Relay protection method and device based on time sequence data prediction

Abstract

The invention discloses a relay protection method and device based on time sequence data prediction, and the method comprises the steps: collecting the operation sequence data of a main network power transformation relay protection device in real time through a wireless sensing node, and transmitting the operation sequence data to a data processing layer through a Zigbee wireless network protocol; historical data features are extracted based on a sliding window method, abnormal data are recognized and removed through a quartile method, and missing data are filled up through a linear interpolation algorithm; constructing a CNN-GRU prediction model fused with the attention mechanism, dynamically weighting key features by the attention mechanism, inputting the preprocessed data into the model for training, and generating an operation state evaluation value based on a prediction result, and when the evaluation value exceeds a preset threshold, triggering a multi-level early warning mechanism. And the prediction precision is verified by using a decision coefficient R2, the R2 is required to be greater than or equal to 0.95, and model parameters are dynamically updated to adapt to the working condition change of the power grid.

Description

Technical Field

[0001] The present application relates to the field of relay protection technology, and in particular to a relay protection method and device based on time series data prediction. Background Art

[0002] As a core component of the power system, the safe and stable operation of the main grid plays a key role in supporting the reliable power supply of the entire power system. Substation relay protection devices, as loyal guardians of the main grid, monitor the operating status of power equipment in real time and quickly act when faults or anomalies are detected, isolating the faulty area and preventing further escalation.

[0003] However, with the continued expansion of power systems and the increasing complexity of operating conditions, traditional monitoring methods for substation relay protection devices have gradually exposed many limitations. Traditional monitoring methods rely on regular manual inspections, which not only consumes a large amount of manpower and material resources, but also has difficulty capturing instantaneous changes in the equipment's operating status, resulting in monitoring lags. Furthermore, faced with massive amounts of operating data, traditional monitoring methods lack efficient data processing and analysis capabilities, making it impossible to deeply explore the potential information behind the data and providing accurate support for equipment operation and maintenance decisions. Therefore, there is an urgent need to explore a new, more intelligent and efficient relay protection method to meet the actual needs of modern power systems. Summary of the Invention

[0004] The present application provides a relay protection method and device based on time series data prediction to solve the above problems.

[0005] In one aspect, the present application provides a relay protection method based on time series data prediction, the method comprising the following steps:

[0006] Step S1: The operation sequence data of the main network substation relay protection device, including current, voltage and temperature data, is collected in real time through the wireless sensor node, and transmitted to the data processing layer using the Zigbee wireless network protocol;

[0007] Step S2: Perform anomaly detection and missing data repair on the collected time series data: extract historical data features based on the sliding window method, identify and eliminate abnormal data using the interquartile range method, and use the linear interpolation algorithm to fill in missing data;

[0008] Step S3: Construct a CNN-GRU prediction model integrating attention mechanism: the CNN layer extracts local spatial features of time series data through convolution kernels, the GRU layer captures long-term temporal dependencies, and the attention mechanism dynamically weights key features. The model output expression is:

[0009] y t =Attention(CNN(X t-n:t),GRU(X t-m:t ))

[0010] Among them, X t-n:t is the input data in the current time window, X t-m:t is historical time series data;

[0011] Step S4: Input the preprocessed data into the model for training. The optimization goal is to minimize the mean square error, as shown in the following formula:

[0012]

[0013] Step S5: Generate an operating status evaluation value based on the prediction result. When the evaluation value exceeds a preset threshold, a multi-level early warning mechanism is triggered, including local sound and light alarms and push notifications to the remote operation and maintenance platform;

[0014] Step S6: Using the coefficient of determination R 2 Verify the prediction accuracy, requiring R 2 ≥0.95, and dynamically update model parameters to adapt to changes in grid conditions.

[0015] In one implementation of the present application, in step S2, anomaly detection adopts an improved isolation forest algorithm to identify outliers by randomly dividing the data space, and combines the variational autoencoder (VAE) to reconstruct the data distribution to enhance the robustness of anomaly detection.

[0016] In one implementation of the present application, in the CNN-GRU fusion model, the attention weight is calculated using a scaled dot product attention mechanism, and the formula is:

[0017]

[0018] Among them, Q, K, and V are generated by mapping the feature matrices output by CNN and GRU respectively, and d k is the feature dimension.

[0019] In one implementation of the present application, a Dropout layer and L2 regularization are introduced in the training process of step S4 to prevent overfitting. The optimizer adopts the Adam algorithm, and the learning rate is set to 0.001-0.005.

