Power plant black-start unit monitoring and fault early warning method and readable storage medium

Abstract

The invention proposes a power plant black-start unit monitoring and fault early warning method, and the method comprises the steps: deploying a sensor array at a key part of a unit through a high-precision sensor network, constructing a neural network model through an LSTM deep learning algorithm based on collected key state data, carrying out the modeling analysis of multi-dimensional time series data, and carrying out the early warning of a fault of the unit. The method comprises the following steps: acquiring a dynamic characteristic mode of a unit operation state, reconstructing a characteristic vector into a two-dimensional characteristic graph by combining a CNN convolutional neural network, performing fault mode identification, acquiring a potential fault type and risk level assessment, and establishing a dynamic threshold adjustment mechanism by adopting an adaptive threshold fault diagnosis algorithm based on the fault type and risk level assessment. And abnormal state determination is carried out, and accurate fault early warning signals and disposal suggestions are obtained. According to the method, the data processing capacity and the system expansibility are improved, the prediction model and the optimization model are constructed through the big data analysis technology, scientific selection and starting time sequence arrangement of the black-start unit are achieved, and the success rate and reliability of black-start are improved.

Description

Technical Field

[0001] This invention relates to the field of black start technology in power systems, specifically to a method for real-time monitoring and fault early warning of black start units in power plants. Background Technology

[0002] Black start refers to the process of gradually restoring the entire power system through self-starting generators in the event of a large-scale power outage. As a key resource for grid restoration, the operational status of black start generators directly affects the overall grid restoration efficiency and reliability.

[0003] Currently, the monitoring of black-start units in power plants mainly adopts traditional SCADA systems, which collect unit operating parameters through a limited number of sensors and rely on simple threshold judgments for fault diagnosis. This approach suffers from problems such as low data acquisition accuracy, incomplete monitoring point coverage, and untimely fault diagnosis, and cannot meet the needs for real-time and accurate monitoring of unit status during black start-up.

[0004] Traditional black-start generator selection relies primarily on experience and static analysis, which cannot adapt to changes in generator status in real time. This can lead to generator malfunctions or even startup failures during the black start process, severely impacting grid recovery efficiency. Therefore, there is an urgent need for a method that can monitor the status of black-start generators in real time, accurately predict potential faults, and scientifically select the optimal black-start generator. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides a method for real-time monitoring and fault early warning of black-start generator units in power plants, comprising: deploying sensor arrays at key locations of the black-start generator unit based on a high-precision sensor network and connecting them to a data acquisition gateway via an industrial Ethernet network for millisecond-level data acquisition, obtaining key status data including temperature, pressure, vibration, and electrical parameters; based on the key status data, constructing a neural network model using an LSTM deep learning algorithm to model and analyze multi-dimensional time-series data, obtaining dynamic feature patterns of the unit's operating status; and based on the dynamic feature patterns, reconstructing the feature vectors using a CNN convolutional neural network. The system generates a two-dimensional feature map and performs fault mode recognition to obtain potential fault types and risk level assessments for the units. Based on the fault type and risk level assessments, an adaptive threshold fault diagnosis algorithm is used to establish a dynamic threshold adjustment mechanism for abnormal state determination, obtaining accurate fault warning signals and handling suggestions. Based on the distributed architecture of the cloud computing platform, monitoring data from multiple units are centrally stored and processed in parallel to obtain a comprehensive operational status assessment of all black-start units in the plant. Based on the comprehensive operational status assessment, a unit operational status prediction model and a multi-objective optimization model are constructed using big data analysis technology to obtain the optimal black-start unit selection scheme and startup sequence arrangement.

[0006] In one embodiment of the present invention, the step of deploying a sensor array at key locations of the black-start unit based on a high-precision sensor network to collect key status data includes: Based on the key components of the black starter unit, a high-precision sensor array is established by deploying turbine bearing temperature sensors, cylinder pressure sensors, generator stator winding temperature sensors, and rotor vibration sensors. Based on the high-precision sensor array, the data including temperature, pressure, vibration, and key electrical status are acquired.

[0007] In one embodiment of the present invention, based on the key status data, a neural network model is constructed using the LSTM deep learning algorithm to model and analyze multi-dimensional time-series data, thereby obtaining dynamic characteristic patterns of the unit's operating status, including: Construct an LSTM neural network model containing an input layer and hidden layers, and establish a temporal feature extraction network based on the key state data; Based on the aforementioned temporal feature extraction network, the input data is preprocessed by normalization and training samples are constructed using a sliding time window method to generate a standardized training dataset. Based on the standardized training dataset, the network weights are trained using the backpropagation algorithm to learn the temporal variation patterns of the unit's operating parameters, thereby obtaining the dynamic feature patterns of the unit's operating status.

[0008] In one embodiment of the present invention, the key state data is used as the input of the model, and temporal feature extraction and analysis are performed through an LSTM neural network model, wherein the hidden layer contains 128 LSTM units to establish a temporal feature extraction network. The number of nodes in the input layer of the LSTM neural network model is the same as the number of sensors used for data acquisition. Each node corresponds to a parameter acquired by a sensor, which can effectively capture the dependencies in long-term series data. The number of nodes in the network output layer is the same as the feature dimension, and it is used to output the extracted dynamic features.

[0009] In one embodiment of the present invention, based on the dynamic feature pattern, the feature vector is reconstructed into a two-dimensional feature map using a CNN convolutional neural network and fault mode recognition is performed to obtain the potential fault types and risk levels of the unit, including: Based on the dynamic feature pattern, the feature vector output by LSTM is reconstructed into a two-dimensional feature map, which is then used as input to the CNN network to generate a two-dimensional feature representation. Based on the two-dimensional feature representation, a CNN network containing 3 convolutional layers and 2 fully connected layers is constructed. The convolutional layers use 3×3 convolutional kernels to extract local features and spatial feature patterns. Based on the spatial feature patterns, a fault mode database is established, and the probability distribution of each fault type is calculated by pattern matching of the feature maps using a Softmax classifier, thereby obtaining the potential fault types and risk level assessments of the unit.

