Electric power system stability on-line monitoring method based on convolutional neural network

Abstract

The invention relates to an electric power system stability online monitoring method based on a convolutional neural network, and the method comprises the steps: collecting target data of all generator buses in real time, and carrying out the preprocessing of the target data, and the target data comprise a voltage amplitude, a phase angle and a frequency; generating a thermodynamic diagram according to the pre-processed target data, performing feature extraction and pattern classification on the thermodynamic diagram by using a convolutional neural network to obtain a category prediction result, the convolutional neural network being obtained by training through a training set, the training set being generated by a simulation fault sample, and the training set being obtained by training the simulation fault sample; each sliding window sample is endowed with a category label; according to the category prediction result, judging whether the state is an instability early warning state or not; and after an instability early warning state is determined, identifying a key generator set through a principal component analysis (PCA) method or a double-branch convolutional neural network. According to the invention, the accuracy and timeliness of transient stability prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system safety and stability analysis, in particular to a power system stability online monitoring method based on a convolutional neural network. BACKGROUND

[0002] At present, transient stability evaluation of power systems is crucial for the safe operation of power grids, and there are some challenges to be optimized. The existing transient stability evaluation methods of power systems mainly include time domain simulation method, transient energy function method and single machine equivalent method. The time domain simulation method can accurately simulate the dynamic response of the power system, but it has large calculation amount and poor real-time performance, which is difficult to meet the demand of large-scale power grid. The transient energy function method is simple to calculate, but it has high requirements for initial state and system parameters, and it is difficult to effectively respond to dynamic disturbances in large-scale power grids. The single machine equivalent method reduces the calculation amount by simplifying the model, but its accuracy and reliability are poor in complex systems. SUMMARY

[0003] The purpose of the present application is to provide a power system stability online monitoring method based on a convolutional neural network, which aims to solve the problem that the existing power system stability monitoring method has strong dependence on fault occurrence and clearance time, insufficient real-time performance, and cannot provide instability key factor positioning.

[0004] To achieve the above purpose, the present application provides the following scheme: A power system stability online monitoring method based on a convolutional neural network, comprising: Real-time acquisition of target data of all generator busbars and preprocessing, wherein the target data includes voltage amplitude, phase angle and frequency; Generating a heat map according to the preprocessed target data, using a convolutional neural network to extract features and classify patterns of the heat map, and obtaining a category prediction result, wherein the convolutional neural network is obtained by training a training set, and each sliding window sample is assigned a category label; According to the category prediction result, it is determined whether it is an instability early warning state; After determining that it is an instability early warning state, a principal component analysis (PCA) method or a double-branch convolutional neural network is used to identify a key generator set.

[0005] Optionally, the preprocessing of the target data of all generator busbars includes: Using a median filter to smooth the phase angle signal noise, and using the arithmetic mean of all generator phase angles as a reference angle to obtain initial preprocessing data, wherein the initial preprocessing data includes data points containing voltage amplitude, phase angle and frequency of G generator at each sampling time. The initial preprocessed data is analyzed using a sliding time window to obtain the preprocessed target data. The sliding window contains s sets of data, including the previous s-1 samples and the current time.

[0006] Optionally, generating a heatmap based on the preprocessed target data includes: The data within the sliding window is represented as a matrix, and the elements in the matrix are mapped to the red, green, and blue channels of the image to obtain the heat map. Each element of the matrix contains a vector of the voltage amplitude, phase angle, and frequency of the generator at the corresponding sampling time.

[0007] Optionally, the convolutional neural network includes convolutional layers, pooling layers, and fully connected layers; The convolutional layer extracts features from the heatmap and outputs a feature map; The feature map is processed by a non-linear activation function to obtain activation values; The pooling layer downsamples the activation value or takes local extrema to output high-dimensional features; The flattened high-dimensional features are input into the fully connected layer for pattern classification to obtain the category prediction result.

[0008] Optionally, assigning category labels to each sliding window sample includes: labeling all windows during the stable operation period before the fault as class 0, labeling windows covering the moment the fault occurs as class I, labeling windows after the fault occurs but before the fault is cleared as class II, labeling windows covering the moment the fault is cleared as class III, and labeling all windows after the fault is cleared as class IV.

