Diesel generator longitudinal protection method based on waveform similarity

Abstract

The present invention relates to the technical field of generator protection methods, and specifically to a diesel generator longitudinal protection method based on waveform similarity, comprising the following steps: deploying current transformers and voltage transformers at key nodes of the diesel generator to collect waveform data of each key node in real time; synchronously transmitting the waveform data collected from each key node via a wireless network communication channel, so that each protection device of the generator can exchange and share waveform data in real time; using a convolutional neural network to generate waveform feature vectors, the waveform feature vectors including waveform shape and amplitude information; calculating the waveform similarity between the waveform feature vectors of different key nodes to determine whether there is a waveform anomaly; performing fault type analysis on nodes with waveform anomalies; and executing a longitudinal protection action. The present invention improves the adaptability to complex power grid fault scenarios, and has significant advantages in handling multi-node concurrent faults.

Description

Technical Field

[0001] The present invention relates to the technical field of generator protection methods, and in particular to a diesel generator longitudinal protection method based on waveform similarity. Background Art

[0002] Diesel generator systems are widely used in industry, electricity, transportation and other fields. As important power supply equipment, ensuring their stable operation and rapid fault response is crucial. Longitudinal protection is a common protection method in power systems. It is usually used to determine whether the system operation is abnormal by real-time monitoring of electrical quantities between multiple nodes and to take timely protective measures. However, with the increasing complexity of power systems, traditional longitudinal protection methods face many challenges when dealing with complex fault scenarios.

[0003] First, most existing longitudinal protection methods rely on fixed fault current and voltage thresholds. When faults such as short circuit, open circuit or overload occur in the system, the protection action is triggered by current or voltage exceeding the limit. However, this method based on simple threshold judgment is difficult to cope with more complex fault modes and waveform changes. Especially in the case of concurrent faults at multiple nodes or small changes in waveform characteristics, the detection accuracy of traditional methods is low, and it is easy to make misjudgments or missed judgments, which cannot guarantee the stable operation of the system.

[0004] Secondly, traditional fault location methods usually rely on single-node waveform data and lack comprehensive analysis of multi-node waveforms. This method is very limited in the accuracy of positioning in complex power grid faults, such as line faults or when multiple nodes are affected simultaneously, which can easily lead to inaccurate fault location identification and thus affect the execution efficiency of protection actions.

[0005] Finally, traditional longitudinal protection systems lack an effective coordination mechanism when executing protection actions, resulting in inconsistent responses from various node protection devices and an inability to ensure synchronized protection measures when a fault occurs. Existing systems typically rely on fixed protection strategies and lack the flexibility to adjust the order and priority of protection actions based on fault type and location. This results in poor system response speed and effectiveness when a fault occurs, and can even trigger secondary faults. Summary of the Invention

[0006] The present invention provides a diesel generator longitudinal protection method based on waveform similarity.

[0007] The diesel generator longitudinal protection method based on waveform similarity includes the following steps:

[0008] S1. Multi-node waveform data acquisition: Current transformers and voltage transformers are deployed at key nodes of the diesel generator, including output ports, both ends of tie lines, and load nodes. Waveform data from each key node is collected in real time, including current and voltage waveforms. Waveform data acquired from each node is subjected to wavelet transform denoising.

[0009] S2. Synchronous Transmission and Communication of Waveform Data: Waveform data collected from each key node is synchronously transmitted via wireless network communication channels, enabling the generator's various protection devices to exchange and share waveform data in real time, ensuring timing synchronization between protection points within the system.

[0010] S3. Waveform feature vector generation: Based on the preprocessed waveform data, a convolutional neural network is used to generate waveform feature vectors. The waveform feature vectors include waveform shape and amplitude information. The feature vector generation process considers waveform changes at multiple time scales.

[0011] S4. Waveform Similarity Calculation: Using the Improved Dynamic Time Warping (IDTW) algorithm, combined with spatiotemporal dynamic correlation, the waveform similarity between the waveform feature vectors of different key nodes is calculated to determine whether there are waveform anomalies. If the waveform similarity between a node and other nodes is lower than a preset threshold, the node is considered to be faulty.

[0012] S5. Fault type and location determination: Based on the waveform similarity calculation results and a preset fault signature library, the system analyzes the fault type of nodes with abnormal waveforms, automatically identifying short circuit, open circuit, and overload fault types, and simultaneously determining the specific location of the fault.

[0013] S6. Execution of longitudinal protection actions: Once the fault location and type are determined, a signal is sent to the protection devices at other nodes through the longitudinal protection communication network to coordinate the simultaneous execution of corresponding protection actions by each protection device, including generator tripping, tie line disconnection, or load shedding. The priority and sequence of protection actions are adjusted according to the actual operating environment.

[0014] Optionally, the S3 specifically includes:

[0015] S31, build a multi-layer convolutional neural network. The network structure includes several convolutional layers, pooling layers, and fully connected layers. The convolutional layers extract local features of the waveform, including the shape and amplitude changes of the waveform. The pooling layers are used for dimensionality reduction, reducing feature redundancy and retaining key features.

[0016] S32 scans and extracts features from waveform data using convolution kernels (convolution windows of different sizes) at multiple time scales to ensure that the convolutional neural network captures the changing characteristics of the waveform at short and long time scales. The size of the convolution kernel is dynamically adjusted according to the sampling rate and frequency characteristics of the waveform signal to ensure that rich waveform information is obtained at different time scales.

[0017] S33, waveform shape and amplitude feature vector generation: After multiple layers of convolution and pooling, the local features of the waveform are extracted and combined layer by layer. Finally, the feature map extracted by the convolution layer will be processed by the fully connected layer to generate multiple feature vectors V containing waveform shape and amplitude information. The feature vector V not only captures the overall shape characteristics of the waveform, but also includes information on the fluctuation amplitude.