[0020] In one implementation of the present application, the early warning mechanism includes three levels of classification: the first-level early warning is a minor abnormality of the equipment, which triggers a local prompt; the second-level early warning is a potential failure risk, which is pushed to the regional monitoring center; the third-level early warning is an emergency failure, which directly links the protection device to execute a tripping command.

[0021] In one implementation of the present application, data acquisition uses the CC2530 microcontroller as the core, integrates a multi-channel ADC module, has a sampling frequency of 1kHz, supports simultaneous acquisition of 8 analog signals, and interacts with the wireless communication module through the SPI interface.

[0022] In one implementation of the present application, the missing data repair adopts a time series data completion method based on a generative adversarial network (GAN). The generator learns the historical data distribution to generate the completion value, and the discriminator distinguishes between real and generated data, and improves the completion accuracy through adversarial training.

[0023] In one implementation of the present application, the model is deployed on an edge computing node and uses TensorRT to accelerate reasoning to ensure that the prediction delay is less than 50ms to meet real-time requirements.

[0024] The present application also provides a relay protection device based on time series data prediction, which includes: a front-end data acquisition module: composed of multiple wireless sensor nodes, each node has a built-in CC2530 microcontroller, a Zigbee communication module and a high-precision sensor for collecting current, voltage and temperature time series data; a data processing module: equipped with an NVIDIA Jetson Nano processor, running a CNN-GRU fusion model to complete data cleaning, prediction and status evaluation; an early warning execution module: integrating an audible and visual alarm, a 4G communication unit and a relay control interface, and executing graded early warnings or protection actions according to the evaluation results.

[0025] In one implementation of the present application, the data processing module adopts a containerized architecture, deploys model training and inference services through Docker, supports dynamic loading of configuration parameters of different substations, and reserves an API interface for seamless connection with the superior energy management system.

[0026] The present application provides a relay protection method and device based on time series data prediction, which has the following beneficial effects:

[0027] (1) The fusion model can accurately capture the local characteristics and long-term dependencies of time series data, and its prediction accuracy is significantly better than traditional methods;

[0028] (2) The system can effectively delete abnormal data and repair missing data to ensure data integrity and reliability;

[0029] (3) The system can generate early warning information in real time and provide scientific decision-making support for operation and maintenance personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0031] Figure 1 A flow chart of a relay protection method based on time series data prediction provided in an embodiment of the present application;

[0032] Figure 2 A diagram of the composition of a relay protection system based on time series data prediction provided in an embodiment of the present application. DETAILED DESCRIPTION

[0033] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0034] The embodiment of the present application provides a relay protection method and device based on time series data prediction. The technical solution proposed in the embodiment of the present application is described in detail below with reference to the accompanying drawings.

[0035] Example 1

[0036] Figure 1 This is a flow chart of a relay protection method based on time series data prediction provided in an embodiment of the present application. Figure 1 As shown, the method mainly includes the following steps:

[0037] Step S1: The operation sequence data of the main network substation relay protection device, including current, voltage and temperature data, is collected in real time through the wireless sensor node, and transmitted to the data processing layer using the Zigbee wireless network protocol;

[0038] Step S2: Perform anomaly detection and missing data repair on the collected time series data: extract historical data features based on the sliding window method, identify and eliminate abnormal data using the interquartile range method, and use the linear interpolation algorithm to fill in missing data;

[0039] Step S3: Construct a CNN-GRU prediction model integrating attention mechanism: the CNN layer extracts local spatial features of time series data through convolution kernels, the GRU layer captures long-term temporal dependencies, and the attention mechanism dynamically weights key features. The model output expression is:

[0040] y t =Attention(CNN(Xt-n:t ),GRU(X t-m:t ))

[0041] Among them, X t-n:t is the input data in the current time window, X t-m:t is historical time series data;

[0042] Step S4: Input the preprocessed data into the model for training. The optimization goal is to minimize the mean square error, as shown in the following formula:

[0043]

[0044] Step S5: Generate an operating status evaluation value based on the prediction result. When the evaluation value exceeds a preset threshold, a multi-level early warning mechanism is triggered, including local sound and light alarms and push notifications to the remote operation and maintenance platform;

[0045] Step S6: Using the coefficient of determination R 2 Verify the prediction accuracy, requiring R 2 ≥0.95, and dynamically update model parameters to adapt to changes in grid conditions.

[0046] In one implementation of the present application, in step S2, anomaly detection adopts an improved isolation forest algorithm to identify outliers by randomly dividing the data space, and combines the variational autoencoder (VAE) to reconstruct the data distribution to enhance the robustness of anomaly detection.

