A fault classification and adaptive relay protection coordination optimization method, system and medium based on time sequence imaging and convolutional neural network for a "double-high" type new power distribution network

Abstract

The application discloses a fault processing technology for a "double-high" type novel power distribution network. The method first constructs a fault database covering multiple scenes based on short-circuit calculation and a fault sample generation mechanism; then, voltage / current time series are normalized and polar coordinate mapping is performed, and Gramian Angular Summation Field (GASF) is used to convert the time series into a two-dimensional image; further, a convolutional neural network is used to extract multi-channel image features and output fault categories; finally, according to the fault categories and short-circuit current levels, the pickup current and time scale coefficient of the directional overcurrent relay are updated online, and a nonlinear optimization model is established under the condition of constraining the coordination time interval of the main and standby relays to solve the optimal relay setting value of the minimum total action time. Compared with the traditional fixed setting value scheme, the application can significantly improve the fault classification accuracy and protection action speed of the power distribution network and realize adaptive protection coordination under a complex source-network-load-storage scene.

Description

Technical Field

[0001] This invention belongs to the field of distribution network relay protection and intelligent fault diagnosis technology, and particularly relates to a fault classification and adaptive relay protection coordination optimization method, system, device and computer-readable storage medium for a new type of distribution network with "high proportion of distributed power sources and high proportion of power electronic equipment". Specifically, it involves the coordination optimization of timing signal imaging (GAF / GASF), convolutional neural network (CNN) classification and directional overcurrent relay (DOCR). Background Technology

[0002] With the large-scale integration of distributed power sources (such as wind and solar power), the operation mode of distribution networks is becoming increasingly complex. In the new "high-voltage and high-efficiency" distribution networks, the characteristics of faults become more dynamic and complex due to the complexity of the interweaving and mixing of power sources, grids, loads, and storage, as well as the integration of AC and DC distribution networks. Traditional power system protection methods mostly rely on fixed parameter settings and cannot cope with such complex and dynamic fault conditions.

[0003] In existing technologies, fault identification and relay protection coordination are often two separate links: the fault identification part focuses on waveform characteristic engineering or a single classifier; the relay protection part focuses on fixed curves and heuristic setting. The lack of a closed-loop linkage of "identification-setting-coordination" leads to excessively long protection action times or coordination failures under complex operating conditions. Therefore, how to efficiently and accurately classify faults in the power system and optimize relay parameters based on fault types is a key issue in improving the performance of distribution network protection.

[0004] To address these issues, this invention proposes an adaptive relay protection technology based on convolutional neural network (CNN) and Gramian angular field (GAF) image transformation. This technology uses deep learning to classify faults with high accuracy and improves the protection efficiency and reliability of power systems by dynamically adjusting relay parameters. Summary of the Invention

[0005] To address the challenges posed by the complex and varied fault characteristics in high-voltage, high-load, and high-energy-consumption (HVDC) new distribution networks, resulting from source-grid-load-storage interleaving and AC / DC hybrid structures, including inaccurate fault type identification, mismatch between relay settings and fault currents, and difficulty in meeting the main / standby relay coordination time interval (CTI) constraints, this invention aims to provide an adaptive relay protection technology based on convolutional neural network (CNN) and Gramian angular field (GAF) image transformation. This technology enhances the fault response capability of the power system by combining fault classification with relay coordination optimization. The specific technical solution is as follows: The method of the present invention includes at least the following steps (S1-S6), and can be executed online in a cyclical manner: S1 Distribution network modeling and relay master / standby relationship identification; S2 fault scenario generation and short-circuit analysis are performed to form a training / verification sample library; S3 Multi-source voltage and current timing sampling, segmentation, normalization and polar coordinate mapping; S4 Temporal Imaging and Multichannel Image Construction Based on GASF; S5 CNN feature extraction and fault classification output; S6 Relay adaptive setpoint update and coordinated optimization solution based on classification results. Attached Figure Description