[0018] Optionally, the S3 also includes the introduction of a cross-layer feature fusion mechanism. The waveform features extracted by multi-scale convolution need to be fused to ensure that the waveform information at different time scales can comprehensively reflect the overall changes of the waveform. This is accomplished through cross-layer feature fusion. The short-scale features extracted by different convolution layers are fused with the long-scale features to improve the ability to analyze complex waveform changes and ensure the accuracy of waveform changes at different time scales. The fused comprehensive feature vector F fusion is expressed as:

[0019] F fusion =α1Y1+α2Y2+…+α n Y n , where α1, α2, …, α n are weight coefficients of features at different time scales. The weights are automatically adjusted through historical data training. Y1, Y2, ..., Y n They are the convolution outputs at different time scales respectively.

[0020] Optionally, the waveform similarity calculation in S4 specifically includes:

[0021] S41, input comprehensive feature vector: the waveform comprehensive feature vector F generated by each key node fusion As input;

[0022] S42, performing an improved dynamic time warping algorithm on each comprehensive feature vector:

[0023] D(i,j)=min{D(i-1,j),D(i,j-1),D(i-1,j-1)}+dist(F fusion1 (i),F fussion2 (j)), where D(i,j) is the minimum cumulative distance between the waveform feature vectors of two nodes, dist(F fusion1 (i),F fussion2(j)) is the local distance calculation between the feature vectors of two nodes at time i and j;

[0024] S43, introduces spatiotemporal dynamic association: Based on spatiotemporal dynamic association, combined with the spatial position of each node and the time series characteristics of the waveform data, dynamically adjusts the time weight w of the similarity calculation t and spatial weight w s , w t and w s Impact factors that represent temporal and spatial correlations:

[0025] D adjusted (i,j)=w t time diff(i,j)+w s D(i,j), where D adjusted (i,j) is the knot

[0026] The cumulative distance after spatiotemporal weight adjustment, w t is the time weight that is dynamically adjusted according to the waveform time difference, w s It is the spatial weight adjusted according to the spatial distance between nodes, and time diff represents the time step difference between two nodes at time i and time j;

[0027] S44, after calculating the waveform similarity between different nodes, the calculation result is compared with the preset similarity threshold θ. If the similarity of a node is lower than the threshold, that is: sim(F fusion1 ,F fusion2 )<θ, it is determined that the node has a waveform anomaly or fault.

[0028] Optionally, the time weight w t Adjusted based on the differences between waveforms at different time points, it is defined as: Among them, α is the adjustment parameter of the time difference, which is a constant used to control the influence of the time difference and determine the weight of the time difference on the final similarity calculation. Δt is the time difference between the two nodes, which is expressed as the offset of the two feature vectors on the time axis. Δt = |t i -t j |, where t i and t j is the corresponding time point of two nodes on the time axis, w t The larger the value, the smaller the impact of time difference on similarity; the smaller the value, the greater the impact of time difference on similarity.

[0029] Optionally, the spatial weight w s Adjusted according to the physical space distance between nodes, it is defined as:

[0030] Among them, β is the adjustment parameter of spatial difference, which is a constant used to control the influence of physical distance between nodes on similarity calculation. space Indicates the physical distance between two nodes.

[0031] Optionally, the S5 specifically includes:

[0032] A preset fault signature library includes waveform feature templates for diesel generator system fault types, including short circuit, open circuit, and overload. Each fault type in the fault signature library corresponds to a specific waveform feature, including the fluctuation pattern, duration, and amplitude change information of current and voltage. The fault signature library is formed through the analysis and induction of historical fault data and is continuously updated to cover new fault modes.

[0033] Abnormal judgment based on waveform similarity: According to the waveform similarity calculation results, when the waveform similarity Sim(F fusion1 ,F fusion2 ) is lower than the preset threshold θ, the waveform of the node is judged to be abnormal, and the abnormal waveform will be compared with the feature template in the fault feature library to analyze the specific fault type.

[0034] Optionally, the fault type is identified by a support vector machine classifier, and the waveform feature vector F of the abnormal node is fusion The waveform feature templates extracted from the fault feature library are used for training, and the support vector machine classifier automatically classifies the waveform features and identifies the fault type.

[0035] Optionally, S5 further includes comparing and analyzing the waveform similarities of multiple nodes, combining the spatial positions of each node and the differences in waveform characteristics, to determine the specific location of the fault, specifically including:

[0036] S51, multi-node waveform similarity comparison: Calculate the waveform similarity between every two nodes in the diesel generator system. By comparing the waveform characteristics of each node, the degree of similarity between them is determined. The more similar, the more consistent the waveform changes. The greater the difference, the greater the probability of waveform abnormality.

[0037] S52, waveform difference analysis combined with spatial position: Combined with the physical position between nodes, analyze the similarity changes of waveforms;

[0038] S53, Fault Location Inference: Infer the fault location based on waveform similarity and node location. If the waveform of a node is significantly different from that of other nodes, it indicates a fault at that node. If the waveforms of multiple adjacent nodes are different, it indicates a line or regional fault. If multiple distant nodes exhibit abnormalities simultaneously, it indicates a concurrent multi-node fault.

[0039] S54, waveform feature difference analysis: Analyze the differences between the waveform feature vectors of abnormal nodes. By comparing the changes in waveform features, verify the fault location. The greater the waveform feature difference, the more obvious the node waveform abnormality.

[0040] S55, final decision on fault location: Determine the specific location of the fault by integrating waveform similarity, spatial position and waveform feature differences.

[0041] Optionally, the S6 specifically includes:

[0042] S61, Fault Signal Generation and Transmission: When the fault type and location of a node are detected, a fault alarm signal is generated. The fault alarm signal includes the location information of the fault node, the fault type (such as short circuit, open circuit, overload), and priority information. The fault signal is transmitted in real time via the pilot protection communication network.

[0043] S62, coordinated transmission of the pilot protection communication network: Fault information is propagated to all protection devices via the pilot protection communication network, which is synchronized with the central control unit via a high-speed communication protocol (IEC 61850);

[0044] S63, triggering and synchronous execution of protection actions: After the fault information is transmitted to each node protection device, the corresponding protection action is automatically triggered according to the severity and type of the fault. The protection actions include:

[0045] Generator tripping: When the fault type involves generator overload or short circuit, the generator protection device performs a tripping operation to stop the generator output and prevent the fault from expanding;

[0046] Tie line disconnection: If a fault occurs on a tie line or affects lines between multiple nodes, the tie line protection device is triggered to disconnect the fault line and prevent current from being transmitted to other areas;

[0047] Load shedding: In severe fault scenarios, a load shedding signal will be sent to the load protection device based on load priority, prioritizing the shedding of non-critical loads.