[0010] In one embodiment of the present invention, based on the fault type and risk level assessment, an adaptive threshold fault diagnosis algorithm is used to establish a dynamic threshold adjustment mechanism for abnormal state determination, thereby obtaining accurate fault warning signals and handling suggestions, including: Based on the aforementioned fault type and risk level assessment, the normal operating range of each parameter is dynamically calculated according to the statistical analysis of the unit's historical operating data, and an adaptive threshold adjustment mechanism is established. Based on the aforementioned adaptive threshold adjustment mechanism, the threshold is corrected by adopting the 3σ principle in combination with the unit's operating conditions, taking into account factors such as load changes and ambient temperature, and generating a dynamic threshold range. Based on the dynamic threshold range, when the fault identification result exceeds the dynamic threshold, a multi-level early warning mechanism is triggered, and corresponding handling suggestions are generated according to the fault type based on the expert knowledge base, thereby obtaining the accurate fault early warning signal and handling suggestions.

[0011] In one embodiment of the present invention, a microservice architecture including data access service, storage service, computing service and analysis service is constructed based on a cloud computing platform. Each service interacts with data through API interfaces to establish a distributed processing platform. Based on the aforementioned distributed processing platform, Apache Kafka message queues are used to receive real-time data from each unit and massive data storage is performed through the distributed file system HDFS, thereby achieving centralized management of multi-unit data. Based on the centralized management of multi-unit data, the Apache Spark cluster is used for parallel data processing and a unit status assessment model is established to comprehensively consider factors such as unit availability, startup time, and output capacity, thereby obtaining a comprehensive operational status assessment of the entire plant's black-start units.

[0012] In one embodiment of the present invention, based on the comprehensive operating status assessment, a machine learning regression algorithm is used to construct a unit status prediction model to predict the future availability probability of the unit based on historical operating data, and generate unit availability prediction results. Based on the unit availability prediction results, a multi-objective optimization model is established with the objective functions of maximizing the black start success rate and minimizing the start-up time. The constraints include unit technical characteristics, fuel supply, and auxiliary equipment status. An optimization solution model is constructed. Based on the aforementioned optimization solution model, a genetic algorithm is used to solve the optimization problem, and combined with the power grid topology and load distribution, the optimal black-start unit selection scheme and start-up sequence arrangement are obtained.

[0013] In one embodiment of the present invention, when evaluating fitness, a multi-level evaluation function is constructed: the feasibility of the solution is checked, and a penalty term is applied to the solution that does not meet the constraints; the two objective function values ​​of black start success rate and total start time are calculated; the network coverage efficiency of the start-up scheme is evaluated in combination with the power grid topology, and the electrical distance from the initial black start unit to each key load point is analyzed.

[0014] In another embodiment of the present invention, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions for causing a computer to execute the above-described method for real-time monitoring and fault early warning of power plant black start unit selection.

[0015] This invention achieves comprehensive and real-time monitoring of key operating parameters of black-start units through millisecond-level data acquisition using a high-precision sensor network, improving the accuracy and timeliness of data acquisition. It employs LSTM and CNN deep learning algorithms to analyze unit operating data, capturing complex temporal and spatial characteristics of unit operating status and improving the accuracy of fault diagnosis. By applying an adaptive threshold fault diagnosis algorithm, considering the influence of unit operating conditions and environmental factors, it achieves accurate judgment and timely early warning of abnormal unit states. Based on a distributed architecture of a cloud computing platform, it enables efficient processing and analysis of monitoring data from multiple units, improving data processing capabilities and system scalability. Through big data analytics, it constructs predictive and optimization models, enabling the scientific selection and startup sequence arrangement of black-start units, improving the success rate and reliability of black starts.

[0016] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for real-time monitoring and fault early warning of black start generator units in power plants according to the present invention.

[0019] Figure 2 This is a flowchart illustrating the data acquisition process based on a high-precision sensor network, as described in this invention.

[0020] Figure 3 This is a flowchart illustrating the dynamic feature patterns of unit operating status obtained by the present invention using the LSTM deep learning algorithm.

[0021] Figure 4 This is a flowchart of the CNN convolutional neural network fault mode recognition process of the present invention.

[0022] Figure 5 This is a flowchart illustrating the process of obtaining fault warning signals and handling suggestions through an adaptive threshold fault diagnosis algorithm, as described in this invention.

[0023] Figure 6 This invention provides a flowchart for evaluating the overall operating status of all black start units in the plant.

[0024] Figure 7 This invention provides a flowchart for obtaining the optimal black start unit selection scheme and start-up timing arrangement.

[0025] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. Detailed Implementation

[0026] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0027] It should be understood that the following specific examples illustrate the implementation of this disclosure, and those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0028] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0029] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0030] like Figure 1 As shown, a method for real-time monitoring and fault early warning of black-start generator units in power plants includes the following steps: Step S1: Deploy sensor arrays at key locations of the black-start generator set and connect them to a data acquisition gateway via industrial Ethernet for millisecond-level data acquisition, obtaining key status data including temperature, pressure, vibration, and electrical parameters. Specifically, this includes: deploying a PT100 temperature sensor at the turbine bearing with an accuracy of ±0.1℃; deploying a pressure transmitter at the cylinder with a range of 0-20MPa and an accuracy of 0.1%FS; deploying a fiber optic temperature sensor at the generator stator winding with a sampling frequency of up to 1kHz; and deploying an acceleration vibration sensor at the rotor with a frequency response range of 0.5Hz-10kHz. All sensors are connected to the data acquisition gateway via an industrial Ethernet switch, using a 24-bit high-precision A / D converter to digitize analog signals at a sampling frequency of up to 10kHz, and performing noise reduction processing on the signals using digital filtering algorithms. Simultaneously, a dual-redundant communication link is constructed, with the primary link using industrial Ethernet and the backup link using an industrial wireless network to ensure the reliability and integrity of data transmission.

[0031] Step S2: Construct a neural network model using the LSTM deep learning algorithm to model and analyze multi-dimensional time-series data, obtaining dynamic feature patterns of the unit's operating status. Specifically, this includes: constructing an LSTM neural network model containing an input layer and hidden layers, where the number of nodes in the input layer is the same as the number of sensors, and the hidden layer contains 128 LSTM units, each containing three control units: an input gate, a forget gate, and an output gate; normalizing and preprocessing the input data and constructing training samples using a sliding time window method, with a window length of 60 time steps and a sliding step size of 10, and using forward imputation to handle missing data; training the network using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64, while applying an early shutdown strategy and dropout technology (with a ratio of 0.3) to prevent overfitting, and finally extracting a 64-dimensional dynamic feature vector representing the unit's operating status.