[0006] Figure 1 This is a confusion matrix for testing the system, used to visually present the recognition status of each category and whether cross-confusion exists. Detailed Implementation

[0007] Step S1: Distribution network modeling and relay master / slave relationship identification: This step addresses the fundamental questions of "what is the protected object, where are the relays installed, and how is the primary / backup relationship established." This is because subsequent short-circuit calculations (S2) require knowledge of the fault location and the measurement points of each relay; CNN classification (S5) requires specific sampling channels; and coordination optimization (S6) requires knowledge of the primary / backup relay pairs. Therefore, this step outputs: a network topology model, a set of measurement points, a set of relays, and a set of primary / backup relay pairs. 1) Equivalent modeling of distribution networks:

[0008] This step primarily addresses the topology modeling problem of the distribution network. The distribution network topology determines the location of each relay and the current measurement points in response to faults. By modeling the distribution network, a complete power grid topology diagram is formed, and sets of relays and primary / standby relay pairs are established.

[0009] Representing the power distribution network as a graphical model: in: Distribution network topology diagram; : Set of nodes (buses); : A collection of branch circuits such as lines, switches, and transformers.

[0010] ; in: : Nodal admittance matrix; , These represent the conductance and susceptance components of the admittance matrix, respectively. Imaginary unit. 2) Relay object set and master / slave relationship set:

[0011] Define the set of directional overcurrent relays: in: : Relay set; : No. One relay; Number of relays.

[0012] Define the set of primary and backup relay pairs: in: Main and backup relay pairs; Main relay; : Backup relay.

[0013] The master-slave relationship can be generated by topology traversal, protected segment coverage relationship, or linked list / adjacency list method; for networks with distributed power sources, the master-slave relationship needs to consider the direction criterion under power flow reversal to avoid mismatch of protected segments due to misjudgment of direction.

[0014] Step S1 clarifies the network model Admittance matrix Relay set With primary and backup sets These outputs will serve as the input for short-circuit analysis in S2, the basis for sampling channel configuration in S5, and the foundation for coordination optimization constraints in S6.

[0015] Step S2: Fault Scenario Generation and Short Circuit Analysis (Sample Library Construction): This step addresses the issue of "ensuring the model has seen a sufficient number of fault modes." Because fault characteristics in high-voltage power distribution networks vary significantly with operating modes: changes in DG output alter short-circuit current; network reconfiguration changes fault current distribution. To enable the CNN to learn robust features and provide accurate fault current levels for relay optimization, this step generates fault samples covering multiple types, locations, impedances, and operating modes, and calculates the visible current for each relay. 1) Fault type set:

[0016] To simulate different types of faults, this invention defines a set of fault types: in: SLG: Single-phase ground fault; LL: Two-phase short circuit fault; LLG: Two-phase ground fault; LLL: Three-phase short circuit fault. 2) Short-circuit current calculation:

[0017] The formula for calculating short-circuit current is: in: : Short-circuit current at the fault point; Equivalent power supply voltage at the fault point; Equivalent Thevenin impedance at the fault point; Fault impedance.

[0018] In sample generation, fault impedance can be sampled by interval: in: Uniform distribution; : Fault impedance upper and lower limits. 3) Relay measurement current mapping:

[0019] A current transformer (CT) converts the current measured by a relay into a secondary current, and the relationship between the relay-measured current and the primary current is as follows: ; in: Primary current; : Sampled current on the secondary side of the current transformer; CT ratio.

[0020] Step S2 obtains a fault sample library covering changes in operating mode (containing Each relay measures current. (etc.), and provides a signal source for the timing sampling of S3, and provides for the tuning and optimization of S6. Critical current boundary quantities.

[0021] Step S3: Timing sampling, segmentation, normalization, and polar coordinate mapping; This step addresses the problem of "transforming electrical waveforms into a normalized sequence suitable for imaging and learning." This is because GASF generation requires the signal to be compressed into a range suitable for angle mapping (typically...). This step requires a fixed-length window to form a uniformly sized image. The output of this step is a time-series segment of uniform length and scale, providing input for S4 imaging. 1) Discrete sampling and windowing:

[0022] Sampling frequency can be represented as: in: : Sampling time interval; Sampling frequency (e.g., 4800 Hz).