[0048] Beneficial effects of the present invention:

[0049] The present invention adopts a fault detection method based on waveform similarity, and realizes high-precision fault detection through real-time collection and feature extraction of multi-node waveform data and an improved dynamic time warping algorithm combined with a spatiotemporal dynamic correlation model. Compared with the traditional waveform comparison method, the introduced multi-scale convolution kernel dynamic adjustment technology greatly improves the accuracy of waveform feature extraction, ensuring that it can cope with waveform changes at different time scales. By flexibly aligning the waveform feature vectors in time sequence, it can more sensitively capture waveform anomalies, improve the adaptability to complex power grid fault scenarios, and has significant advantages in dealing with multi-node concurrent faults.

[0050] The present invention combines waveform similarity analysis with comprehensive consideration of spatial position differences to achieve accurate fault location judgment. By analyzing the waveform differences between multiple nodes and combining them with information on physical spatial distance, the location of the fault can be quickly located, supporting the precise positioning of concentrated faults, line faults, and multi-node concurrent faults. The waveform feature difference analysis ensures that when the waveforms of multiple nodes are abnormal, the fault severity of each node can be comprehensively evaluated and the fault location can be further verified, thereby achieving more effective fault isolation and rapid response, and significantly reducing the scope of fault impact.

[0051] The present invention, once the fault type and location are determined, synchronously sends a fault signal to each node protection device through the longitudinal protection communication network, realizing automatic coordination of system protection actions. Through the high-speed communication protocol, it ensures the synchronous execution of protection actions such as generator tripping, tie line disconnection and load shedding, avoiding inconsistent responses between systems. It introduces a protection action priority management mechanism to ensure that the sequence and execution time of protection actions are optimized in different fault scenarios, thereby minimizing system damage and downtime. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0054] Figure 2 Schematic diagram of waveform similarity calculation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0056] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0057] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0058] like Figure 1-Figure 2 As shown, the diesel generator longitudinal protection method based on waveform similarity includes the following steps:

[0059] S1. Multi-node waveform data acquisition: Current transformers and voltage transformers are deployed at key nodes of the diesel generator, including output ports, both ends of tie lines, and load nodes. Waveform data from each key node is collected in real time. Waveform data, including current and voltage waveforms, is then subjected to wavelet transform denoising. The pre-processed waveform data is then used for feature extraction to generate waveform feature vectors for each key node.

[0060] A current transformer is installed at the output port of the diesel generator to monitor the output current waveform of the generator in real time. A voltage transformer is installed in the parallel circuit of the generator output port to obtain the output voltage waveform.

[0061] Current transformers are installed at both ends of the tie line to ensure that the current waveforms at both ends of the tie line can be monitored simultaneously. Voltage transformers are also used to collect voltage waveform data at both ends of the tie line to facilitate fault location and tie line anomaly judgment during subsequent similarity calculations.

[0062] A current transformer is installed at the load node to monitor changes in the load current, and a voltage transformer is installed at the load connection end to obtain real-time changes in the voltage waveform at the load end to ensure that the operating status and abnormal conditions of the load end can be accurately reflected;

[0063] S2. Synchronous transmission and communication of waveform data: The essence of longitudinal protection lies in the collaborative work between key nodes. Each node does not make protection decisions independently, but needs to cooperate with other nodes to ensure that when a node fails, other nodes can respond promptly and cooperate to perform protection actions. Synchronous transmission of waveform data allows each node to exchange information in real time, facilitates comprehensive analysis based on global waveform changes, and coordinates the protection devices of each node to ensure a rapid linkage response of the entire system. Therefore, the waveform data collected by each key node is synchronously transmitted through the wireless network communication channel, allowing the various protection devices of the generator to exchange and share waveform data in real time, ensuring the timing synchronization between the protection points within the system.

[0064] S3. Waveform Feature Vector Generation: Based on preprocessed waveform data, a convolutional neural network is used to generate waveform feature vectors. The waveform feature vectors include waveform shape and amplitude information. The feature vector generation process considers waveform variations at multiple time scales, improving the ability to respond to long- and short-term fluctuations.

[0065] S4. Waveform Similarity Calculation: Using the Improved Dynamic Time Warping (IDTW) algorithm, combined with spatiotemporal dynamic correlation, the waveform similarity between waveform feature vectors at different key nodes is calculated to determine whether there are waveform anomalies. If the waveform similarity between a node and other nodes falls below a preset threshold, the node is considered faulty. Compared to traditional similarity algorithms, this algorithm introduces spatiotemporal dynamic correlation parameters, which dynamically adjust the similarity threshold based on waveform changes in time and space, allowing for more accurate waveform anomaly determination.

[0066] S5. Fault type and location determination: Based on the waveform similarity calculation results and a preset fault signature library, the system analyzes the fault type of nodes with abnormal waveforms, automatically identifying short circuit, open circuit, and overload fault types, and simultaneously determining the specific location of the fault.

[0067] S6. Execution of longitudinal protection actions: Once the fault location and type are determined, a signal is sent to the protection devices at other nodes through the longitudinal protection communication network to coordinate the protection devices to simultaneously execute corresponding protection actions, including generator tripping, tie line disconnection, or load shedding. The longitudinal protection action needs to be completed within milliseconds to ensure rapid response to fault isolation. Unlike traditional fixed protection action schemes, the priority and sequence of protection actions are adjusted according to the actual operating environment to ensure that the system's operating risks and losses are minimized.

[0068] The wavelet transform denoising process specifically includes:

[0069] Decomposition of waveform data: The current waveform and voltage waveform data obtained from each node are input into the wavelet transform processing module, and the original waveform is decomposed into multi-scale using discrete wavelet transform. The waveform data is decomposed into different frequency band levels through wavelet basis functions to obtain detail signals and approximate signals. The higher frequency components contain noise information, while the lower frequency components retain the main characteristics of the waveform.