[0032] Step S3: Combine the CNN convolutional neural network to reconstruct the feature vector into a two-dimensional feature map and perform fault mode recognition to obtain the potential fault types and risk level assessment of the unit. Specifically, this includes: reconstructing the 64-dimensional feature vector output by LSTM into an 8×8 two-dimensional feature map, preserving the correlation between features through spatial rearrangement; constructing a CNN network containing 3 convolutional layers and 2 fully connected layers. The first layer uses 32 3×3 convolutional kernels, the second layer uses 64 3×3 convolutional kernels connected to a max pooling layer, and the third layer uses 128 3×3 convolutional kernels connected to a global average pooling layer; the fully connected layer contains 256 nodes, and the number of nodes in the output layer matches the number of fault types; establishing a pattern database containing 10 common fault types such as bearing overheating, abnormal cylinder pressure, and excessive vibration; calculating the probability distribution of samples belonging to each fault type using a Softmax classifier, and determining a fault type as such when the probability of a certain fault type exceeds 0.7; and classifying the risk level into three levels: low (L1), medium (L2), and high (L3) based on the fault probability value and severity.

[0033] Step S4: Establish a dynamic threshold adjustment mechanism through an adaptive threshold fault diagnosis algorithm to determine abnormal states and obtain accurate fault warning signals and handling suggestions. Specifically, this includes: dynamically calculating the normal operating range of each parameter based on statistical analysis of historical unit operating data; collecting at least one year of normal unit operating data; dividing the dataset according to different load ranges and seasons; statistically analyzing each monitoring parameter in each dataset; calculating the mean μ and standard deviation σ; and initially determining the normal operating range of the parameters. ;use The principle is to adjust the threshold based on the unit's operating conditions, taking into account factors such as load changes and ambient temperature. For example, when the load increases, the upper limit threshold of vibration parameters is appropriately relaxed, and when the ambient temperature is high, the upper limit thresholds of bearing temperature and stator temperature are appropriately adjusted. A three-level early warning mechanism is set up: when the parameter approaches the threshold boundary (reaching 80% of the threshold), a level one early warning is triggered (reminding attention); when the parameter exceeds the threshold but does not reach the danger value, a level two early warning is triggered (recommending inspection); and when the parameter exceeds the threshold and reaches the danger value, a level three early warning is triggered (recommending shutdown for maintenance). In addition, corresponding handling suggestions are generated based on the fault type using an expert knowledge base.

[0034] Step S5: Based on the distributed architecture of the cloud computing platform, centrally store and process the monitoring data of multiple units in parallel to obtain a comprehensive operational status assessment of all black-start units in the plant. Specifically, this includes: constructing a microservice architecture based on the cloud computing platform, comprising data access services, storage services, computing services, and analysis services. Each service interacts with data through API interfaces, establishing a distributed processing platform; based on the distributed processing platform, using an Apache Kafka message queue to receive real-time data from each unit and storing massive amounts of data through the HDFS distributed file system, achieving centralized management of multi-unit data; based on the centralized management of multi-unit data, using an Apache Spark cluster for parallel data processing and establishing a unit status assessment model that comprehensively considers factors such as unit availability, startup time, and output capacity to obtain a comprehensive operational status assessment of all black-start units in the plant.

[0035] Step S6: Construct a unit operation status prediction model and a multi-objective optimization model using big data analytics to obtain the optimal black-start unit selection scheme and startup sequence. Specifically, this includes: using machine learning regression algorithms to construct a unit status prediction model, predicting the unit's availability probability within the next 24 hours based on historical operating data, and generating unit availability prediction results; establishing a multi-objective optimization model with the objective functions of maximizing black-start success rate and minimizing startup time, with constraints including unit technical characteristics, fuel supply, and auxiliary equipment status, and constructing an optimization solution model; using a genetic algorithm to solve the optimization problem and combining it with the power grid topology and load distribution to obtain the optimal black-start unit selection scheme and startup sequence.

[0036] like Figure 2 As shown, a sensor array is deployed at key locations of the black starter unit based on a high-precision sensor network and connected to a data acquisition gateway via industrial Ethernet for millisecond-level data acquisition. This acquires key status data including temperature, pressure, vibration, and electrical parameters, among others. S11: Sensors are deployed based on key components of the black-start unit, including core components such as turbine bearings, cylinders, generator stator windings, and rotor. These components are critical nodes in the unit's operation, and their condition directly affects the reliability and success rate of black starts. Faults can lead to start-up failures or operational anomalies, thus requiring close monitoring. The deployed sensors include turbine bearing temperature sensors, cylinder pressure sensors, generator stator winding temperature sensors, and rotor vibration sensors, forming a high-precision sensor array. A PT100 temperature sensor with an accuracy of ±0.1℃ is deployed at the turbine bearing; a pressure transmitter with a range of 0-20MPa and an accuracy of 0.1%FS is deployed at the cylinder; a fiber optic temperature sensor with a sampling frequency of up to 1kHz is deployed at the generator stator winding; and an acceleration vibration sensor with a frequency response range of 0.5Hz-10kHz is deployed at the rotor. These sensors constitute a high-precision sensor array covering the key components of the unit.

[0037] S12: Based on the high-precision sensor array, the sensor is connected to a data acquisition gateway via industrial Ethernet for A / D conversion, digital filtering, and protocol conversion to generate standardized digital signals. An industrial Ethernet switch connects each sensor to the data acquisition gateway, which uses a 24-bit high-precision A / D converter to digitize the analog signals with a sampling frequency of up to 10kHz. Digital filtering algorithms are applied to reduce noise in the signals, including low-pass filtering and median filtering. The data formats output from different sensors are uniformly converted, and the OPC UA protocol is used as the standard communication protocol to ensure data consistency and interoperability.

[0038] S13: Based on the standardized digital signal, establish redundant communication links to ensure data transmission reliability and acquire key status data including temperature, pressure, vibration, and electrical parameters. Construct dual redundant communication links: the primary link uses industrial Ethernet, and the backup link uses an industrial wireless network; apply network status monitoring technology to monitor the communication link status in real time, automatically switching to the backup link when the primary link fails; employ data caching technology to locally store collected data during network interruptions and automatically upload it after network recovery, ensuring data integrity; ultimately acquire key status data including bearing temperature, cylinder pressure, stator winding temperature, and rotor vibration, with data acquisition accuracy reaching the millisecond level.