[0023] The discrete sequence is: in: Discrete sampled values; : Sampling point number.

[0024] Window length: in: Number of sampling points in the window; Window duration. 2) Normalization:

[0025] Normalize the sequence to : ; in: Normalization to sequence; Minimum / maximum value within the window.

[0026] Remapped to : ; in: Mapped to sequence. 3) Polar coordinate angle mapping:

[0027] Angle definition: in: : No. Each sampling point corresponds to an angle; Inverse cosine function.

[0028] Radius (timestamp normalized): ; in: Normalized time radius; : Sampling point number; Window length.

[0029] Step S3 unifies the original voltage / current waveforms into a discrete sequence. By normalizing and completing polar coordinate mapping, the subsequent GASF imaging of S4 is ensured to be mathematically feasible and scale-consistent, thereby improving the training stability and cross-condition generalization ability of CNN.

[0030] Step S4: GASF temporal imaging and multi-channel image construction: This step addresses the crucial aspect of "transforming a temporal problem into an image problem." Using GASF, temporal correlations are encoded into a two-dimensional matrix, enabling CNNs to automatically extract "temporally correlated textures" using convolution operators. Furthermore, to accommodate three-phase voltage and current information, this step constructs multi-channel images, forming a unified input format. 1) GASF image generation:

[0031] in: Gramian Angular Summation Field matrix; Angle in polar coordinates; : The length of the sequence used for imaging within the window.

[0032] For ease of engineering implementation, the element expression can be written as: ; in: : No. Line number Column pixel values; : Corresponding angle.

[0033] Expanding further, this can be expressed as an equivalent form without the need for trigonometric functions: ; in: : Normalized sequence value.

[0034] 2) Multi-channel image construction: By stitching together the GASF images of three-phase voltage and current along channels, a multi-channel image is obtained: in: Multi-channel input tensor; : The matrix obtained from the time-series imaging of phase A voltage; : The matrix obtained from the time-series imaging of phase A current; : Assembled by channel dimension.

[0035] Step S4 maps the normalized time series to Images are generated, and a multichannel tensor containing multiphase voltage / current information is constructed. This output is the direct input to the CNN in S5 for feature extraction and fault classification, making the temporal fault features expressed in the form of "convolutional texture", which is more conducive to robust classification.

[0036] Step S5: CNN Feature Extraction and Fault Classification Output: This step addresses the problem of "automatically extracting features from images and outputting fault categories." Because fault features in new distribution networks are significantly affected by distribution generation (DG), operating mode, and fault impedance, traditional manual feature extraction is unstable. This step uses CNN for automatic learning to extract multi-scale texture features from GASF images and outputs fault category probabilities, providing a basis for updating relay settings in S6. 1) Convolution calculation:

[0037] Feature extraction is performed using a convolutional neural network (CNN). The convolution operation formula is as follows: ; in: : No. Convolutional layer output feature map in Linear response at the point; : No. Layer convolution kernel weights; : Input feature map of the previous layer; : Bias term; : Convolution kernel index.

[0038] The activation function (ReLU) is: in: Features after activation.

[0039] Pooling (Max Pooling): in: Pooled output; The set of indexes covered by the pooled window. 2) Feature vector concatenation:

[0040] The flattened eigenvectors can be represented as: in: : The flattened feature vector; Tensor flattening operator; : The last layer of pooling output.

[0041] Multi-channel feature stitching: ; in: : The concatenated total feature vector; :Depend on Features extracted from channel images; the rest are similar. 3) Fully Connected vs. Softmax Classification:

[0042] Fully connected: ; in: : Classification of logits; : Weight matrix of fully connected layer; Bias vector; : Concatenate feature vectors.