[0070] Threshold processing: Perform threshold processing on the decomposed high-frequency detail components to remove noise. Set a hard threshold and select the threshold size based on the characteristics of the waveform data and the noise level. High-frequency components smaller than the threshold will be regarded as noise and removed, while components larger than the threshold will be retained to reduce waveform distortion caused by noise.

[0071] Reconstruction of wavelet coefficients: After threshold processing, the retained high-frequency and low-frequency components are used to reconstruct the wavelet coefficients. The decomposed signal is reconstructed back to the time domain through inverse wavelet transform to obtain the denoised waveform signal, and the waveform data after denoising is restored, while retaining its main features and removing noise interference.

[0072] Waveform smoothing and refinement: The waveform data after wavelet transform denoising will be further smoothed, and multiple reconstruction and signal fusion techniques will be used to eliminate the edge effects caused by wavelet transform to ensure that the denoised waveform signal has high smoothness and integrity.

[0073] The processed waveform data will be input into the feature vector generation as clear current and voltage waveforms to generate waveform feature vectors for subsequent similarity calculation and fault judgment.

[0074] The synchronous transmission and communication of waveform data are as follows:

[0075] Wireless communication module deployment: Wireless communication modules are deployed at each key node of the diesel generator. The communication module of each node maintains real-time connection with the preset central control unit. The waveform data collected from each node is digitized by the local data processing unit and data packet processing is performed. Each data packet contains the identification, timestamp, waveform data, and waveform characteristic value information of the collection node. In order to ensure the consistency of the timing, the timestamp of the waveform data will be uniformly processed based on the global synchronized clock system (GPS synchronized clock); in order to improve the transmission efficiency and ensure the security of the data, the waveform data will be compressed after being packaged to reduce the size of the data packet. At the same time, the data packet will be processed by AE S encryption ensures that data is protected from external attacks or tampering during wireless network transmission. Data packets collected by each node are transmitted in real time to the central control unit via wireless communication channels. The central control unit receives waveform data from each node in real time and uses a clock synchronization mechanism to align the data from each node to ensure that data from different nodes can be compared and analyzed within a unified time frame. If there is delay or data loss in the network, the automatic retransmission mechanism (ARQ) is activated to ensure the integrity and accuracy of the waveform data. After receiving waveform data from all nodes, the central control unit shares and exchanges this data in real time among the various protection devices of the generator. Through the internal wireless mesh network, each protection device can simultaneously receive global waveform data.

[0076] S3 specifically includes:

[0077] S31, build a multi-layer convolutional neural network. The network structure includes several convolutional layers, pooling layers, and fully connected layers. The convolutional layers extract local features of the waveform, including the shape and amplitude changes of the waveform. The pooling layers are used for dimensionality reduction, reducing feature redundancy and retaining key features.

[0078] In convolutional neural networks, the convolution operation is the core computational step for feature extraction of input waveform data. For the input waveform data X and the convolution kernel W, the convolution operation is expressed as: Among them, X(i) represents the i-th data point of the waveform data, W(k) represents the k-th parameter of the convolution kernel, Y(i) is the output after the convolution operation, reflecting the local waveform characteristics, K is the size of the convolution kernel, that is, the length of the waveform data covered by one operation. In the present invention, the convolution operation is applied to the preprocessed current and voltage waveform data, with the aim of extracting the local shape and amplitude characteristics of the waveform. Through convolution kernels of different sizes, the changes of the waveform on different time scales can be captured. In order to ensure that the features after convolution retain valid information, the ReLU activation function is introduced. The activation function sets negative numbers to zero after the convolution operation and retains non-negative eigenvalues ​​to ensure that the extracted waveform features are valid.

[0079] S32 scans and extracts features from waveform data using convolution kernels (convolution windows of different sizes) at multiple time scales, ensuring that the convolutional neural network captures the changing characteristics of the waveform at both short and long time scales. The size of the convolution kernel is dynamically adjusted based on the sampling rate and frequency characteristics of the waveform signal to ensure that rich waveform information is obtained at different time scales.

[0080] In the present invention, the size of the convolution kernel should be dynamically adjusted according to the sampling rate and frequency characteristics of the waveform signal. The frequency characteristics and sampling rate of the waveform signal determine the distribution of different frequency components in the signal. For high-frequency components, a smaller convolution kernel is required to capture rapidly changing features; while for low-frequency components, a larger convolution kernel is required to capture slower changes.

[0081] Sampling rate: The sampling rate of waveform data determines the temporal resolution of the signal, that is, the number of data points collected per second. A higher sampling rate means that the signal can reflect more subtle fluctuations. Therefore, for data with a high sampling rate, a smaller convolution kernel can be used to capture rapid waveform changes.

[0082] Frequency characteristics: The frequency characteristics of a waveform signal determine the frequency range in which the main energy of the waveform is concentrated. If the high-frequency component dominates the signal, the convolution kernel size should be smaller to accurately capture these rapidly changing features; for low-frequency signals, a larger convolution kernel is required to capture long-term slow changes.

[0083] Dynamically adjust the convolution kernel size as follows:

[0084] Perform spectral analysis on the collected waveform data to calculate the dominant frequency distribution of the signal. Use Fast Fourier Transform (FFT) to obtain the spectrum of the signal and identify the frequency ranges where the energy in the signal is mainly concentrated. Based on the results of the spectral analysis, determine the characteristics of different frequency components, and then assign a suitable convolution kernel size to each frequency range.

[0085] Assume the sampling rate of the waveform signal is f s , the frequency characteristics of the signal are determined by the frequency range f min to f max Definition: The size K of the convolution kernel is dynamically adjusted, expressed as: Among them, f s is the sampling rate of the waveform signal, f target It is the target frequency component that needs to be captured in the signal;

[0086] In order to capture the characteristics of different frequency ranges, the frequency characteristics of the signal are analyzed and multiple target frequencies f are selected. target1 ,f target2 ,…,f targetn, and use different convolution kernel sizes for each target frequency.