[0039] like Figure 3 As shown, based on the key status data, a neural network model is constructed using the LSTM deep learning algorithm to model and analyze multi-dimensional time-series data, obtaining dynamic characteristic patterns of the unit's operating status, including: S21: Based on the aforementioned key state data, an LSTM neural network model containing an input layer and hidden layers is constructed. The key state data serves as the model's input, and the LSTM network extracts and analyzes temporal features. The hidden layer contains 128 LSTM units, establishing a temporal feature extraction network. The number of nodes in the LSTM network model's input layer matches the number of sensors used for data acquisition, with each node corresponding to parameters collected by one sensor. The hidden layer contains 128 LSTM units, each with three control units: an input gate, a forget gate, and an output gate, effectively capturing dependencies in long-term sequence data. The number of nodes in the network's output layer matches the feature dimension, used to output the extracted dynamic features. The LSTM network uses the tanh function as its activation function, and the sigmoid function is used for the forget gate and input gate. Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network structure specifically designed to address the long-term dependency problem in traditional RNNs. By introducing gating mechanisms (including input, forget, and output gates), LSTM can effectively learn and retain important information in long-term sequence data while selectively forgetting irrelevant information. This structure makes LSTM particularly suitable for processing time series data, such as the time-series variation patterns of unit operating parameters.

[0040] S22: Based on the aforementioned temporal feature extraction network, the input data is preprocessed by normalization, and training samples are constructed using a sliding time window method. The window length is set to 60 time steps to generate a standardized training dataset. The data collected by each sensor is subjected to max-min normalization processing, using the formula... Calculate, where x is the original data value, and The historical minimum and maximum values ​​of the parameter are given, and all data are linearly transformed and mapped to the [0,1] interval. Training samples are constructed using a sliding time window method, with a window length of 60 time steps (corresponding to 1 minute of data, assuming a sampling frequency of 1Hz) and a sliding step size of 10. Missing data are processed using a forward imputation method, that is, using the nearest valid value before the missing value to fill the missing position, ensuring the continuity of the time series. For missing values ​​at the beginning of the sequence, a backward imputation method is used, using the nearest valid value after the missing value to fill the missing position. When consecutive missing values ​​exceed a preset threshold (10 time steps), the average value of the historical period is used for imputation to avoid data deviation caused by large missing segments. Through the above preprocessing steps, a standardized training dataset is generated, containing a large number of time window samples. Each sample contains 60 time steps (corresponding to 1 minute of sampling data) of multidimensional sensor data, with an overlap rate of 83.3% between samples (sliding step size of 10), ensuring that the model can fully learn the temporal characteristics of the data.

[0041] S23: Based on the standardized training dataset, the network weights are trained using the backpropagation algorithm to learn the temporal variation pattern of the unit's operating parameters, thereby obtaining the dynamic feature pattern of the unit's operating status. The Adam optimizer was used for network training, with an initial learning rate of 0.001 and a batch size of 64. An early stopping strategy was employed during training: training was stopped when the loss function on the validation set showed no improvement for 10 consecutive epochs. Specifically, the patience parameter was set to 10. After each epoch, the loss value on the validation set was calculated, and the current loss value was compared with the historical best loss value. If the current loss value was lower, the optimal model parameters were updated and the counter was reset; otherwise, the counter was incremented. Training was terminated and the optimal model parameters were restored when the counter reached the patience value. Dropout technology (with a ratio of 0.3) was applied to prevent overfitting. Specifically, a dropout layer was added after the hidden layers of the LSTM network, randomly setting the output of 30% of the neurons to zero. This forced the network to learn more robust feature representations, avoiding overfitting to the training data. Simultaneously, different neuron combinations were activated in different training epochs, effectively training an ensemble model with multiple different network structures. After training, the LSTM network was able to extract the temporal variation patterns of the unit's operating parameters, generating dynamic feature patterns representing the unit's operating state. The feature dimension was 64, containing key temporal features of the unit's operating state.

[0042] like Figure 4 As shown, based on the dynamic feature pattern, the feature vector is reconstructed into a two-dimensional feature map using a CNN convolutional neural network, and fault mode recognition is performed to obtain the potential fault types and risk levels of the unit, including: S31: Based on the dynamic feature pattern, the feature vector output by the LSTM is reconstructed into a two-dimensional feature map as input to the CNN network, generating a two-dimensional feature representation. The 64-dimensional feature vector output by the LSTM network is reconstructed into an 8×8 two-dimensional feature map. The reconstruction process adopts a row-first arrangement, filling the two-dimensional matrix with consecutive feature values ​​sequentially. By converting the one-dimensional feature vector into a two-dimensional feature map, the spatial relationship between features can be better captured, ensuring that adjacent features remain spatially adjacent, which is convenient for the convolution kernel to extract local feature patterns. The reconstructed feature map is subjected to Min-Max normalization, mapping the feature values ​​to the [0,1] interval to ensure that the feature value distribution is suitable for the processing of the CNN network and avoids the gradient explosion problem caused by excessively large numerical ranges. At the same time, batch normalization is applied to standardize the feature map on each mini-batch, accelerating network convergence and improving model stability. The generated two-dimensional feature representation not only retains the temporal feature information extracted by the LSTM, but also provides a spatially correlated structured representation, which is more suitable for the CNN network to perform convolution operations and spatial feature extraction, thereby achieving effective fusion of temporal and spatial features.

[0043] S32: Based on the aforementioned two-dimensional feature representation, a CNN network consisting of three convolutional layers and two fully connected layers is constructed. The convolutional layers use 3×3 convolutional kernels to extract local features and spatial feature patterns. The first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1 and 'same' padding to maintain the feature map size. The activation function is ReLU to introduce non-linear transformation. The second convolutional layer uses 64 3×3 convolutional kernels to increase the number of feature channels, with a stride of 1 to maintain spatial resolution. The activation function is ReLU to improve the model's expressive power. A max-pooling layer (2×2, stride of 2) is then applied to achieve feature dimensionality reduction and position invariance. The third convolutional layer uses 128 3×3 convolutional kernels to further extract high-level features. The activation function is ReLU to enhance the network's deep learning capabilities. The network is followed by a global average pooling layer to convert the feature map into a fixed-length feature vector. The fully connected layer consists of two layers. The first layer has 256 nodes and is used for feature fusion and transformation. The activation function is ReLU to enhance the network's ability to express complex nonlinear relationships. The second layer has the number of nodes corresponding to the number of fault types (10 fault types) and serves as the final classification output layer. The CNN network extracts features layer by layer through this multi-layer convolutional structure, which can effectively capture spatial patterns such as local textures, edges, and shapes in the feature map. It also forms a discriminative high-level spatial feature representation through hierarchical abstraction, providing a reliable basis for fault classification.