[0043] The Softmax function is used to classify features, and the formula is as follows: ; in: : Belongs to the The probability of a type of failure; : No. Logit class; Number of fault categories.

[0044] The cross-entropy loss is: ; in: Single-sample loss; : The one-hot component of the label (1 for correct classes, 0 for others); : Predicted probability.

[0045] In this invention, the design of the hidden layers in the CNN is crucial. A reasonable hidden layer architecture can improve the effect of feature extraction and further enhance the accuracy of fault classification. To optimize this process, we adopted the following network architecture, as shown in Table 1.

[0046] Table 1: Hidden layer structure of convolutional neural networks: ; Table 1 illustrates the specific architecture of the convolutional neural network, which includes two convolutional layers (64 and 128 kernels), two pooling layers, and one fully connected layer. The output layer uses the Softmax function to classify fault categories. Each convolutional layer employs the ReLU activation function to increase the model's non-linear expressive power. Pooling layers are used to reduce computational complexity and prevent overfitting. When processing time-series images, this network structure can effectively extract temporal features from current and voltage signals and automatically learn spatial features through multiple convolutional and pooling layers, ultimately outputting the fault type.

[0047] Step S5 extracts features from the GASF image through convolution, flattens and stitches them together, then passes them through a fully connected layer and Softmax to output the fault category probability, and completes model training using a loss function. This output provides the decision basis for the relay adaptive setting in step S6, which includes "fault type + confidence level + feature-corresponding current level".

[0048] Step S6: Relay Adaptive Setting Update and Coordinated Optimization: This step addresses the core engineering problems of "how the classification results guide protection actions, how to ensure coordination between primary and backup systems, and how to minimize the total operating time under constraints." First, appropriate relay curve parameters are selected based on the fault classification results; second, the pickup current is updated using the fault current level. With time scale Finally, a nonlinear optimization model including CTI constraints and parameter boundary constraints is established and solved, enabling the distribution network to achieve fast and selective fault isolation even under complex operating conditions. 6.1 Relay Action Time Model:

[0049] The relay's operating time model is as follows: ; in: : Relay operating time; Time Dial Setting; , Inverse time curve parameters (related to curve type, such as standard inverse time, extraordinary inverse time, etc.); : The relay measures the current (equivalent current on the primary side); Pick-up current (starting current threshold). 6.2 Main and Backup Relay Coordination Time Interval Constraints

[0050] The coordination time interval between the main and backup relays is: ; in: Backup relay operating time; Main relay operating time; Coordination Time Interval (e.g., 0.2 s). 6.3 Relay setting boundary constraints:

[0051] ; in: : No. The time scale of each relay; Its upper and lower boundaries.

[0052] ; in: : No. Each relay picks up the current; Its upper and lower boundaries. 6.4 Calculation of the upper and lower bounds of the pickup current:

[0053] ; in: : Pick up the lower bound of the current; Engineering lower limit constant (to avoid false starts due to excessively low constant value); : Maximum load current of the line; The minimum fault current visible in the relay.

[0054] in: : Pick up the upper limit of the current; Engineering upper limit constant (to avoid excessively high values ​​that may cause failure to operate); The minimum fault current visible in the relay. 6.5 Objective function for minimizing total motion time:

[0055] The final objective function is to minimize the total operating time of the relay: ; in: Minimize; : Number of main relays or number of relays involved in optimization; , : No. The set value variable of each relay; Curve parameters; : The current measured by the relay in the corresponding fault scenario. 6.6 Explicitly inject the "classification results" into the tuning update:

[0056] Classification output categories With confidence level : ; in: Predict the type of failure; Softmax outputs the probability.

[0057] Map fault categories to curve type parameters: ; in: : Category-to-curve parameter mapping function; Curve parameters.

[0058] Step S6 transforms the CNN classification results into executable setting updates for relay protection. Under the constraints of coordination time interval and setting boundary, the objective function is used to minimize the total action time, thereby achieving coordination between main and backup relays and faster action, thus forming a closed-loop adaptive protection of "identification-setting-coordination".