[0087] For high frequency components, choose a smaller K, such as In order to capture the details of fast fluctuations; for low-frequency components, choose a larger K, such as To capture long-term waveform trends.

[0088] S33, waveform shape and amplitude feature vector generation: After multiple layers of convolution and pooling, the local features of the waveform are extracted and combined layer by layer. Finally, the feature map extracted by the convolution layer will be processed by the fully connected layer to generate multiple feature vectors V containing waveform shape and amplitude information. The feature vector V not only captures the overall shape characteristics of the waveform, but also includes information on the amplitude of the fluctuation, ensuring that it can fully reflect the changes in the waveform at each time point. The generation expression of the feature vector V is as follows: V = W fc ·Y pool +b, where W fc is the weight matrix of the fully connected layer, b is the bias term, and Y pool It is the waveform feature map after convolution and pooling.

[0089] S3 also includes the introduction of a cross-layer feature fusion mechanism. The waveform features extracted by multi-scale convolution need to be fused to ensure that the waveform information at different time scales can comprehensively reflect the overall changes of the waveform. This is achieved through cross-layer feature fusion. The short-scale features extracted by different convolution layers are fused with the long-scale features to improve the ability to analyze complex waveform changes and ensure the accuracy of waveform changes at different time scales. The fused comprehensive feature vector F fusion is expressed as:

[0090] F fusion =α1Y1+α2Y2+…+α n Y n , where α1, α2, …, α n are weight coefficients of features at different time scales. The weights are automatically adjusted through historical data training. Y1, Y2, ..., Y n They are the convolution outputs of different time scales. Through feature fusion, the important features of the waveform at multiple time scales are effectively extracted, and a comprehensive feature vector is generated for subsequent waveform similarity calculation and fault identification.

[0091] The waveform similarity calculation in S4 specifically includes:

[0092] S41, input comprehensive feature vector: the waveform comprehensive feature vector F generated by each key node fusion As input;

[0093] S42, performing an improved dynamic time warping algorithm on each comprehensive feature vector:

[0094] D(i,j)=min{D(i-1,j),D(i,j-1),D(i-1,j-1)}+dist(F fusion1 (i),F fussion2 (j)), where D(i,j) is the minimum cumulative distance between the waveform feature vectors of two nodes, dist(F fusion1 (i),F fussion2 (j)) is the local distance calculation between the feature vectors of two nodes at time i and j, using the Euclidean distance metric. The algorithm allows the waveform feature vectors to be flexibly aligned in time series by dynamically adjusting the time axis to capture the temporal and spatial feature differences of the waveforms of each node;

[0095] S43, introduces spatiotemporal dynamic association: Based on spatiotemporal dynamic association, combined with the spatial position of each node and the time series characteristics of the waveform data, dynamically adjusts the time weight w of the similarity calculation t and spatial weight w s , w t and w s Impact factors that represent temporal and spatial correlations:

[0096] D adjusted (i,j)=w t time diff(i,j)+w s D(i,j), where D adjusted (i, j) is the cumulative distance after adjusting the spatiotemporal weights, w t is the time weight that is dynamically adjusted according to the waveform time difference, w s The spatial weight is adjusted based on the spatial distance between nodes. Time diff represents the time step difference between two nodes at time i and time j, and is used to measure whether two feature vectors are aligned on the time axis. Generally, the greater the time step difference, the lower the waveform similarity. By combining spatiotemporal features, waveform changes across time and nodes can be more accurately captured.

[0097] S44, after calculating the waveform similarity between different nodes, the calculation result is compared with the preset similarity threshold θ. If the similarity of a node is lower than the threshold, that is: sim(F fusion1 ,F fusion2 )<θ, it is judged that the node has abnormal waveform or fault, sim(F fusion1 ,F fusion2 ) is the comprehensive feature vector F of the two nodes fusion1 ,F fusion2 The similarity between

[0098] Preset similarity threshold It is an important criterion for judging whether the waveforms of two nodes are similar. The similarity of the waveforms between different nodes is calculated under normal and fault conditions, and a reasonable threshold is set based on the experimental results.

[0099] Under normal operating conditions: waveform data under normal conditions is collected and similarity is calculated. Usually, the similarity value of this data will be high (close to 1). Based on this normal data, a high similarity threshold that can identify normal waveforms is selected.

[0100] Under fault conditions: Similarly, waveform data from the fault is collected and similarity is calculated. This data typically has a low similarity value (close to 0). By analyzing this data, an appropriate threshold is set to ensure that the similarity is below the threshold when the waveform is abnormal, triggering a fault judgment.

[0101] Time weight w t Adjusted based on the differences between waveforms at different time points, it is defined as: Among them, α is the adjustment parameter of the time difference, which is a constant used to control the influence of the time difference and determine the weight of the time difference on the final similarity calculation. Δt is the time difference between the two nodes, which is expressed as the offset of the two feature vectors on the time axis. Δt = |t i -t j |, where t i and t j is the corresponding time point of two nodes on the time axis, w t The larger the value, the smaller the impact of time difference on similarity; the smaller the value, the greater the impact of time difference on similarity.

[0102] Spatial weight w s Adjusted according to the physical space distance between nodes, it is defined as:

[0103] Among them, β is the adjustment parameter of spatial difference, which is a constant used to control the influence of physical distance between nodes on similarity calculation. space Represents the physical distance between two nodes, measured by their actual geographic locations, which is similar to the control parameter of the time weight, and determines the degree of influence of spatial differences on the final similarity. s The larger the value, the smaller the impact of spatial distance on similarity; the smaller the value, the greater the impact of spatial distance on similarity.

[0104] Through a large number of experiments and data set verification, the values ​​of α and β are adjusted to obtain the best results in actual scenarios. The specific steps are as follows:

[0105] 1. Dataset preparation: By collecting waveform data under different operating conditions, a set of training and validation datasets is formed. The dataset should contain data from normal operation and data under fault scenarios. The corresponding fault type and spatial location should also be marked in the dataset to ensure that the accuracy of fault detection can be evaluated by adjusting parameters.