[0044] S33: Based on the aforementioned spatial feature patterns, a fault mode database is established. A Softmax classifier is used to perform pattern matching on the feature maps to calculate the probability distribution of each fault type, thereby obtaining the potential fault types and risk levels of the unit. A fault mode database containing 10 common fault types is established, including bearing overheating, abnormal cylinder pressure, excessive vibration, abnormal stator winding temperature, insufficient bearing lubricating oil pressure, turbine speed control system malfunction, generator excitation system fault, vacuum system leakage, cooling water system malfunction, and auxiliary equipment control system fault. Each fault type is associated with a specific feature pattern and severity rating. A Softmax classifier is applied to the output layer of the CNN network, using the formula... Calculate samples Belonging to various fault types The probability distribution, where This is the raw output score for the corresponding category. When the probability value of a certain fault type exceeds a preset threshold of 0.7, the current state is determined to belong to that fault type. For cases where the probability is close to but does not exceed the threshold, it is marked as "potential risk" and continuously monitored. Based on the fault probability value and the inherent severity of the fault type, the risk level is quantified into three levels: Low (L1) indicates minor anomalies that require attention but do not affect unit startup and operation, with a probability between 0.7 and 0.8 or low severity; Medium (L2) indicates moderate anomalies that may affect unit performance but do not require immediate shutdown, with a probability between 0.8 and 0.9 or moderate severity; High (L3) indicates severe anomalies that require immediate handling and may cause the unit to fail to start, with a probability greater than 0.9 or high severity. An assessment report containing fault type, fault probability, risk level, and confidence level is generated to provide data support for subsequent fault warning and handling decisions.

[0045] like Figure 5 As shown, based on the fault type and risk level assessment, an adaptive threshold fault diagnosis algorithm is used to establish a dynamic threshold adjustment mechanism for abnormal state determination, obtaining accurate fault warning signals and handling suggestions, including: S41: Based on the aforementioned fault type and risk level assessment, and through statistical analysis of historical unit operating data, dynamically calculate the normal operating range of each parameter and establish an adaptive threshold adjustment mechanism. Collect at least one year of normal unit operating data, dividing it into 16 data subsets according to different load ranges (0-30%, 30%-60%, 60%-90%, 90%-100%) and different seasons (spring, summer, autumn, winter); perform statistical analysis on each monitoring parameter (bearing temperature, cylinder pressure, vibration amplitude, etc.) in each data subset, and calculate the mean. and standard deviation The preliminary determination of the normal operating range of the parameters is as follows: This interval theoretically covers 99.73% of normal operating data points; considering the time correlation of parameters, an autoregressive moving average model is adopted. Establish a time series model of the parameters, where The model parameters are estimated using the least squares method to capture the daily, weekly, and seasonal variations of the parameters. Based on the above statistical analysis and time series modeling, an adaptive threshold adjustment mechanism is established that considers the effects of load changes, seasonal variations, start-up and shutdown processes, and ambient temperature, so as to achieve dynamic optimization and adaptive updating of the threshold.

[0046] S42: Based on the aforementioned adaptive threshold adjustment mechanism, adopt... The principle is to adjust the threshold based on the unit's operating conditions, taking into account factors such as load changes and ambient temperature, to generate a dynamic threshold range. In principle, the initial threshold range is set as follows: The threshold values ​​are dynamically adjusted based on the current unit load status. When the load increases from 60% to 90% of the rated value, the upper limit threshold for vibration parameters increases by 10%. When the load exceeds 95% of the rated value, the upper limit threshold is further increased to 15%. A temperature correction coefficient is established to account for the influence of ambient temperature. ,in For ambient temperature, when When the temperature exceeds 25°C, the upper limit threshold for bearing temperature is increased. The upper limit threshold of stator temperature is increased by times. The system is divided into three phases for unit startup: preheating, acceleration, and grid connection. During the preheating phase, the upper limit of bearing temperature is allowed to increase by 20%; during the acceleration phase, the upper limit of vibration parameters is increased by 30%; and during the grid connection phase, the fluctuation range of electrical parameters is expanded by 15%. Thresholds are automatically recalculated every 5 minutes based on real-time collected operating parameters and environmental data, and then smoothed using an exponential smoothing method. The threshold change curve is smoothed to avoid false alarms caused by sudden threshold changes. Through the above dynamic adjustment mechanism, a dynamic threshold range that changes in real time with the operating conditions and environmental conditions of the random group is generated, thereby improving the adaptability and rationality of the threshold.

[0047] S43: Based on the dynamic threshold range, when the fault identification result exceeds the dynamic threshold, a multi-level early warning mechanism is triggered, and corresponding handling suggestions are generated according to the fault type based on the expert knowledge base, thereby obtaining the accurate fault early warning signal and handling suggestions. A three-level early warning mechanism is set up: when the parameter approaches the threshold boundary (reaching 80% of the threshold), a level one early warning is triggered (prompting attention); when the parameter exceeds the threshold but does not reach the danger value, a level two early warning is triggered (recommending inspection); when the parameter exceeds the threshold and reaches the danger value, a level three early warning is triggered (recommending shutdown for maintenance); an expert knowledge base containing experience in handling various faults is established, automatically generating corresponding handling suggestions for different fault types and severity; the handling suggestions include emergency measures, inspection items, maintenance plans, etc.; through the multi-level early warning mechanism and expert knowledge base, accurate early warning and timely handling guidance for abnormal unit states are achieved, improving the efficiency and accuracy of fault handling.

[0048] like Figure 6 As shown, based on the distributed architecture of the cloud computing platform, monitoring data from multiple units are centrally stored and processed in parallel to obtain a comprehensive operational status assessment of all black-start units in the plant, including: S51: Based on a cloud computing platform, a microservice architecture is constructed, comprising modules such as data access services, storage services, computing services, and analytics services, establishing a distributed processing platform. The microservice architecture is built using Docker container technology and the Kubernetes orchestration platform. The data access service is responsible for receiving data from various sensors, supporting multiple industrial protocols such as Modbus and OPC UA. The storage service is responsible for distributed storage and management of data, including real-time and historical databases. The computing service is responsible for parallel processing and analysis of data, supporting both stream and batch processing modes. The analytics service is responsible for unit status assessment and fault diagnosis, integrating various machine learning models. The microservices communicate with each other through RESTful APIs and message queues, achieving a loosely coupled architecture. Through this design, a highly available and highly scalable distributed processing platform is established.