[0059] Symbol Summary Table: Network topology and sets of nodes and branches; : Nodal admittance matrix and its real and imaginary parts; : Relay set, relay number, quantity; : Main and backup relay pairs and their elements; Fault type set (SLG / LL / LLG / LLL); Thevenin equivalent and fault current parameters; CT ratio and primary / secondary current; Sampling frequency, sampling interval, number of window points, window length, discrete sequence; Normalized correlation coefficient; Polar coordinate angles and normalized radius; : GASF matrix and pixel values; CNN convolution kernel, bias, linear output, activation output, pooling output; Feature vector, concatenated vector, fully connected weights, bias, logits; : Probability, labels, number of categories, loss; : Relay operating time, time scale, curve parameters, current and pickup current; Main and backup action time and coordination interval; : Upper and lower bounds of the constant value; Maximum load current, minimum fault current; Example: To verify the applicability of the method of the present invention in a typical distribution network equivalent system, a 9-bus test system was built for simulation. The system voltage level is 35kV, including 12 lines and 1 power source with a capacity of 100 MVA connected to bus 1. The remaining buses connect to 8 loads, with load sizes ranging from 11.5 MW to 3.3 MW. To cover the bidirectional characteristics that may exist in power flow and fault current, directional overcurrent relays are configured at both ends of each line, with a total of 24 directional overcurrent relays deployed in the system. The target line between bus 6 and bus 7 was selected as the fault injection line, with a line length of 200 km, and implemented using two sections of three-phase π-type distributed parameter line modules connected in series to enhance the characterizability of the transient response of the long line. To improve the observability of fault characteristics, three-phase voltage and three-phase current signals were collected at the bus 7 side as fault identification inputs.

[0060] To train and validate the fault classifier, this embodiment constructs a dataset with 8 operating conditions (7 fault conditions + 1 normal condition) based on the settings. To ensure consistency between the category definition and the label, the fault categories and codes in this embodiment are shown in Table 2.

[0061] Table 2 Fault Categories and One-hot Codes: ; Once the category system is determined, simulations are needed to generate samples covering different fault locations to ensure that the classifier can not only distinguish types but also adapt to waveform changes when the same type occurs in different locations. To avoid overestimating the training results, the samples must be explicitly divided into training, validation, and test sets to ensure the independence and objectivity of the test evaluation. Based on the above objectives, the dataset size and allocation table is shown in Table 3.

[0062] Table 3 Dataset Allocation: ; Table 3 shows that the total number of samples is 320, of which the test set has 96 samples, the validation set has 45 samples used to monitor generalization ability and overfitting risk during training, and the training set has 179 samples to provide basic learning capacity for the deep model.

[0063] To transform the imaged information into fault category output, it is necessary to define the structure and parameter size of the classification network, especially the hidden layer dimension setting and regularization measures. Too small a hidden layer will result in insufficient expressive power, while too large a hidden layer will easily lead to overfitting and high computational cost. Therefore, in this embodiment, a progressively compressed fully connected layer structure is adopted, and Dropout is introduced at key locations to balance expressive power and generalization ability. Based on the above objectives, the hidden layer structure of the classification network is shown in Table 4 below.

[0064] Table 4 Hidden Layer Structure: ; Table 4 shows that the input dimensions are gradually reduced from 38,400 to 8,192, 4,096, and 512, before outputting 8 categories of probabilities. This structure helps compress the high-dimensional representation after multi-channel information fusion into more stable discriminative features, while avoiding information loss caused by a large-scale dimensionality reduction at once. The two 25% Dropouts demonstrate control over overfitting and are suitable for scenarios where changes in fault location cause fluctuations in waveform details, improving the model's ability to stably identify samples from different fault points.