[0106] 2. Parameter adjustment process: Experiment with different values ​​of α and β. Start with smaller values, such as α = 0.1 and β = 0.1, and then gradually increase them. Observe the similarity calculation results of the system and the effectiveness of fault detection. In each experiment, evaluate the false alarm rate and the missed alarm rate. Ideally, adjust these parameters to find a balance point so that both the false alarm rate and the missed alarm rate are low.

[0107] 3. Grid search: Using the grid search method, we systematically search for α and β within a predetermined range to find the optimal combination. For example:

[0108] Set the value range of α to [0.1, 0.5, 1.0, 1.5];

[0109] Set the value range of β to [0.1, 0.5, 1.0, 1.5];

[0110] Perform similarity calculation and fault detection under each combination, observe the performance changes, and select the combination with the best false alarm rate and false negative rate as the final parameter value.

[0111] S5 specifically includes:

[0112] A preset fault signature library includes waveform feature templates for diesel generator system fault types, including short circuit, open circuit, and overload. Each fault type in the fault signature library corresponds to a specific waveform feature, including the fluctuation pattern, duration, and amplitude change information of current and voltage. The fault signature library is formed through the analysis and induction of historical fault data and is continuously updated to cover new fault modes.

[0113] Abnormal judgment based on waveform similarity: According to the waveform similarity calculation results, when the waveform similarity Sim(F fusion1 ,F fusion2 ) is lower than the preset threshold θ, the waveform of the node is judged to be abnormal, and the abnormal waveform will be compared with the feature template in the fault feature library to analyze the specific fault type.

[0114] The fault type is identified by the support vector machine classifier, and the waveform feature vector F of the abnormal node is fusion The waveform feature template extracted from the fault feature library is trained, and the support vector machine classifier automatically classifies the waveform features and identifies the fault type. The discriminant formula of the support vector machine classifier is as follows:

[0115]

[0116] in, represents the identified fault type label, α i is the Lagrange multiplier of the support vector machine, used for the support vector weight, y i is the true label of the fault type, is the waveform feature F fusion Templates in the fault signature library The kernel function, Gaussian kernel or linear kernel, is used to measure the similarity between the current waveform and the fault template, and b is the bias term of the support vector machine.

[0117] S5 also includes comparing and analyzing the waveform similarities of multiple nodes, combining the spatial locations of each node and the differences in waveform characteristics to determine the specific location of the fault. By analyzing the degree of waveform anomaly of multiple nodes, it is determined whether the fault is concentrated at a certain node or spreads across a certain section of the line or area. Specifically, it includes:

[0118] S51, multi-node waveform similarity comparison: Calculate the waveform similarity between every two nodes in the diesel generator system. By comparing the waveform characteristics of each node, the degree of similarity between them is determined. The more similar, the more consistent the waveform changes. The greater the difference, the greater the probability of waveform abnormality.

[0119] S52, waveform difference analysis combined with spatial position: Combined with the physical position between nodes, analyze the similarity changes of waveforms;

[0120] S53, Fault Location Inference: Infer the fault location based on waveform similarity and node location. If the waveform of a node is significantly different from that of other nodes, it indicates a fault at that node. If the waveforms of multiple adjacent nodes are different, it indicates a line or regional fault. If multiple distant nodes exhibit abnormalities simultaneously, it indicates a concurrent multi-node fault.

[0121] S54, waveform feature difference analysis: Analyze the differences between the waveform feature vectors of abnormal nodes. By comparing the changes in waveform features, verify the fault location. The greater the waveform feature difference, the more obvious the node waveform abnormality.

[0122] S55, final decision on fault location: Determine the specific location of the fault by integrating waveform similarity, spatial position and waveform feature differences.

[0123] The detailed plan is as follows:

[0124] 1. Multi-node waveform similarity comparison, there are N nodes, their comprehensive waveform feature vectors are:

[0125] Ffusion1 ,F fusion2 ,…,F fusionN For each pair of nodes, the waveform similarity is calculated to evaluate whether their waveform changes are consistent. The similarity calculation formula in S44 is: in;

[0126] Sim(F fusioni ,F fusionj ) represents the waveform similarity between node i and node j;

[0127] D adjusted (i, j) is the cumulative minimum distance after combining the spatiotemporal weights, which is used to measure the waveform difference between node i and node j (see the formula given above);

[0128] By calculating the similarity matrix S between every two nodes, we can get:

[0129]

[0130] Each matrix element Sim(F fusioni ,F fusionj ) represent the similarity of the waveforms of two nodes. The closer the value is to 1, the more similar the waveforms of the two nodes are, and the closer the value is to 0, the greater the difference in the waveforms.

[0131] 2. In order to further combine the physical space location information of the node, the spatial distance matrix D of the node is introduced space , where D space (i, j) represents the physical distance between nodes i and j. A comprehensive analysis matrix C is constructed using the following formula, combining waveform similarity and spatial distance to determine the fault location: Among them, C(i,j) is a comprehensive value combining waveform similarity and spatial distance, which is used to describe the correlation between the waveforms of two nodes. If the waveforms of two nodes are very different and the spatial distance is far, C(i,j) will be small; if the waveforms are similar and the distance is close, C(i,j) will be large.

[0132] 3. Inference of fault location: By analyzing the values ​​in the comprehensive matrix C, the location of the fault can be inferred. The following are the inference rules:

[0133] 3.1 Single-node concentrated failure: If a node has low similarity with multiple other nodes, that is: And the spatial distance between this node and other nodes is not large (ie D space (i,j) is smaller), then it is inferred that the fault is concentrated on this node;

[0134] 3.2 Extended fault in a certain line or area:

[0135] If the waveform similarity of multiple adjacent nodes is low at the same time, that is:

[0136] sim(F fusioni ,F fusioni+1 )<θ and sim(F fusioni+1 ,F fusioni+2 )<θ, indicating that the fault occurs between these adjacent nodes and is a line fault or a regional fault;

[0137] 3.3 Concurrent failures of multiple nodes:

[0138] If multiple nodes in different regions have low similarity at the same time, that is:

[0139] sim(F fusioni ,F fusionj )<θ and sim(F fusionk ,F fusionl )<θ, and the physical distance between these nodes is far (i.e., D space (i,j) is larger), it is judged that multiple nodes fail at the same time, F fusionk is the waveform comprehensive feature vector of node k, F fusionl is the waveform comprehensive feature vector of node l.