[0009] Optionally, identifying the set of key generators using principal component analysis (PCA) includes: Select The time series data of frequency and phase angle of all generators within the time period constitutes the dimension of The data matrix, where, G represents the number of sampling points from fault clearing to the early warning interval, where G is the total number of generators. The initial moment of the fault. This is the moment when CNN first identified it as a Class 4 warning. Regarding the Principal component analysis (PCA) is performed on the data matrix to extract the directions of the first two principal components, and the data vector of each generator is projected onto the two principal component planes. A binary Gaussian distribution is used to fit and model the projection points. By utilizing the minimum covariance row and column, the generator corresponding to the projection point with the minimum likelihood value under the fitted probability distribution is obtained, which is the set of key generators.

[0010] Optionally, the identifying the key generator set through the double-branch convolutional neural network comprises: The features of the heat map are extracted through the CNN feature extractor, the extracted features are input into the first full-connection classification branch and the second full-connection classification branch, and the circuit system state and the key generator set are respectively output.

[0011] Optionally, the output layer of the second full-connection classification branch contains G+1 nodes, adopts a Sigmoid activation function, and outputs the probability indication of the key generator for each generator.

[0012] The application has the advantages that: the application can realize continuous monitoring and instability early warning of the transient stability state of the power system without knowing the fault time in advance, and can automatically identify the key generator causing instability when instability is predicted. The method improves the accuracy and timeliness of transient stability prediction, and helps the operation personnel to take targeted control measures to prevent the spread of accidents. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description only constitute some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0014] Figure 1 It is a general structure schematic diagram of an online monitoring method for power system stability based on a convolutional neural network according to an embodiment of the application. Figure 2 It is a CNN structure schematic diagram according to an embodiment of the application. Figure 3 It is a Y-Net basic architecture diagram according to an embodiment of the application. Figure 4 It is a sensitivity analysis curve diagram of the influence of N pairs of system performance according to an embodiment of the application. It is a sensitivity analysis curve diagram of the influence of N pairs of system performance according to an embodiment of the application. Figure 5 It is a robustness analysis diagram of OMS under type II noise according to an embodiment of the application. Figure 6 It is a projection to Measurement results and outlier identification of space, wherein (a) is 118 bus system: unstable case 1, (b) is 118 bus system: unstable case 2, (c) is 145 bus system: unstable case 1, (d) is 145 bus system: unstable case 2; Figure 7 Measurement results of space, wherein (a) is 118 bus system: stable case 1, (b) is 118 bus system, stable case 2. Measurement results of space, wherein (a) is 118 bus system: stable case 1, (b) is 118 bus system, stable case 2. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0016] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0017] The embodiment utilizes the generator bus voltage data collected in real time by the synchronous phasor measurement, identifies the "heat map" representing the system operating state through the convolutional neural network, judges whether the system is stable or not, and issues an alarm when detecting the instability risk.

[0018] The online monitoring method for power system stability based on the convolutional neural network provided in the embodiment comprises: Real-time collection of target data of all generator buses and preprocessing, wherein the target data comprises voltage amplitude, phase angle and frequency; Heat map generation according to the preprocessed target data, feature extraction and pattern classification of the heat map by the convolutional neural network, and acquisition of the category prediction result, wherein the convolutional neural network is obtained by training the training set, the training set is generated by the simulation fault sample, and each sliding window sample is assigned a category label; According to the category prediction result, it is determined whether it is in the instability early warning state or not; After determining that it is in the instability early warning state, the key generator set is identified by the principal component analysis (PCA) method or the double-branch convolutional neural network.

[0019] Specifically, as Figure 1The shown includes an input stage, a convolutional neural network classifier, and a key generator identification. The input stage preprocesses the PMU measurement data, combines the voltage measurements of multiple generators to form a two-dimensional matrix, and generates a heat map on a sliding time window; the convolutional neural network is used as the core classifier to extract features and classify patterns of the heat map corresponding to each time window; once the instability warning state is determined, the system enters the key generator identification stage, and a principal component analysis and a second neural network are combined to determine the key generator set leading to instability, and an alarm signal is triggered to notify the operator.