[0047] In one implementation of the present application, in the CNN-GRU fusion model, the attention weight is calculated using a scaled dot product attention mechanism, and the formula is:

[0048]

[0049] Among them, Q, K, and V are generated by mapping the feature matrices output by CNN and GRU respectively, and d k is the feature dimension.

[0050] In one implementation of the present application, a Dropout layer and L2 regularization are introduced in the training process of step S4 to prevent overfitting. The optimizer adopts the Adam algorithm, and the learning rate is set to 0.001-0.005.

[0051] In one implementation of the present application, the early warning mechanism includes three levels of classification: the first-level early warning is a minor abnormality of the equipment, which triggers a local prompt; the second-level early warning is a potential failure risk, which is pushed to the regional monitoring center; the third-level early warning is an emergency failure, which directly links the protection device to execute a tripping command.

[0052] In one implementation of the present application, data acquisition uses the CC2530 microcontroller as the core, integrates a multi-channel ADC module, has a sampling frequency of 1kHz, supports simultaneous acquisition of 8 analog signals, and interacts with the wireless communication module through the SPI interface.

[0053] In one implementation of the present application, the missing data repair adopts a time series data completion method based on a generative adversarial network (GAN). The generator learns the historical data distribution to generate the completion value, and the discriminator distinguishes between real and generated data, and improves the completion accuracy through adversarial training.

[0054] In one implementation of the present application, the model is deployed on an edge computing node and uses TensorRT to accelerate reasoning to ensure that the prediction delay is less than 50ms to meet real-time requirements.

[0055] The above is a relay protection method based on time series data prediction provided by an embodiment of the present application. Based on the same inventive concept, an embodiment of the present application also provides a relay protection device based on time series data prediction. Figure 2 A diagram showing the composition of a relay protection device based on time series data prediction provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the device mainly includes: a front-end data acquisition module 201: composed of multiple wireless sensor nodes, each node has a built-in CC2530 single-chip microcomputer, a Zigbee communication module and a high-precision sensor for collecting current, voltage and temperature time series data; a data processing module 202: equipped with an NVIDIA Jetson Nano processor, running a CNN-GRU fusion model to complete data cleaning, prediction and status evaluation; an early warning execution module 203: integrating an audible and visual alarm, a 4G communication unit and a relay control interface, and executing graded early warning or protection actions according to the evaluation results.

[0056] The data processing module adopts a containerized architecture, deploys model training and inference services through Docker, supports dynamic loading of configuration parameters of different substations, and reserves an API interface for seamless connection with the superior energy management system.

[0057] Example 2

[0058] The following specifically demonstrates the application of the solution of the embodiment of the present application in a specific scenario. The main grid substation relay protection device of a certain 220kV substation was used as the test object to verify the effectiveness of the relay protection method based on time series data prediction. The test objectives include: (1) verifying the real-time and reliability of the data collected by the wireless sensor node; (2) evaluating the performance of the anomaly detection and missing data repair algorithm; (3) testing the prediction accuracy of the CNN-GRU fusion model; and (4) verifying the response efficiency of the multi-level early warning mechanism.

[0059] The experimental equipment for this example is as follows: the wireless sensor node is designed based on the CC2530 microcontroller, integrated with a 16-bit ADC module (ADS1115), supports 8-channel synchronous acquisition, and has a sampling frequency of 1kHz; the communication module adopts the Zigbee protocol (CC2531 chip), with a transmission frequency of 2.4GHz and a transmission distance of ≤100m; the data processing server uses the NVIDIA Jetson Nano (quad-core ARMCortex-A57, 128-core GPU) with 4GB of memory; the early warning execution module specifically includes a local sound and light alarm (buzzer + LED), a 4G communication module (SIM7600), and a relay control interface (supporting RS485 protocol).

[0060] The data acquisition parameters of this example are: (1) Acquisition variables: three-phase current (phase A, phase B, phase C), line voltage (AB, BC, CA), device temperature; (2) Sampling frequency: 1kHz (per channel); (3) Data storage: local storage at the edge node (SD card), and simultaneous upload to the cloud database.

[0061] The experimental scenarios for this example are as follows: Normal operating conditions: 72 hours of continuous data collection to simulate steady-state operation; Abnormal operating conditions: artificial injection of fault signals such as current surge (+20%), voltage drop (-15%), and temperature exceeding the limit (>85°C).