[0049] S52: Based on the aforementioned distributed processing platform, an Apache Kafka message queue is used to receive real-time data from each unit and massive data is stored through the distributed file system HDFS, achieving centralized management of multi-unit data. An Apache Kafka cluster is deployed as the data access layer, with each unit's data acquisition gateway acting as a producer to send data to Kafka; multiple topics are configured, categorized according to unit type and data type; a Consumer Group is deployed to read data from Kafka and write it to HDFS; a tiered storage strategy is adopted on HDFS, with hot data (recent data) stored on SSDs and cold data (historical data) stored on HDDs; automatic data partitioning and compression strategies are implemented, with partitions stored according to time and unit ID; time-series databases such as InfluxDB are deployed to store recent real-time data, providing efficient time-series query capabilities; through the above design, centralized management of multi-unit monitoring data is achieved, supporting the access and storage of hundreds of thousands of data entries per second.

[0050] S53: Based on the centralized management of multi-unit data, an Apache Spark cluster is used for parallel data processing, and a unit status assessment model is established to comprehensively consider factors such as unit availability, start-up time, and output capacity to obtain a comprehensive operational status assessment of all black-start units in the plant. An Apache Spark cluster is deployed for parallel data processing. The cluster consists of one master node and eight worker nodes, each configured with 32 CPU cores and 128GB of memory, supporting dynamic resource allocation. Computational resources are automatically expanded according to real-time data volume and processing requirements. During peak data periods, the number of worker nodes can be expanded to 16, and during idle periods, it is automatically reduced to 4, achieving optimal resource utilization. A real-time data processing application based on Spark Streaming is developed, adopting a micro-batch processing mode with a batch processing interval set to 100 milliseconds. By optimizing the serialization method and memory management strategy, the end-to-end data processing latency is controlled within 200 milliseconds, meeting the real-time monitoring requirements in black-start scenarios. A Spark-based application is also developed. MLlib's distributed machine learning application integrates random forest, gradient boosting tree and deep learning models to build a model training and inference pipeline; the model training adopts an incremental learning method, and updates the model parameters with the latest data every 6 hours to ensure that the model adapts to changes in equipment status; the inference process adopts a batch processing architecture to evaluate the unit status in real time, and the evaluation results are updated once per second; a multi-dimensional comprehensive evaluation model is established, considering the following key factors: (1) Unit availability: based on the fault warning results, it is quantified as a continuous value of 0-100%, with a weight coefficient of 0.35; (2) Start-up time: based on historical start-up data and current hot and cold status prediction, it is divided into three categories: fast (<30 minutes), medium speed (30-60 minutes) and slow speed (>60 minutes), with a weight coefficient of 0.25; (3) Output capacity: based on the unit The maximum available output and ramp rate are estimated based on the type and current status, with a weighting coefficient of 0.20; (4) Fuel supply status: the continuous operation capability is assessed based on fuel inventory data and supply chain status, with a weighting coefficient of 0.10; (5) Auxiliary equipment status: the reliability of key auxiliary equipment such as cooling system and lubrication system is assessed based on auxiliary equipment monitoring data, with a weighting coefficient of 0.10; The weight of each factor is determined by the analytic hierarchy process (AHP), and the comprehensive score (0-100 points) of each unit is calculated by combining the fuzzy comprehensive evaluation method. The units are divided into three levels according to the score: preferred (>85 points), alternative (70-85 points) and not recommended (<70 points); the scoring results are updated every minute, and the comprehensive operation status assessment results of all potential black start units in the plant are displayed through a visual dashboard, providing intuitive support for scheduling decisions.

[0051] like Figure 7As shown, based on the comprehensive operational status assessment, a unit operational status prediction model and a multi-objective optimization model are constructed using big data analytics to obtain the optimal black-start unit selection scheme and startup sequence arrangement, including: S61: Based on the comprehensive operational status assessment, a machine learning regression algorithm is used to construct a unit status prediction model. This model predicts the unit's availability probability within the next 24 hours based on historical operational data, generating unit availability prediction results. A random forest regression algorithm is used to construct the prediction model. Input features include current unit status parameters, historical fault frequencies, and environmental conditions. Specifically, the input feature set includes the following types of data: real-time unit status parameters (such as bearing temperature, vibration values, pressure parameters, and other key monitoring indicators), historical fault statistics (such as fault frequency, average fault duration, and fault recovery time in the past 30 days), environmental condition parameters (such as ambient temperature, humidity, and atmospheric pressure), and time-series features such as unit operating time and maintenance records. The training dataset contains at least two years of historical operational data and fault records to ensure the model can learn the unit's operational patterns under different seasons and load conditions. The data preprocessing stage includes outlier detection and processing, missing value imputation, feature standardization, and time window feature construction to improve data quality. The optimal model parameters were determined through 5-fold cross-validation, and key parameters were determined after grid search optimization: 500 trees, a maximum depth of 10, a minimum number of leaf node samples of 5, and a feature sampling ratio of 0.7. Parallel computing technology was employed for model training, fully utilizing multi-core processors to accelerate the training process. The model output is a time series of the unit availability probability for the next 24 hours, with a time resolution of 1 hour, quantifying uncertainty through probabilistic prediction. To improve prediction accuracy, an ensemble learning method was applied, constructing a multi-model ensemble system including Random Forest, Gradient Boosting Tree (XGBoost), Support Vector Regression (SVR), and Long Short-Term Memory (LSTM). The prediction results of each base model were fused using weights determined by Bayesian optimization. The weight allocation was dynamically adjusted based on the performance of each model on the validation set, adapting to different operating conditions. The model was automatically updated monthly using the latest data, while historical model versions were retained for performance comparison analysis. The model update process included both incremental learning and full retraining modes, automatically selecting the update strategy based on the amount of new data. The model evaluation employs multiple metrics, including root mean square error (RMSE), mean absolute error (MAE), and R², to ensure stable and reliable predictive performance. This design enables the generation of high-precision predictions of the availability probability of each potential black start unit within the next 24 hours, providing reliable data support for black start unit selection.