[0020] Further, the target data of all generator buses is preprocessed, including: The median filter is used to smooth the noise of the phase angle signal, and the arithmetic average of all generator phase angles is used as the reference angle to obtain the initial preprocessed data, wherein the initial preprocessed data includes data points obtained at each sampling time, including the voltage amplitude, phase angle and frequency of G generators; The initial preprocessed data is analyzed by using a sliding time window to obtain the preprocessed target data, wherein the sliding window contains s groups of data before s-1 sampling and the current time.

[0021] Specifically, the input data preprocessing includes: The online monitoring system OMS (Online Monitoring System) assumes that the PMU measurement of all generator buses is available at the center position, and can transmit the amplitude and phase angle of each bus voltage and other measurements to the central monitoring system in real time. The input data includes the instantaneous amplitude and phase angle of each generator bus voltage, as well as the frequency of three measurements. In order to reduce the influence of numerical differential amplification noise, the sampled phase angle signal is passed through a median filter with a kernel size of 3 before calculating the phase angle derivative, to smooth the noise spikes. The training data of the CNN is obtained by detailed simulation of the power system. In the simulator, all phase angles need to refer to a common reference point. Considering that selecting different generators as the reference phase angle may introduce bias, in order to ensure the robustness of the method to system parameter selection, this embodiment does not use a specific generator as the reference phase angle, but uses the arithmetic average of all generator phase angles as the reference angle. Assuming that the system has G generators, the reference phase angle is defined as: ; In the formula, is the kth phase angle sample corresponding to the ith generator, and G is the total number of generators in the system under consideration. If synchronous PMU measurement is available, this step is not needed.

[0022] In this way, the phase angle of each generator relative to the global average reference is obtained to eliminate the reference system failure problem caused by the possible instability of the single-machine reference. After the above processing, a data point containing the voltage amplitude, phase angle and frequency of G generators is obtained at each sampling time t.

[0023] Since the time of the disturbance is unknown and can occur at any time, the system continuously analyzes the data using a sliding time window mechanism. The length of the time window is set to s sampling periods, and each time new PMU sampling data is obtained, the data s periods before the current time is used to form a window for analysis. The window is overlapped and slides one sampling interval at a time, which ensures the continuity of the monitoring. For any given time, the sliding window contains s groups of data, including the previous (s-1) sampling and the current time.

[0024] Further, the heat map generation according to the preprocessed target data comprises: The data in the sliding window is represented as a matrix, and the elements in the matrix are mapped to the red, green and blue channels of the image, respectively, to obtain the heat map, wherein each element of the matrix contains a vector of the voltage amplitude, phase angle and frequency of the generator at the corresponding sampling time.

[0025] Specifically, the heat map generation and feature representation: The measurements of all generators in the sliding window form a matrix, which can construct the heat map representation of the time window. Specifically, for G generators and window length s, the data of each window can be represented as a matrix with a dimension of , where each element of the matrix contains a vector of three measurements of the generator at the corresponding sampling time. By mapping the three measurements to the red, green and blue channels of the image, respectively, the matrix is rendered as a color image with a size of , which is called a "heat map". In the heat map, different colors or brightness represent different measurement values, which can intuitively reflect the pattern of the dynamic changes of the voltage amplitude and phase angle of each generator in the time window.

[0026] In a long time range, the voltage amplitude and phase angle change pattern of each generator in a stable situation has a similar overall pattern, while the color distribution of the heat map in an unstable situation shows significant differences and irregular changes. The visual difference between the stable and unstable heat maps inspires the use of CNN's image pattern recognition capability to automatically distinguish between them. It is important to note that since online monitoring can only use data within a short time after the fault occurs, the difference between stable and unstable heat maps is not easily directly distinguishable by the naked eye, so it is necessary to introduce a deep learning model to mine subtle features.

[0027] Further, the convolutional neural network comprises a convolutional layer, a pooling layer and a fully connected layer; Convolutional layers extract features from the heatmap and output a feature map. The feature map is processed using a non-linear activation function to obtain activation values; Pooling layers downsample activation values ​​or extract local extrema to output high-dimensional features; The flattened high-dimensional features are input into a fully connected layer for pattern classification to obtain the category prediction results.