[0140] 4. Determine the fault location by combining the waveform feature difference: When determining the fault location, the location can be further verified by the difference in waveform feature vectors. For each node’s comprehensive feature vector F fusioni, , analyze the difference between its eigenvector and other nodes:

[0141] ΔF ij =F fusioni -F fusionj , where ΔF ij Indicates the difference in waveform characteristics between node i and node j. If ΔF ij The value is large, and the similarity Sim(F fusioni ,F fusionj ) is low, it means that the waveform anomaly at node i is more serious, and it is inferred that the failure probability of this node is greater.

[0142] In general, the comprehensive waveform similarity matrix S and spatial distance matrix D space And the characteristic difference analysis ΔF is used to determine the specific location of the fault through the following decision process:

[0143] First, by analyzing the nodes in the similarity matrix whose values ​​are lower than the threshold θ, we determine which nodes may be abnormal.

[0144] Then, based on the spatial distance, we determine whether these abnormal nodes are concentrated in a certain physical area;

[0145] Finally, the specific node or line location of the fault is determined by analyzing the differences in the waveform characteristic vectors.

[0146] S6 specifically includes:

[0147] S61, Fault Signal Generation and Transmission: When the fault type and location of a node are detected, a fault alarm signal is generated. The fault alarm signal includes the location information of the fault node, the fault type (such as short circuit, open circuit, overload), and priority information. The fault signal is transmitted in real time via the pilot protection communication network.

[0148] S62, Coordinated transmission of the pilot protection communication network: Fault information is propagated to all protection devices via the pilot protection communication network. This is synchronized with the central control unit via a high-speed communication protocol (IEC 61850). The fault signal received by each protection device includes a synchronized clock to ensure that all system nodes respond to the fault signal at the same time, avoiding delays or malfunctions.

[0149] S63, triggering and synchronous execution of protection actions: After the fault information is transmitted to each node protection device, the corresponding protection action is automatically triggered according to the severity and type of the fault. The protection actions include:

[0150] Generator tripping: When the fault type involves generator overload or short circuit, the generator protection device performs a tripping operation to stop the generator output and prevent the fault from expanding;

[0151] Tie line disconnection: If a fault occurs on a tie line or affects lines between multiple nodes, the tie line protection device is triggered to disconnect the fault line and prevent current from being transmitted to other areas;

[0152] Load shedding: In the event of a serious fault, a load shedding signal will be sent to the load protection device based on load priority, with non-critical loads being shelved first.

[0153] Coordinate the order and priority of executing protection actions:

[0154] Each protection device determines the corresponding action sequence and priority based on the fault signal it receives. For example, when a generator fault occurs, the tripping operation is triggered first, and then the need to further disconnect the interconnecting line or remove the load is determined based on the situation. All protection actions are executed under the unified coordination of the system to ensure fault isolation and safe system operation.

[0155] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0156] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A diesel generator longitudinal protection method based on waveform similarity, characterized in that: The following steps are involved: S1. Multi-node waveform data acquisition: Current transformers and voltage transformers are deployed at key nodes of the diesel generator, including output ports, both ends of tie lines, and load nodes. Waveform data from each key node is collected in real time, including current and voltage waveforms. Waveform data acquired from each node is subjected to wavelet transform denoising. S2. Synchronous Transmission and Communication of Waveform Data: Waveform data collected from each key node is synchronously transmitted via wireless network communication channels, enabling the various protection devices of the generator to exchange and share waveform data in real time; S3. Waveform feature vector generation: Based on the preprocessed waveform data, a convolutional neural network is used to generate waveform feature vectors. The waveform feature vectors include waveform shape and amplitude information. The feature vector generation process considers waveform changes at multiple time scales. Specifically, the following steps are performed: S31, build a multi-layer convolutional neural network. The network structure includes several convolutional layers, pooling layers, and fully connected layers. The convolutional layers extract local features of the waveform, including the shape and amplitude changes of the waveform. The pooling layers are used for dimensionality reduction and feature redundancy reduction. S32 scans and extracts features from waveform data using convolution kernels at multiple time scales, ensuring that the convolutional neural network captures the changing characteristics of the waveform at both short and long time scales. The size of the convolution kernel is dynamically adjusted based on the sampling rate and frequency characteristics of the waveform signal. S33, waveform shape and amplitude feature vector generation: The feature map extracted by the convolution layer will be processed by the fully connected layer to generate multiple feature vectors V containing waveform shape and amplitude information. The feature vector V not only captures the overall shape characteristics of the waveform, but also includes information about the amplitude of the fluctuation; S4. Waveform Similarity Calculation: Using an improved dynamic time warping algorithm, combined with spatiotemporal dynamic correlation, the waveform similarity between the waveform feature vectors of different key nodes is calculated to determine whether there are waveform anomalies. If the waveform similarity between a node and other nodes is lower than a preset threshold, the node is judged to be faulty. Specifically, this includes: S41, input comprehensive feature vector: the waveform comprehensive feature vector F generated by each key node fusion As input; S42, performing an improved dynamic time warping algorithm on each comprehensive feature vector: D(i,j)=min{D(i-1,j),D(i,j-1),D(i-1,j-1)}+ dist(F fusion1 (i),F fussion2 (j)), where D(i,j) is the minimum cumulative distance between the waveform feature vectors of two nodes, dist(F fusion1 (i),F fussion2 (j)) is the local distance calculation between the feature vectors of two nodes at time i and j; S43, introduces spatiotemporal dynamic association: Based on spatiotemporal dynamic association, combined with the spatial position of each node and the time series characteristics of the waveform data, dynamically adjusts the time weight w of the similarity calculation t and spatial weight w s , w t and w s Impact factors that represent temporal and spatial correlations: D adjusted (i,j) = w t ·time diff(i,j) + w s ·D(i,j), where D adjusted (i, j) is the cumulative distance after adjusting the spatiotemporal weights, w t is the time weight that is dynamically adjusted according to the waveform time difference, w s It is the spatial weight adjusted according to the spatial distance between nodes, and time diff represents the time step difference between two nodes at time i and time j; S44, after calculating the waveform similarity between different nodes, the calculation result is compared with the preset similarity threshold θ. If the similarity of a node is lower than the threshold, that is: sim(F fusion1 ,F fusion2 )<θ, it is judged that the node has abnormal waveform or fault; S5. Fault type and location determination: Based on the waveform similarity calculation results and a preset fault signature library, the system analyzes the fault type of nodes with abnormal waveforms, automatically identifying short circuit, open circuit, and overload fault types, and simultaneously determining the specific location of the fault. S6. Execution of longitudinal protection actions: Once the fault location and type are determined, a signal is sent to the protection devices at other nodes through the longitudinal protection communication network to coordinate the simultaneous execution of corresponding protection actions by each protection device, including generator tripping, tie line disconnection, or load shedding. The priority and sequence of protection actions are adjusted according to the actual operating environment.