[0065] During training, the model accuracy increased with the number of training epochs and eventually stabilized, while the overall loss decreased with slight fluctuations in some stages, demonstrating that the model gradually learned separable features and achieved convergence. In the testing phase, independent test sets were evaluated. The classifier could stably distinguish eight operating conditions, effectively differentiating between ground faults and two-phase short circuits of different phases. This indicates that "temporal imaging + deep feature learning" can capture the feature differences caused by phase combinations and has consistent discrimination ability for samples at different fault locations. The predicted distribution of each class of samples in the test set was statistically analyzed, and the correspondence between "true category - predicted category" was displayed in a matrix. Figure 1 As shown, this is used to visually present the identification status of each category and whether there is any overlap or confusion.

[0066] from Figure 1 It can be seen that the identification results of all eight operating conditions are concentrated on the main diagonal, with off-diagonal elements being 0, indicating that the classifier has achieved stable differentiation of each class of samples without cross-referencing. This result directly reflects the synergistic effect of the two key innovations of this invention: First, by imaging and representing the time-series information of three-phase voltage and three-phase current and inputting it in a multi-channel manner, the differences in fault phases (such as two-phase short circuits of AB / AC / BC and single-phase grounding of A / B / C) are significantly amplified in the image texture and time-series correlation structure, avoiding the "confusion of different phases of the same type" that is easy to occur when based on a single image threshold or a small number of manual features in the traditional method; Second, by adopting a feature extraction and hierarchical hidden layer fusion structure, the multi-channel features form a more robust discriminative representation under progressive compression and regularization constraints, so as to maintain clear class boundaries even under different fault locations and operating disturbances. Therefore, the "full diagonal" phenomenon of the confusion matrix not only indicates high overall accuracy, but also demonstrates that the present invention has significant technical effects in fine-grained phase differentiation, fault structure differentiation, and generalization stability in engineering scenarios, providing a reliable preliminary criterion for subsequent adaptive setting and coordinated optimization of relay protection based on classification results.

[0067] After fault category identification, to achieve selectivity and speed of protection actions, it is necessary to model the coordination relationships between relays. Specifically, it is necessary to identify one or more backup relays corresponding to each main relay, so as to check the sequence of main and backup actions, evaluate whether the coordination relationship meets the coordination requirements, and constrain the actions of backup relays during the optimization process. Based on the above objectives, the main and backup relay relationship table 5 is established as follows.

[0068] Table 5. Relationship between main and backup relays in the 9-bus system: ; Table 5 shows that the system exhibits both single-level and multi-level backups: some main relays correspond to only one backup relay, resulting in a simpler structure; while some main relays correspond to two or even three levels of backup relays, reflecting overlapping power supply paths and protection coverage within the network structure. Multi-level backups improve reliability, but also mean that coordination and optimization must be more cautious: the more backup levels there are, the more necessary it is to control the action range to avoid "backup relays operating beyond their designated level" at certain fault locations, leading to wider-area disconnection. Therefore, this table is used not only to establish constraints but also to identify potentially complex coordination sections.

[0069] When performing relay setting and coordination optimization, if a reasonable range for the pickup current is not set, the optimization may result in excessively low pickup values ​​leading to malfunctions due to load fluctuations, or excessively high pickup values ​​leading to failure to trip due to faults. To ensure the engineering usability of the setting results, it is necessary to set upper and lower bounds for the pickup current for each relay and use them as constraints within the feasible region of optimization. Based on the above objectives, the relay pickup current boundaries are shown in Table 6 below.

[0070] Table 6 Upper and lower limits of relay pickup current: ; Table 6 shows significant differences in the pickup boundaries of different relays, reflecting the differences in load levels and fault visibility in their respective line sections. For example, the boundaries of relays 13–16 are significantly lower, indicating that their operating current levels and visible fault current levels are relatively small. In contrast, the boundaries of relays 23–24 are the highest, indicating that the current magnitude in their corresponding sections is larger, requiring higher pickup settings to avoid malfunctions. The significance of this table lies not only in providing specific feasible regions, but more importantly, in providing a basis for "zonal differentiated setting" for subsequent optimization, avoiding protection mismatch caused by using a uniform threshold.