[0028] Furthermore, assigning category labels to each sliding window sample includes: labeling all windows during the stable operation period before the fault as class 0, labeling windows covering the moment the fault occurs as class I, labeling windows having all samples after the fault occurs but before the fault is cleared as class II, and labeling windows covering the moment the fault is cleared as class III.

[0029] Specifically, the construction and training of convolutional neural network classifiers: This embodiment uses a convolutional neural network (CNN) to classify heatmaps and determine the system state corresponding to the current sliding window. A CNN is a deep learning model adept at processing gridded topological data, efficiently extracting features through local perception and parameter sharing mechanisms. A typical CNN structure includes convolutional layers, pooling layers, and fully connected layers. A convolutional layer consists of multiple convolutional kernels, each sliding on the input feature map and performing convolution operations with local regions to extract local features; its output is called a feature map. By stacking multiple convolutional layers, more abstract high-level features can be extracted progressively. Let the weight vector and bias of the k-th convolutional kernel in the l-th layer be... and , No. Layer input at image location The local region vector at is The position of the output feature map of the convolutional layer The value can be: ; In the formula and To represent the first The weight vector and bias term of the k-th convolutional kernel in the layer; Indicated by position The input local region is centered. Here, a feature map is generated. convolution kernel For all locations All are the same. Then, they are processed by a nonlinear activation function. Get activation value , as the input to the next layer. The activation function can be Sigmoid, Tanh or ReLU, etc. The convolutional layer usually also connects a pooling layer, which down-samples or takes local extreme values of the feature map, reduces the feature dimension and thus improves the robustness. Finally, the high-dimensional features output by the last several convolutional / pooling layers are flattened and input into a fully connected network to realize the final classification and discrimination. The final output of the last layer of the entire network is usually obtained through the softmax operator: ; where is the output of the Jth neuron in the output layer, z is a vector of real-valued scores as the input to the last layer, and the function is called the softmax function, with components . The vector has all components between 0 and 1, and the sum of its components is 1.

[0030] Let denote the vector of all parameters of the CNN. Considering n data samples, denoted as , the optimal parameters are obtained by minimizing the empirical risk, defined as: ; where denotes the true or desired output, is the output obtained by the CNN on the input vector x, is the loss of the i-th sample. For classification problems, the cross-entropy loss is usually taken as .

[0031] Gradient descent and its variations are usually used to minimize the empirical risk. The gradient descent method updates the weights as follows: ; where denotes the parameter value at time step t, is the gradient of with respect to the parameter at t, denotes the step size. The gradient is efficiently calculated using the so-called backpropagation algorithm.

[0032] All windows in the pre-fault stable operating period are labeled as class 0. If a window covers the moment of fault occurrence, it is labeled as class I. If a window has all samples after the fault occurs but before the fault is cleared, it is labeled as class II. If a window covers the moment of fault clearance, it will be labeled as class III. In the current configuration file , if the system is stable after fault clearance, all windows after fault clearance are again classified as class 0; otherwise, all windows after fault clearance will be labeled as class IV. For unstable configuration files Each window marked as Class IV is also associated with the key generator set. Related. All other windows are assigned an empty one. In the training data of voltage distribution obtained through simulation, the instant of fault occurrence, fault clearance, etc., are known; this information will be used to mark the window. As described above, a window... , marked as : ; in, The end time of the time window, The initial time of the fault The start time of the time window, For fault clearing time, This is the transient stability coefficient.

[0033] The training data is generated from a large number of simulated fault samples, including different load levels, fault types and locations, topology changes, etc., to ensure that the model is robust to various operating conditions. In the training set, each sliding window sample is assigned a category label based on the simulation results for supervised learning.

[0034] The output layer of a CNN is a softmax normalization layer, which maps the output to probability values ​​for each class. This output must be converted into a prediction or alarm for instability. If, within N consecutive windows, the normalized score of Class-4 in the softmax layer output exceeds a certain threshold... This will trigger an alert indicating impending instability. Hyperparameters And N can be determined through cross-validation. In this embodiment, N = 15 and = 0.9.

[0035] The CNN architecture design in this embodiment is as follows: the convolutional feature extraction part includes two 2D convolutional layers and appropriate pooling layers, followed by a fully connected classification part. In the specific implementation, two convolutional layers are used to extract features, and then two fully connected layers are used for classification, as follows: Figure 2 As shown.