[0062] The experimental steps and key technologies of this example are as follows:

[0063] (1) Data collection and transmission. (a) Deployment of wireless sensor nodes: 6 sensor nodes (2 per phase) are installed on the high-voltage side of the substation to collect current, voltage, and temperature data in real time; (b) Data transmission protocol configuration: The Zigbee network adopts a Mesh topology, sets channel 11 (2.405GHz), sets the packet interval to 10ms, and uses CRC checksum to ensure transmission integrity; (c) Data format: A single data packet contains a timestamp (Unix milliseconds), 8-channel analog quantity (16-bit precision), and a checksum. The format is: [Timestamp][A phase current][B phase current][C phase current][AB voltage][BC voltage][CA voltage][temperature][CRC].

[0064] (2) Anomaly detection and missing data repair. (a) Sliding window method feature extraction: The window size is set to 1 second (1000 sampling points), and statistics such as mean, variance, and peak-to-peak value are extracted; (b) Anomaly detection algorithm: Improved Isolation Forest (iForest): Set the tree depth limit to 8 and the number of subsamples to 256; Variational Autoencoder (VAE): The encoder structure is [64-32-16], the latent space dimension is 8, and the reconstruction error threshold ε = 0.05; Anomaly judgment: If the iForest score is greater than 0.6 and the VAE reconstruction error is greater than ε, it is marked as abnormal data; (c) Missing data repair: GAN completion network: Generator (LSTM layer × 2, hidden unit 64), Discriminator (CNN layer × 2, convolution kernel 3 × 1), adversarial training iteration 1000 times; Completion effect evaluation: Use mean square error (MSE) and dynamic time warping (DTW) indicators, requiring MSE < 0.01 and DTW < 0.1. The performance comparison of anomaly detection and repair is shown in Table 1 below.

[0065] Table 1. Comparison of anomaly detection and repair performance

[0066] Scenario Number of outliers Detection rate False positive rate Complete MSE Current surge 120 98.3 1.2 0.008 Voltage drop 85 96.5 1.8 0.009 Temperature exceeds limit 45 99.1 0.7 0.007

[0067] (3) Construction and training of CNN-GRU fusion model. (a) The model structure is as follows: CNN layer: 2 layers of convolution (kernel size 3×1, number of channels 32 and 64), ReLU activation, maximum pooling (pooling size 2); GRU layer: 128 hidden units, time step m=60; attention mechanism: scaled dot product attention, feature dimension d_k=64; output layer: fully connected layer (32 neurons) + Sigmoid activation; (b) The training parameters are as follows: input window n=30 (historical 30 seconds data), output window 1 second; optimizer: Adam (learning rate 0.003, β1=0.9, β2=0.999); regularization: Dropout rate 0.2, L2 coefficient λ=0.001; training set: 80% (57600 items), validation set: 20% (14400 items); (c) The loss function is as follows:

[0068]

[0069] (4) Multi-level warning mechanism. The thresholds are set as follows: Level 1 warning: evaluation value > 0.8 (minor equipment abnormality); Level 2 warning: evaluation value > 0.9 (potential failure risk); Level 3 warning: evaluation value > 0.95 (emergency failure); The response time test table is shown in Table 2 below.

[0070] Table 2 Response time test table

[0071]

[0072]

[0073] Experimental results analysis: (1) Prediction accuracy verification, evaluation index: determination coefficient R 2 , root mean square error (RMSE); the results are shown in Table 3 below.

[0074] Table 3 Evaluation indicators

[0075] Data Type <![CDATA[R 2 ]]> RMSE Current 0.972 0.021 Voltage 0.985 0.015 temperature 0.961 0.032

[0076] (2) Real-time verification. Edge inference latency: After using TensorRT acceleration, a single prediction takes ≤45ms (meeting the <50ms requirement); end-to-end latency (data collection → warning execution): ≤100ms.

[0077] (3) Early warning effectiveness. Fault injection test: (a) Current surge scenario: Level 3 early warning trigger rate 100%, tripping time 92ms; (b) Voltage drop scenario: Level 2 early warning trigger rate 98%, remote alarm push successful; (c) Temperature over-limit scenario: Level 1 early warning trigger rate 95%, local alarm response normal.

[0078] Experimental conclusion: This example verifies the feasibility and superiority of the relay protection method based on time series data prediction through a complete experimental process. The wireless sensor nodes and Zigbee network achieve high-precision and low-latency data acquisition; the improved isolation forest + VAE algorithm is robust in anomaly detection (detection rate > 96%); the CNN-GRU fusion model has significant prediction accuracy (R 2 >0.95), and edge deployment meets real-time requirements. The multi-level warning mechanism effectively differentiates fault levels, and response latency is below design requirements. This approach provides reliable technical support for relay protection in smart grids and has high engineering application value.

