obtain the association weights between the standardized features and the associated data table. Assign the association weights to the feature covariance matrix to obtain the weighted covariance matrix. Load the weighted covariance matrix, perform eigenvalue decomposition on the weighted covariance matrix based on principal component analysis to obtain at least one set of principal components, calculate the variance contribution rate and cumulative contribution rate of the principal components, sort the principal components based on the cumulative contribution rate of the principal components, and output the important queue of principal components. Using the association weights of standardized features and related data tables as prior conditions for the dynamic feature screening mechanism, the importance queue of principal components is traversed, and the importance of principal components is comprehensively ranked to obtain the re-ranked importance queue of principal components. The cumulative contribution rate of principal components in the important principal component queue is identified, and the dynamic feature screening mechanism combines the correlation weight and the cumulative contribution rate to dynamically define the threshold of the cumulative contribution rate. Select at least one principal component from the principal component importance queue that exceeds the cumulative contribution rate threshold, integrate at least one principal component and the standardized features before principal component decomposition into a feature extraction set, and output the feature extraction set.

6. The unattended monitoring method for smart power plants based on LoRa networks as described in claim 1, characterized in that: The cascaded propagation model is based on a graph neural network architecture. The cascaded propagation model also includes an input layer and an output layer. Between the input layer and the output layer, there are two graph convolutional layers, a node risk assessment layer, a cascaded impact analysis layer, and a path generation layer. The path generation layer is a risk propagation tree structure, and the Dijkstra algorithm is introduced into the path generation layer. The path generation layer is used to perform risk propagation analysis on distributed monitoring nodes by combining the cascaded failure impact indicators of associated monitoring nodes, and generate the node risk propagation path corresponding to the smart power station.

7. The unattended monitoring method for smart power plants based on LoRa networks as described in claim 6, characterized in that: The method for real-time risk assessment of feature extraction sets based on a pre-built cascading propagation model includes: The feature extraction set is obtained and graph convolution is performed on the feature extraction set. The graph convolution-processed feature extraction set is then updated to the graph neural network structure, and the graph convolution information that integrates the information of neighboring nodes is output. The risk assessment layer uses an attention mechanism to perform multi-dimensional weighting on graph convolutional information that integrates neighboring node information, and then performs attention aggregation on the multi-dimensional weighted features to output the node risk index after attention aggregation. Based on a multi-task learning algorithm, the node risk index after attention aggregation is jointly learned by multiple objectives, and the node risk value representing the comprehensive risk of distributed monitoring nodes is output. Bayesian networks calculate node risk attributes based on probability density estimation functions to determine node risk indices.

8. A LoRa-based unattended monitoring system for smart power plants, used to implement the LoRa-based unattended monitoring method for smart power plants as described in any one of claims 1-7, characterized in that: The LoRa-based smart power plant unattended monitoring system includes: LoRa terminals are used to collect raw monitoring data from distributed monitoring nodes in power plants in real time, preprocess the raw monitoring data, and aggregate the preprocessed raw monitoring data to the edge computing gateway based on the LoRa network. The edge computing gateway is used to acquire pre-processed raw monitoring data in real time, extract edge features from the raw monitoring data in real time, obtain real-time features associated with distributed monitoring nodes, and use a dynamic feature filtering mechanism based on principal component analysis to filter the real-time features and output the feature extraction set. The risk assessment module is used to obtain a feature extraction set and perform real-time risk assessment on the feature extraction set based on a pre-built cascading propagation model to obtain the node risk value and node risk attribute that characterize the comprehensive risk of the distributed monitoring nodes. The propagation path generation module calculates the cascading failure impact index of associated monitoring nodes based on the fault propagation path analysis algorithm, and performs risk propagation analysis on distributed monitoring nodes in combination with the cascading failure impact index of associated monitoring nodes to generate the node risk propagation path corresponding to the smart power station.

9. The unattended monitoring system for smart power plants based on LoRa networks as described in claim 8, characterized in that: The LoRa terminal includes: The distributed sensing group collects real-time monitoring data from the sensing devices and the environment from the distributed monitoring nodes based on distributed sensors, and integrates the monitoring data from the sensing devices and the environment into raw monitoring data. The outlier handling unit is used to load the raw monitoring data, clean outliers and unify timestamps on the raw monitoring data to obtain raw monitoring data with unified timestamps. The adaptive noise reduction unit performs categorized filtering and noise reduction on the original monitoring data based on an adaptive filter to obtain the original monitoring data after filtering and noise reduction. The data association unit is used to acquire the original monitoring data after filtering and noise reduction, perform association analysis on the original monitoring data, establish mapping relationships between data, and mark the mapping relationships of the original monitoring data to obtain the original monitoring data containing the mapping relationships.

Images