[0036] Furthermore, the key generator set identified using principal component analysis (PCA) includes: Select The time series data of frequency and phase angle of all generators within the time period constitutes the dimension of The data matrix, where, G represents the number of sampling points from fault clearing to the early warning interval, where G is the total number of generators. The initial moment of the fault. This is the moment when CNN first identified it as a Class 4 warning. right Principal component analysis (PCA) is performed on the data matrix to extract the directions of the first two principal components, and the data vector of each generator is projected onto the two principal component planes. A binary Gaussian distribution is used to fit and model the projection points. By utilizing the minimum covariance row and column, the generators corresponding to the projection points with the minimum likelihood values ​​under the fitted probability distribution are obtained, which are the key generator sets.

[0037] Furthermore, identifying the set of key generators using a dual-branch convolutional neural network includes: Features of the heatmap are extracted using a CNN feature extractor. The extracted features are then input into the first fully connected classification branch and the second fully connected classification branch, which output the circuit system state and the set of key generators, respectively.

[0038] Furthermore, the output layer of the second fully connected classification branch contains G+1 nodes and uses the Sigmoid activation function to output the probability indication of the key generator for each generator, where G is the total number of generators.

[0039] Specifically, key generator identification: After the system issues an instability warning, it is necessary to further identify the generators that first lost synchronization and were most affected by the disturbance during the instability incident—the critical generator set. This helps dispatchers take targeted stability control measures to prevent the instability from spreading. This embodiment provides two complementary methods for identifying critical generators: an offline method based on PCA and an online method based on a dual-branch neural network Y-Net.

[0040] (1) Principal Component Analysis (PCA) based method: This method uses the differences in the dynamic response of each generator during the period from the occurrence of a fault to the early warning determination to identify "outlier" units. The specific steps are as follows: Assume The initial moment of the fault. This is the moment when CNN first determined it to be a Class 4 warning. The time series data of frequency and phase angle of all generators within the time period constitutes the dimension of The data matrix, where This represents the number of sampling points between fault clearance and warning. Each row of the matrix corresponds to a generator's three-dimensional measurement sequence during that time period. Principal Component Analysis (PCA) is performed on this matrix to extract the directions of the first two principal components, and the data vector of each generator is projected onto these two principal component planes. At this point, each generator corresponds to a two-dimensional projected coordinate point. Subsequently, a bivariate Gaussian distribution is used to fit and model these projected points, utilizing the minimum covariance row and column. In the fitted probability distribution, the generator corresponding to the projected point with the minimum likelihood value under this distribution is identified as an outlier and defined as the "most critical generator." The two-dimensional projection map also allows for the visual identification of other possible outlier units. Therefore, the PCA-based method not only automatically identifies the earliest out-of-synchronization unit but also provides visualization tools to assist manual identification of multiple potentially unstable units, thereby reducing the risk of false alarms.

[0041] (2) Y-Net Neural Network-Based Method: To automatically detect all instability-related units, this embodiment further proposes a dual-branch convolutional neural network architecture, namely Y-Net, to simultaneously achieve system stability classification and critical generator set prediction. A schematic diagram of the Y-Net structure is shown below. Figure 3 As shown.

[0042] The convolutional layers of the CNN serve as shared feature extractors. After extracting the high-dimensional features through convolution, the network splits into two fully connected classification branches: one, Classifier-1, is the system state classifier mentioned above, and the other, Classifier-2, is used for predicting the set of key generators.

[0043] Unlike Classifier-1, Classifier-2's output layer contains G+1 nodes and uses the Sigmoid activation function to provide a probability indication of whether each generator belongs to a critical generator. The extra "+1" node represents "no instability". For any input window, Classifier-2 outputs a binary vector of length G+1, where generators with an output of 1 are identified as critical generators; if all G generators output 0 and only the last "no instability" node is 1, it indicates that there is currently no instability. Since multiple units may lose synchronization simultaneously in an instability situation, this problem belongs to multi-label classification, meaning a single window sample can correspond to multiple critical generators. To train Y-Net, in the offline simulation data, each window, in addition to a state label, is also assigned a corresponding multi-label vector as a supervision signal for the second branch based on the actual set of critical units obtained from the simulation. For example, the label vector of the unstable sample window is 1 at the position corresponding to the unstable unit, and 0 elsewhere; the stable sample window only has a "no instability" label of 1. Y-Net's loss function consists of two parts, derived from state classification and multi-label classification, respectively. The errors from these two parts are alternately backpropagated during training to update the parameters. This allows the trained Y-Net model to predict instability while simultaneously identifying the set of aircraft that caused the instability, thus providing a more comprehensive identification of all critical aircraft compared to the PCA method.