[0052] S62: Based on the unit availability prediction results, establish a multi-objective optimization model with the objective functions of maximizing the black start success rate and minimizing the start-up time. Constraints include unit technical characteristics, fuel supply, and auxiliary equipment status. Construct an optimization solution model. The multi-objective optimization model is designed with the objective functions of: maximizing the black start success rate (calculated based on the unit availability prediction results) and minimizing the total start-up time (the sum of the start-up times of each unit). Specifically, the first objective function... It is expressed as a weighted product of the availability probabilities of all selected units, i.e. , where P i Let be the availability probability of the i-th unit. The importance weight of the unit in the black start sequence; the second objective function It represents the total time required for all units to go from cold start to grid connection, i.e. ,in For the first The startup time of the generating unit. Constraints include: unit technical characteristics constraints (such as minimum startup power not less than the load required at the initial stage of grid recovery, maximum ramp rate meeting load recovery speed requirements, etc.), fuel supply constraints (ensuring fuel supply meets the continuous operation requirements for at least 24 hours during startup), auxiliary equipment status constraints (the status score of necessary auxiliary equipment such as startup power supply, cooling water system, lubrication system, etc., not less than 85 points), and grid topology constraints (ensuring effective grid expansion after startup to cover critical load areas). A weighted objective method is used to transform the multi-objective problem into a single-objective problem, and the comprehensive objective function is expressed as follows: ,in The weighting coefficient has a range of values. Adjustments will be made dynamically based on the level of urgency. In extreme emergency situations, A value close to 1 prioritizes a high startup success rate; under normal recovery conditions, The value is set to 0.6-0.7 to balance success rate and time efficiency. The problem is formalized as a mixed integer programming problem, with decision variables including unit selection variables. (0-1 variables, 1 indicates the unit is selected, 0 indicates it is not selected) and start-up time variable (A continuous variable, representing the start-up time of the i-th unit). Priority constraints are introduced. Ensure that the selected options include a sufficiently high priority fleet, among which Assign priority scores to the generator units. This is the lowest priority requirement. Through the above design, a complete optimization model has been constructed, capable of comprehensively considering factors such as unit availability, start-up time, technical characteristics, fuel supply, auxiliary equipment status, and grid topology for decision optimization, providing a scientific basis for the selection of black-start units.

[0053] S63: Based on the aforementioned optimization solution model, a genetic algorithm is used to solve the optimization problem, combined with the power grid topology and load distribution, to obtain the optimal black-start unit selection scheme and startup sequence arrangement. A genetic algorithm solver is implemented, with a population size of 100 and a maximum number of iterations of 200. Integer encoding is used, with each chromosome consisting of two parts: a unit selection part and a startup sequence part. The former uses binary bits to indicate whether a unit is selected (1 indicates selected, 0 indicates not selected), and the latter uses an integer sequence to represent the startup order of the selected units. The initial population is generated through heuristic rules to ensure that the initial solution has a certain quality while maintaining population diversity. The crossover operation employs an improved two-point crossover, performing crossover on both the unit selection and startup sequence components, with a crossover probability set to 0.8. To maintain solution feasibility, a repair operation is performed after crossover on the startup sequence component to ensure no duplicate or missing sequence numbers occur. Mutation operations include bit-flip mutation (for unit selection) and swap mutation (for startup sequence), with a mutation probability set to 0.1. The selection operation uses tournament selection, randomly selecting three individuals for comparison each time, and choosing the individual with the highest fitness to proceed to the next generation. When evaluating fitness, a multi-level evaluation function is constructed: first, the feasibility of the solution is checked, and a penalty term is applied to solutions that do not meet the constraints; then, the black start success rate and total startup time are calculated as two objective function values; finally, the network coverage efficiency of the startup scheme is evaluated in conjunction with the grid topology, and the electrical distance from the initial black-start unit to each critical load point is analyzed to ensure that the startup scheme can efficiently restore the core area of ​​the grid. Considering a load priority classification system, loads are divided into three categories: extremely important loads (such as hospitals and data centers), first-level important loads (such as water plants and communication facilities), and ordinary loads. Priority is given to restoring high-priority load areas. An improved elite strategy is applied, retaining 10% of the best individuals from each generation directly into the next generation. A population diversity maintenance mechanism is also introduced, triggering a reorganization operation when population diversity falls below a threshold. Through this algorithm design, the algorithm converges to a stable solution after approximately 150 iterations, yielding the optimal black-start unit selection scheme and startup sequence arrangement. This includes: the recommended black-start unit sequence (sorted by priority), the optimal startup time for each unit (accurate to the minute), the power output curve during startup (including ramp rate constraints), the load restoration order (implemented in three stages), and detailed information such as the power grid topology evolution path at each stage. Visual decision support charts are generated to assist operators in execution.

[0054] This invention achieves comprehensive and real-time monitoring of key operating parameters of black-start units through millisecond-level data acquisition using a high-precision sensor network, improving the accuracy and timeliness of data acquisition. It employs LSTM and CNN deep learning algorithms to analyze unit operating data, capturing complex temporal and spatial characteristics of unit operating status and improving the accuracy of fault diagnosis. By applying an adaptive threshold fault diagnosis algorithm, considering the influence of unit operating conditions and environmental factors, it achieves accurate judgment and timely early warning of abnormal unit states. Based on a distributed architecture of a cloud computing platform, it enables efficient processing and analysis of monitoring data from multiple units, improving data processing capabilities and system scalability. Through big data analytics, it constructs predictive and optimization models, enabling the scientific selection and startup sequence arrangement of black-start units, improving the success rate and reliability of black starts.

[0055] Those skilled in the art will understand that, in order to solve the technical problem of how to achieve a good user experience, this embodiment may also include well-known structures such as communication buses and interfaces, and these well-known structures should also be included within the protection scope of this disclosure.

[0056] like Figure 8 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. It illustrates a structural schematic diagram suitable for implementing the computer device in the embodiments of the present disclosure. Figure 8 The computer device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0057] like Figure 8 As shown, a computer device may include a processor (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) or a program loaded from a storage device into random access memory (RAM). The RAM also stores various programs and data required for the operation of the computer device. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus. The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the inventive concept of the present invention and the description and drawings, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

[0058] A computer device according to embodiments of the present disclosure includes a memory and a processor. The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0059] The processor may be a central processing unit (CPU) or other form of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the computer device to perform desired functions. In one embodiment of this disclosure, the processor is used to execute computer-readable instructions stored in the memory, causing the computer device to perform all or part of the steps of the learning outcome prediction method based on learning behavior data mining of the foregoing embodiments of this disclosure.