2. The diesel generator longitudinal protection method based on waveform similarity according to claim 1 is characterized in that: The S3 also includes the introduction of a cross-layer feature fusion mechanism to fuse the short-scale features extracted by different convolutional layers with the long-scale features. The fused comprehensive feature vector F fusion is expressed as: F fusion =α1Y1+α2Y2+…+α n Y n , where α1, α2, …, α n are weight coefficients of features at different time scales. The weights are automatically adjusted through historical data training. Y1, Y2, ..., Y n They are the convolution outputs at different time scales respectively.

3. The diesel generator longitudinal protection method based on waveform similarity according to claim 1 is characterized in that: The time weight w t Adjusted based on the differences between waveforms at different time points, it is defined as: Among them, α is the adjustment parameter of the time difference, which is a constant, Δt is the time difference between the two nodes, which is expressed as the offset of the two feature vectors on the time axis, Δt=|t i -t j |, where t i and t j It is the corresponding time point of two nodes on the timeline.

4. The diesel generator longitudinal protection method based on waveform similarity according to claim 1 is characterized in that: The spatial weight w s Adjusted according to the physical space distance between nodes, it is defined as: Among them, β is the adjustment parameter of spatial difference, which is a constant used to control the influence of physical distance between nodes on similarity calculation. spac Indicates the physical distance between two nodes.

5. The diesel generator longitudinal protection method based on waveform similarity according to claim 4 is characterized in that: The S5 specifically includes: A preset fault feature library includes waveform feature templates for diesel generator system fault types, including short circuit, open circuit, and overload. Each fault type in the fault feature library corresponds to a specific waveform feature, including the fluctuation pattern, duration, and amplitude change information of current and voltage. The fault feature library is formed by analyzing and summarizing historical fault data; Abnormal judgment based on waveform similarity: According to the waveform similarity calculation results, when the waveform similarity Sim(F fusion1 ,F fusion2 ) is lower than the preset threshold θ, the waveform of the node is judged to be abnormal, and the abnormal waveform will be compared with the feature template in the fault feature library to analyze the specific fault type.

6. The diesel generator longitudinal protection method based on waveform similarity according to claim 5 is characterized in that: The fault type is identified by the support vector machine classifier, and the waveform feature vector F of the abnormal node is fusion The waveform feature templates extracted from the fault feature library are used for training, and the support vector machine classifier automatically classifies the waveform features and identifies the fault type.

7. The diesel generator longitudinal protection method based on waveform similarity according to claim 1, characterized in that: S5 also includes comparing and analyzing the waveform similarities of multiple nodes, combining the spatial positions of each node and the differences in waveform characteristics, to determine the specific location of the fault, specifically including: S51, multi-node waveform similarity comparison: Calculate the waveform similarity between every two nodes in the diesel generator system. By comparing the waveform characteristics of each node, the degree of similarity between them is determined. The more similar, the more consistent the waveform changes. The greater the difference, the greater the probability of waveform abnormality. S52, waveform difference analysis combined with spatial position: Combined with the physical position between nodes, analyze the similarity changes of waveforms; S53, Fault Location Inference: Infer the fault location based on waveform similarity and node location. If the waveform of a node is significantly different from that of other nodes, it indicates a fault at that node. If the waveforms of multiple adjacent nodes are different, it indicates a line or regional fault. If multiple distant nodes exhibit abnormalities simultaneously, it indicates a concurrent multi-node fault. S54, waveform feature difference analysis: Analyze the differences between the waveform feature vectors of abnormal nodes. By comparing the changes in waveform features, verify the fault location. The greater the waveform feature difference, the more obvious the node waveform abnormality. S55, final decision on fault location: Determine the specific location of the fault by integrating waveform similarity, spatial position and waveform feature differences.

8. The diesel generator longitudinal protection method based on waveform similarity according to claim 1, characterized in that: The S6 specifically includes: S61, Fault Signal Generation and Transmission: When the fault type and location of a node are detected, a fault alarm signal is generated. The fault alarm signal includes the location information, fault type, and priority information of the fault node. The fault signal is transmitted in real time via the pilot protection communication network. S62, coordinated transmission of the pilot protection communication network: Fault information is propagated to all protection devices via the pilot protection communication network, which is synchronized with the central control unit via a high-speed communication protocol; S63, triggering and synchronous execution of protection actions: After the fault information is transmitted to each node protection device, the corresponding protection action is automatically triggered according to the severity and type of the fault. The protection actions include: Generator tripping: When the fault type involves generator overload or short circuit, the generator protection device performs a tripping operation to stop the generator output and prevent the fault from expanding; Tie line disconnection: If a fault occurs on a tie line or affects lines between multiple nodes, the tie line protection device is triggered to disconnect the fault line and prevent current from being transmitted to other areas; Load shedding: In severe fault scenarios, a load shedding signal will be sent to the load protection device based on load priority, prioritizing the shedding of non-critical loads.

Images