[0044] The method in this embodiment also includes system alarms and performance evaluation: When the CNN continuously determines that the system has entered a warning state and meets the alarm conditions, the system immediately issues an alarm signal and presents the list of critical generators identified by Y-Net to the operators so that measures can be taken (such as disconnecting the unit or reducing its output). Simultaneously, this embodiment introduces the following metrics when evaluating system performance: false alarm rate (FA), false alarm rate (MD), and Jaccard similarity. The FA rate and MD rate measure the frequency of false alarms under stable conditions and the frequency of false alarms under unstable conditions, respectively, and can be obtained through extensive simulation tests and statistics. The method proposed in this embodiment performs well on both metrics, with both the FA rate and MD rate below 0.5% under zero-noise conditions in a 118-node system. The Jaccard similarity J is used to evaluate the accuracy of the critical generator set prediction and is defined as: ; In the formula This is a set of key generators obtained from simulations of actual instability cases. This is the set predicted by Y-Net. The closer the J value is to 1, the more accurate the predicted set of key units.

[0045] The proportion of unstable samples with J=1 is calculated as the JS accuracy rate of the critical generator prediction. Additionally, the prediction delay is defined as the time difference between the first OMS alarm and the fault clearing, reflecting the system's ability to provide early warnings.

[0046] The verification model in this embodiment uses IEEE 118-node and 145-node standard power systems, including 19 and 50 generators respectively, to simulate three-phase grounding faults. The fault locations are different locations on each bus and transmission line.

[0047] This embodiment tested the performance of OMS under noise-free and baseline topologies, and verified its generalization ability under an N-1 line fault condition. Subsequently, a dataset containing ±5% uniform perturbation of generator parameters was constructed to evaluate the model's robustness under parameter uncertainty. Simultaneously, Type I noise (±1% to ±3% uniform random noise) was superimposed on the training and testing data to examine the impact of measurement error on performance. Finally, the OMS trained on the baseline topology with added noise was tested under topology change and parameter perturbation scenarios to verify its adaptability and robustness. The test results under the IEEE 118-node and IEEE 145-node systems are shown in Tables 1 and 2, respectively.

[0048] Table 1

[0049] Table 2

[0050] For the 118-node system, under baseline topology and noise-free conditions, both the FA rate and MD rate are below 0.5%. As noise levels increase, the FA rate and MD rate rise slightly; however, JS accuracy is less affected, showing only slight fluctuations. With increasing noise... Slightly increased Slightly lower.

[0051] Under the topology change condition, the FA rate and MD rate increased slightly compared with the baseline condition, and the trend of change was consistent. , Both the accuracy of JS and the accuracy of FA are slightly lower than the baseline operating condition, and the trend of change is the same; under the condition of generator parameter variation, the accuracy of FA and MD are slightly higher than the baseline operating condition, while , Even with further reductions, the accuracy of JS remains essentially unchanged, and its trend of change with increasing noise level is consistent with the baseline operating condition.

[0052] Under various operating conditions of the baseline topology, All are less than 7.4 cycles. All of these periods are greater than 25 cycles, indicating that OMS can identify instability within approximately 7.4 cycles after fault clearance and issue an instability warning at least 25 cycles in advance. The main purpose of introducing the waiting period is to reduce the false alarm rate; therefore, OMS needs at least [a certain number of cycles] cycles before the system actually becomes unstable. Detection is performed over several cycles. This is done at different waiting periods N and confidence levels. Under these conditions, the changes in the FA rate and MD rate are as follows: Figure 4 As shown in (a)-(d).