[0079] The various embodiments in this application are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences from other embodiments. In particular, the device embodiments are generally similar to the method embodiments, so the description is relatively simple. For relevant parts, refer to the partial description of the method embodiments.

[0080] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0081] The foregoing is merely an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A relay protection method based on time series data prediction, characterized in that: The method comprises the following steps: Step S1: The operation sequence data of the main network substation relay protection device, including current, voltage and temperature data, is collected in real time through the wireless sensor node, and transmitted to the data processing layer using the Zigbee wireless network protocol; Step S2: Perform anomaly detection and missing data repair on the collected time series data: extract historical data features based on the sliding window method, identify and eliminate abnormal data using the interquartile range method, and use the linear interpolation algorithm to fill in missing data; Step S3: Construct a CNN-GRU prediction model integrating attention mechanism: the CNN layer extracts local spatial features of time series data through convolution kernels, the GRU layer captures long-term temporal dependencies, and the attention mechanism dynamically weights key features. The model output expression is: y t =Attention(CNN(X t-n:t ),GRU(X t-m:t )) Among them, X t-n:t is the input data in the current time window, X t-m:t is historical time series data; Step S4: Input the preprocessed data into the model for training. The optimization goal is to minimize the mean square error, as shown in the following formula: Step S5: Generate an operating status evaluation value based on the prediction result. When the evaluation value exceeds a preset threshold, a multi-level early warning mechanism is triggered, including local sound and light alarms and push notifications to the remote operation and maintenance platform; Step S6: Using the coefficient of determination R 2 Verify the prediction accuracy, requiring R 2 ≥0.95, and dynamically update model parameters to adapt to changes in grid conditions.

2. The relay protection method based on time series data prediction according to claim 1, characterized in that: In step S2, anomaly detection uses an improved isolation forest algorithm to identify outliers by randomly dividing the data space, and combines it with a variational autoencoder (VAE) to reconstruct the data distribution to enhance the robustness of anomaly detection.

3. The relay protection method based on time series data prediction according to claim 1, characterized in that: In the CNN-GRU fusion model, the attention weight is calculated using the scaled dot product attention mechanism, and the formula is: Among them, Q, K, and V are generated by mapping the feature matrices output by CNN and GRU respectively, and d k is the feature dimension.

4. The relay protection method based on time series data prediction according to claim 1, characterized in that: During the training process of step S4, a Dropout layer and L2 regularization are introduced to prevent overfitting. The optimizer adopts the Adam algorithm, and the learning rate is set to 0.001-0.

005.

5. The relay protection method based on time series data prediction according to claim 1, characterized in that: The warning mechanism includes three levels of classification: Level 1 warning is a minor equipment anomaly, triggering a local prompt; Level 2 warnings are potential failure risks and are sent to the regional monitoring center; The third level warning is an emergency fault, which directly links the protection device to execute the tripping command.

6. The relay protection method based on time series data prediction according to claim 1, characterized in that: The data acquisition uses the CC2530 microcontroller as the core, integrates a multi-channel ADC module, has a sampling frequency of 1kHz, supports the simultaneous acquisition of 8-channel analog signals, and interacts with the wireless communication module through the SPI interface.

7. The relay protection method based on time series data prediction according to claim 1, characterized in that: The missing data repair adopts a time series data completion method based on a generative adversarial network (GAN). The generator learns the historical data distribution to generate the completion value, and the discriminator distinguishes between real and generated data, and improves the completion accuracy through adversarial training.

8. The relay protection method based on time series data prediction according to claim 1, characterized in that: The model is deployed on edge computing nodes and uses TensorRT to accelerate reasoning, ensuring that the prediction delay is less than 50ms, meeting real-time requirements.

9. A relay protection device based on time series data prediction, characterized in that: The device includes: a front-end data acquisition module: composed of multiple wireless sensor nodes, each node has a built-in CC2530 single-chip microcomputer, a Zigbee communication module and a high-precision sensor, and is used to collect current, voltage and temperature time series data; a data processing module: equipped with an NVIDIA Jetson Nano processor, running a CNN-GRU fusion model to complete data cleaning, prediction and status evaluation; an early warning execution module: integrating an audible and visual alarm, a 4G communication unit and a relay control interface, and executing graded early warning or protection actions based on the evaluation results.

10. The relay protection device based on time series data prediction according to claim 9, characterized in that: The data processing module adopts a containerized architecture, deploys model training and inference services through Docker, supports dynamic loading of configuration parameters of different substations, and reserves an API interface for seamless connection with the superior energy management system.

Images