[0060] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for power plant black start unit monitoring and fault early warning, characterized in that, The method comprises the following steps: Based on a high-precision sensor network, a sensor array is deployed at a key position of a black-start unit to collect key state data, wherein the key state data includes temperature, pressure, vibration, and electrical parameters; Based on the key state data, a neural network model is constructed through an LSTM deep learning algorithm to model and analyze multi-dimensional time series data, and a dynamic characteristic pattern of the unit operation state is obtained; Based on the dynamic characteristic pattern, a feature vector is reconstructed into a two-dimensional feature map through a CNN convolutional neural network for fault pattern recognition, and a potential fault type and risk level assessment of the unit are obtained; Based on the fault type and risk level assessment, a dynamic threshold adjustment mechanism is established through an adaptive threshold fault diagnosis algorithm for abnormal state judgment, and a precise fault warning signal and disposal suggestion are obtained.

2. The method of claim 1, wherein, Based on the high-precision sensor network, a sensor array is deployed at a key position of a black-start unit to collect key state data, comprising: Based on the key position of the black-start unit, a turbine bearing temperature sensor, a cylinder pressure sensor, a generator stator winding temperature sensor, and a rotor vibration sensor are deployed to establish a high-precision sensor array; Based on the high-precision sensor array, the key state data including temperature, pressure, vibration, and electrical parameters are obtained.

3. The method of claim 1, wherein, Based on the key state data, a neural network model is constructed through an LSTM deep learning algorithm to model and analyze multi-dimensional time series data, and a dynamic characteristic pattern of the unit operation state is obtained, comprising: An LSTM neural network model containing an input layer and a hidden layer is constructed, and a time series feature extraction network is established based on the key state data; Based on the time series feature extraction network, input data is normalized and preprocessed, and a training sample is constructed using a sliding time window method to generate a standardized training data set; Based on the standardized training data set, the network weight is trained through a back propagation algorithm to learn the time series variation law of the unit operation parameters, and the dynamic characteristic pattern of the unit operation state is obtained.

4. The method of claim 3, wherein: The key state data is used as the input of the model, and the LSTM neural network model is used for time series feature extraction and analysis, wherein the hidden layer contains 128 LSTM units, and a time series feature extraction network is established; The number of input layer nodes of the LSTM neural network model is consistent with the number of collected sensors, and each node corresponds to a parameter collected by a sensor, which can effectively capture the dependency of long time series data; The number of network output layer nodes is the same as the feature dimension, which is used to output the extracted dynamic features.

5. The method of claim 1, wherein, Based on the dynamic characteristic pattern, a feature vector is reconstructed into a two-dimensional feature map through a CNN convolutional neural network for fault pattern recognition, and a potential fault type and risk level assessment of the unit are obtained, comprising: Based on the dynamic characteristic pattern, the feature vector output by the LSTM is reconstructed into a two-dimensional feature map as the input of the CNN network to generate a two-dimensional feature representation; Based on the two-dimensional feature representation, a CNN network containing 3 convolutional layers and 2 fully connected layers is constructed, and a 3x3 convolution kernel is used in the convolutional layer to extract local features and spatial feature patterns. Based on the spatial feature mode, a fault mode database is established, and a Softmax classifier is used for pattern matching calculation of the feature map to obtain the probability distribution of each fault type, and the potential fault type and risk level assessment of the unit are obtained.

6. The method of claim 1, wherein, Based on the fault type and risk level assessment, a dynamic threshold adjustment mechanism is established through an adaptive threshold fault diagnosis algorithm to determine the abnormal state, and accurate fault warning signals and disposal suggestions are obtained, including: Based on the fault type and risk level assessment, the normal operation interval of each parameter is dynamically calculated based on historical operation data statistical analysis, and an adaptive threshold adjustment mechanism is established; Based on the adaptive threshold adjustment mechanism, the 3σ principle is used to correct the threshold considering the influence of load change, environmental temperature and other factors, and a dynamic threshold range is generated; Based on the dynamic threshold range, when the fault recognition result exceeds the dynamic threshold, a multi-level warning mechanism is triggered, and corresponding disposal suggestions are generated according to the fault type based on the expert knowledge base, and the accurate fault warning signals and disposal suggestions are obtained.

7. The method of claim 1, wherein, Further comprising: Based on the cloud computing platform, a micro-service architecture including data access service, storage service, computing service and analysis service is constructed, and each service realizes data interaction through API interface to establish a distributed processing platform; Based on the distributed processing platform, Apache Kafka message queue is used to receive real-time data of each unit and store massive data through distributed file system HDFS to realize centralized management of multi-unit data; Based on the centralized management of multi-unit data, Apache Spark cluster is used for data parallel processing and a unit state evaluation model is established to consider the availability, start-up time, output capacity and other factors of the unit to obtain the comprehensive operation state evaluation of the black-start unit.

8. The method of claim 7, wherein, Further comprising: Based on the comprehensive operation state evaluation, a unit state prediction model is constructed using machine learning regression algorithm to predict the availability probability of future units based on historical operation data, and a unit availability prediction result is generated; Based on the unit availability prediction result, a multi-objective optimization model is established with the maximum black-start success rate and the shortest start-up time as the objective function, and the constraint conditions include unit technical characteristics, fuel supply and auxiliary equipment state, and an optimization solving model is constructed; Based on the optimization solving model, genetic algorithm is used to solve the optimization problem and combined with the grid topology structure and load distribution, the optimal black-start unit selection scheme and start-up timing arrangement are obtained.

9. The method of claim 8, wherein, Further comprising, In the fitness evaluation, a multi-level evaluation function is constructed; the feasibility of the solution is checked, and a penalty term is applied to the solution that does not meet the constraint condition; The values of the two objective functions of black-start success rate and total start-up time are calculated; the network coverage efficiency of the start-up scheme is evaluated combined with the grid topology structure, and the electrical distance from the initial black-start unit to each key load point is analyzed.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for causing a computer to execute the power plant black-start unit selection real-time monitoring and fault warning method of any one of claims 1-9.

Images