[0053] When the waiting period N increases from 0 to 30, the MD rate changes by approximately 2%, and the FA rate changes by approximately 10%. The results indicate that selection... This is a good compromise to reduce the false alarm rate. Therefore, considering the waiting period, the actual advance prediction time of the system exceeds 40 cycles.

[0054] The results for the IEEE 145-node system are similar to those for the IEEE 118-node system. This can be seen from the two systems studied. , JS accuracy is a system-related metric, while FA rate and MD rate do not change significantly under different system configurations. This is an ideal characteristic for a stability monitoring system, indicating that both maintain good stability and consistent performance characteristics under different configurations.

[0055] Furthermore, this embodiment also considers Type II noise, meaning that each generator experiences random noise measurement with a probability of ±1% to ±3% at any given time, with a probability of 0.1. The FA rate and MD rate of the 118-node and 145-node systems of the reference topology under different noise levels are as follows: Figure 5 As shown in (a)-(d).

[0056] Figure 5 In (a)-(d), the vertical axis represents the percentage of FA and MD, and the horizontal axis represents the noise level in the training data. Each noise level in the training data corresponds to the results of ±1%, ±2%, and ±3% noise in the test data. As can be seen from the figure, compared with Tables 1 and 2, for the 118-node and 145-node systems, the FA rate under Type II noise is slightly higher than that under Type I noise, meaning the system is slightly more sensitive to sudden disturbances; however, the MD rate does not change significantly, indicating that OMS still maintains good robustness in this type of random spike noise environment.

[0057] Compared with existing technologies, the method in this embodiment has significant advantages. First, it eliminates the need to pre-acquire fault occurrence and clearing times, as the system achieves continuous online monitoring through a sliding window, thus eliminating reliance on fault timing information. Second, by leveraging deep feature extraction from the voltage heatmap using CNN, it can accurately identify early signs of instability and provide rapid warnings, significantly improving prediction accuracy and response speed. Simultaneously, the system can synchronously identify key generators causing instability, providing targeted control basis for dispatching. Extensive simulation results demonstrate that this method exhibits excellent robustness under various load levels, topology changes, parameter disturbances, and measurement noise conditions. Overall, this embodiment outperforms traditional SVM methods in terms of false alarm rate, false negative rate, and accuracy in identifying key generators, demonstrating higher reliability and practicality.

[0058] To better understand the purpose, structure, and function of this embodiment, a method for online stability monitoring of power systems based on convolutional neural networks is described in further detail. The specific steps are as follows: Step 1: System data acquisition and preprocessing; Step 2: CNN Model Construction and Network Design; Step 3: Training dataset construction and label definition; Step 4: Set confidence thresholds and waiting periods to optimize false positive and false negative performance; Step 5: The system uses a sliding window to input CNN to judge stability in real time. When a continuous window is judged to be unstable and the confidence exceeds the threshold, an early warning is triggered and key units are identified. Step 6: The OMS demonstrates high accuracy and early warning capabilities; Step 7: Verify that the system's computational delay is in the microsecond range, meeting the real-time requirements of online monitoring and possessing engineering application potential for power grid stability early warning.

[0059] All measurement data is available during the training phase. Consider the following three scenarios: (1) Measurement data of 10% of generators are randomly missing, with a missing probability of 0.1 at each time step; (2) Measurement data of any stable generator is missing, and measurement data of that generator is missing at all times; (3) Measurement data of any unstable generator is missing, and measurement data of that generator is missing at all times. Table 3 shows the performance of OMS in the EE 118-node system and IEEE 145-node system scenarios.

[0060] Table 3

[0061] As can be seen from Table 3, OMS still exhibits strong robustness in the face of measurement loss.

[0062] Its performance under single-phase ground fault conditions was tested. Figure 6 The figure shows the 2D projection results of the 118-node and 145-node systems after PCA dimensionality reduction. Solid circles and dashed rectangles in the figure represent the key generators identified by PCA and Y-Net methods, respectively. Figure 6 (a) and Figure 6 In (c), there is only one outlier, therefore both PCA and Y-Net identify the same key generator. And... Figure 6 (b) contains two outliers. Although both are visible in the visualization, PCA can only numerically identify one of them, while Y-Net correctly identifies both key generators. Figure 6 (d) contains five outliers. PCA only identified one key generator, while Y-Net was able to identify all five.