[0071] To verify the engineering applicability of this invention in the adaptive setting and coordinated optimization of relays, it is necessary to compare different solution strategies from two dimensions: "setting effect" and "computational cost." On the one hand, the setting effect can be reflected by indicators such as total operating time; the smaller the indicator, the faster the fault clearing and the better the protection speed. On the other hand, the computational cost can be reflected by the solution time; the shorter the time, the more beneficial it is to achieve online and rapid updating of settings in distribution networks with frequently changing operating modes. Based on this, several representative intelligent optimization algorithms and mathematical programming solvers are selected for comparison, including GA (Genetic Algorithm), PSO (Particle Swarm Optimization), DE (Differential Evolutionary Algorithm), HS (Harmony Search Algorithm), SOA (Seagull Algorithm), and CONOPT (Nonlinear Programming Solver called in the GAMS environment). Table 7 summarizes the comparison results of the target value (total operating time) and computational time of the above methods under the same constraints and target settings, so as to intuitively evaluate the trade-off between "optimal effect" and "optimal speed" of different methods.

[0072] Table 7. Comparison of target values ​​and time consumption for different optimization methods: ; Table 7 shows a clear trade-off between "effectiveness" and "speed" among different optimization methods: DE and HS achieve lower total action time, meaning they may be more advantageous in offline tuning; while CONOPT (GAMS) has the shortest computation time, only 0.047 seconds, demonstrating extremely strong fast solution capability and is more suitable for online adaptive tuning scenarios. Although PSO has a low computation time, its target value is not as high as DE / HS; GA and HS have longer computation times, which is not conducive to online updates. Therefore, when implementing fast adaptive tuning in distribution networks with frequent changes in operating modes or high fault risk, fast solution strategies can significantly improve the real-time performance and deployability of tuning updates.

[0073] The beneficial effects of this invention are as follows: 1. Improved Fault Classification Accuracy: This invention utilizes Gramian Angular Field (GAF) technology based on time-series imaging and Convolutional Neural Network (CNN) for fault classification. This enables the extraction of complex temporal features from current and voltage time-series signals in power systems, significantly improving the accuracy of fault type classification. Compared to traditional rule-based fault identification methods, this invention better addresses dynamic and complex fault characteristics, especially in high-voltage, high-efficiency distribution networks where the integration of a high proportion of distributed power sources leads to changes in fault modes.

[0074] 2. Achieving adaptive protection setting adjustment of the relay: This invention can dynamically adjust the relay's pickup current based on fault classification results and short-circuit current calculations. The system adaptively updates its settings based on real-time fault types and system operating states, using a time delay (TDS) mechanism. This adaptive protection mechanism effectively avoids the mismatch between fixed settings and fault types present in traditional relay protection methods, improving the relay's response speed and accuracy.

[0075] 3. Optimize relay coordination and reduce fault isolation time: The relay coordination optimization method proposed in this invention takes into account the coordination between primary and backup relays during the relay setting update process, ensuring that relays can operate quickly and in a coordinated manner when a fault occurs. By minimizing the total operating time and using constraints such as the Coordination Time Interval (CTI), the fault isolation delay caused by improper coordination between relays is reduced, further improving the safety and reliability of the power system.

[0076] 4. Adaptability to Complex Fault Characteristics: The method of this invention can adapt to different fault types and power system configurations, including traditional distribution networks and distribution networks with a high proportion of distributed generation access. When facing complex fault characteristics of source-grid-load-storage interleaving and AC / DC distribution networks, it can achieve precise fault isolation and protection through real-time fault classification and adaptive relay setting adjustment.

[0077] 5. Improved Power System Protection Efficiency and Stability: This invention provides a novel solution through the organic combination of fault classification, adaptive relay setting adjustment, and relay coordination optimization. This significantly improves the protection efficiency and system stability of distribution networks with a high proportion of distributed power sources. When a fault occurs in the power system, it can quickly and accurately detect the fault type and promptly adjust the relay's operating state, reducing system downtime and ensuring the continuity of power supply.