[0063] Simulation results show that one advantage of the PCA method is that its visualization results can be used to help operators identify false alarm examples, thereby effectively reducing the actual false alarm rate of the OMS system.

[0064] Two stable samples that were misjudged as unstable by OMS are as follows: Figure 7 As shown in (a)-(b).

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for online monitoring of power system stability based on convolutional neural networks, characterized in that, include: Real-time acquisition and preprocessing of target data from all generator buses, including voltage amplitude, phase angle, and frequency; A heatmap is generated based on the preprocessed target data. A convolutional neural network is used to extract features and classify patterns from the heatmap to obtain category prediction results. The convolutional neural network is trained using a training set generated from simulated fault samples, and each sliding window sample is assigned a category label. Based on the prediction results of the aforementioned categories, determine whether it is an instability warning state; After determining that an instability warning state has been established, the set of key generators is identified using principal component analysis (PCA) or a two-branch convolutional neural network.

2. The online monitoring method for power system stability based on convolutional neural networks according to claim 1, characterized in that, Preprocessing of target data for all generator buses includes: A median filter is used to smooth the phase angle signal noise, and the arithmetic mean of the phase angles of all generators is used as the reference angle to obtain initial preprocessed data. The initial preprocessed data includes data points of voltage amplitude, phase angle and frequency of G generators obtained at each sampling time. The initial preprocessed data is analyzed using a sliding time window to obtain the preprocessed target data.

3. The online monitoring method for power system stability based on convolutional neural networks according to claim 2, characterized in that, Generating a heatmap based on the preprocessed target data includes: The data within the sliding window is represented as a matrix, and the elements in the matrix are mapped to the red, green, and blue channels of the image to obtain the heat map. Each element of the matrix contains a vector of the voltage amplitude, phase angle, and frequency of the generator at the corresponding sampling time.

4. The online monitoring method for power system stability based on convolutional neural networks according to claim 1, characterized in that, The convolutional neural network includes convolutional layers, pooling layers, and fully connected layers; The convolutional layer extracts features from the heatmap and outputs a feature map; The feature map is processed by a non-linear activation function to obtain activation values; The pooling layer downsamples the activation value or takes local extrema to output high-dimensional features; The flattened high-dimensional features are input into the fully connected layer for pattern classification to obtain the category prediction result.

5. The online monitoring method for power system stability based on convolutional neural networks according to claim 1, characterized in that, The category labels assigned to each sliding window sample are as follows: all windows during the stable operation period before the fault are labeled as class 0, windows covering the moment the fault occurs are labeled as class I, windows after the fault occurs but before the fault is cleared are labeled as class II, windows covering the moment the fault is cleared are labeled as class III, and all windows after the fault is cleared are labeled as class IV.

6. The online monitoring method for power system stability based on convolutional neural networks according to claim 1, characterized in that, The key generator set identified using Principal Component Analysis (PCA) includes: Select The time series data of frequency and phase angle of all generators within the time period constitutes the dimension of The data matrix, where, G represents the number of sampling points from fault clearing to the early warning interval, where G is the total number of generators. The initial moment of the fault. This is the moment when CNN first identified it as a Class 4 warning. Regarding the Principal component analysis (PCA) is performed on the data matrix to extract the directions of the first two principal components, and the data vector of each generator is projected onto the two principal component planes. A binary Gaussian distribution is used to fit and model the projection points. By utilizing the minimum covariance row and column, the generator corresponding to the projection point with the minimum likelihood value under the fitted probability distribution is obtained, which is the set of key generators.

7. The online monitoring method for power system stability based on convolutional neural networks according to claim 1, characterized in that, The key generator set identified using a dual-branch convolutional neural network includes: The features of the heatmap are extracted by a CNN feature extractor. The extracted features are then input into the first fully connected classification branch and the second fully connected classification branch, which output the circuit system state and the set of key generators, respectively.

8. The online monitoring method for power system stability based on convolutional neural networks according to claim 7, characterized in that, The output layer of the second fully connected classification branch contains G+1 nodes and uses the Sigmoid activation function to output a probability indication of the key generator for each generator, where G is the total number of generators.

Images