[0078] 6. Facilitates Engineering Application and Deployment: The method of this invention enables real-time online operation, is applicable to various power system scenarios, and can dynamically adjust according to real-time changes in the system. The implementation of this method does not rely on fixed rules or manual intervention, exhibiting strong automation and intelligence, making fault handling more efficient and flexible, and suitable for large-scale application in smart grids and modern power distribution systems.

[0079] Through the above-mentioned beneficial effects, the present invention can not only improve the accuracy of fault classification and relay protection, but also significantly reduce fault isolation time, improve the safety, stability and operating efficiency of the power system, and meet the needs of the intelligent and automated development of modern power distribution networks.

Claims

1. A fault classification and adaptive relay protection coordination optimization method for a new type of distribution network with high voltage and high efficiency, characterized in that, include: 1) Construct a distribution network topology model and determine the relay set and the main / standby relay pair set; 2) Perform short-circuit analysis on the distribution network based on the preset fault type set and fault scenario parameters to generate a fault sample library containing voltage and current time-series signals; 3) Perform windowed sampling and normalization on the voltage and current time-series signals, and perform polar coordinate angle mapping; 4) Use Gramian Angular Summation Field to convert the normalized time-series signals into two-dimensional images and construct a multi-channel image tensor; 5) Input the multi-channel image tensor into a convolutional neural network for feature extraction and fault classification output; 6) Determine the curve parameters of the inverse-time directional overcurrent relay based on the fault classification results and update the picking current and time scale. Under the condition of satisfying the main and backup relay coordination time interval constraints and setting boundary constraints, establish a nonlinear optimization model, solve for the optimal relay setting value with minimum total action time, and issue it for execution.

2. The method as described in claim 1, characterized in that, Step 2) The short-circuit current is calculated as follows: .

3. The method as described in claim 1, characterized in that, Step 3) Normalization includes: ; in, For sampled values, The minimum / maximum value within the window; The angle mapping described in step 3) satisfies: ; in, For angle, For mapping to The normalized value; The sampling time interval in step 3) satisfies: 。 4. The method as described in claim 1, characterized in that, In step 4), the multi-channel image tensor is obtained by stitching together the GASF images of three-phase voltage and three-phase current along each channel. The Gramian Angular Summation Field mentioned in step 4) is generated according to the following formula: ; in, For pixel values, The angle is obtained by normalizing the sequence and then applying an inverse cosine mapping.

5. The method as described in claim 1, characterized in that, In step 5), the convolutional neural network uses convolution operations: ； in, For convolution kernel weights, For bias; In step 5), the fault classification output is processed via Softmax. ;in, For the first Class failure probability, This represents the number of categories.

6. The method as described in claim 1, characterized in that, In step 6), the relay operating time satisfies: ; in, For the action time, For time scale, To measure the current, To pick up the current, For curve parameters; In step 6), the coordination time interval between the main and backup relays must satisfy: ; in, To coordinate time intervals; In step 6), the constant boundary constraints satisfy: ； The upper bound of the pickup current satisfies: ; The lower bound of the pickup current satisfies: 。 7. The method as described in claim 1, characterized in that, The optimization objective in step 6) is to minimize the total action time: .

8. A fault classification and adaptive relay protection coordination optimization system for a new type of distribution network with high voltage and high efficiency, characterized in that: include: Distribution network modeling and primary / standby relationship identification module; Fault scenario generation and short circuit analysis module; Sampling and normalization, and polar coordinate mapping module; GASF temporal imaging and multichannel building block; CNN Fault Classification Module; Relay adaptive setpoint update and coordination optimization module; Each module is configured to perform the steps of the method described in any one of claims 1-7.

9. A computing device comprising a processor and a memory, the memory storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed, it implements the method described in any one of claims 1-7.

Images