Full-station protection setting management method and system based on NLP
By adopting NLP-based full-station protection fixed-value management method in hydropower plants and combining with a variety of artificial intelligence technologies, the problems of low efficiency, poor accuracy and insufficient intelligence in the existing technology are solved, and the protection fixed-value management is achieved Intelligent and automated management of values improves the reliability and security of the system.
Patent Information
- Application Number
- CN202411096188.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-12
AI Technical Summary
The existing protection fixed value management method relies on manual settings and manual verification, which is inefficient and prone to human errors, cannot adapt to the complex power system operating environment, lacks data sharing and intelligent optimization capabilities, resulting in insufficient system reliability and security. .
Using NLP-based station-wide protection fixed-value management method, combining self-supervised learning, time series analysis, generation adversarial network, variational autoencoder and reinforcement learning technology, data acquisition, automatic feature extraction, prediction trends, data augmentation and Automatically adjust parameters to realize intelligent and automated management of protection fixed values.
Improve the efficiency, accuracy and intelligence of protection fixed value management, ensure the efficiency and reliability of data acquisition and preprocessing, enhance abnormal detection and early warning capabilities, dynamically adjust protection fixed value parameters, and optimize the operating status of the equipment , improves the adaptability and safety of the system.
Smart Images

Figure CN118628289B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of hydropower plants, and in particular relates to a method and system for managing fixed value of whole-station protection based on NLP. Background Art
[0002] With the continuous improvement of the automation and intelligence level of hydropower plants, the role of protection setting management in ensuring the safe and stable operation of power systems is becoming increasingly important. The protection setting is a key parameter of the power system protection device, which determines when to start the protection device to prevent equipment damage or system failure. However, the existing protection setting management method mainly relies on manual setting and manual verification, which is not only inefficient, but also prone to human errors and difficult to adapt to complex and changing operating environments. Traditional protection setting management systems usually use fixed thresholds and preset rules to judge protection actions. This static method is stretched when faced with a dynamic and complex power system operating environment, and is prone to misjudgment or missed judgment, affecting the reliability and safety of the system.
[0003] In addition, traditional protection setting data management is highly dependent on manual entry and processing, and the accuracy and integrity of the data are difficult to guarantee. When processing a large amount of historical data and real-time data, information lag and low processing efficiency are prone to occur. This not only increases the workload of operators, but also leads to insufficient response speed of setting management, and the inability to adjust and optimize protection setting parameters in a timely manner. Most of the existing protection setting management systems are independently operated, lacking effective data sharing and collaborative management mechanisms, and cannot make full use of the experience and data resources between different power plants, resulting in poor overall optimization effects. At the same time, these systems lack the ability to deeply analyze and intelligently optimize protection setting data, and cannot dynamically adjust protection settings according to actual operating conditions. The system's adaptability and flexibility are insufficient. Especially in terms of abnormality detection and early warning, the existing systems generally lack early fault warning mechanisms, making it difficult to detect potential problems before faults occur, thus affecting the safe and stable operation of hydropower plants.
[0004] In view of these problems, there is an urgent need for an innovative protection setting management method and system that can overcome the defects of the existing technology to improve the efficiency, accuracy and intelligence level of protection setting management. The present invention aims to comprehensively improve the intelligence and automation level of protection setting management by introducing a variety of advanced artificial intelligence technologies and solve many deficiencies in the existing technology. Summary of the invention
[0005] The purpose of this invention is to propose a station-wide protection setting management method and system based on NLP. By combining self-supervised learning, time series analysis, generative adversarial networks (GANs), variational autoencoders (VAEs) and reinforcement learning technology, it aims to comprehensively improve the intelligence and automation level of protection setting management and solve many deficiencies in the prior art.
[0006] In order to achieve the above object, the present invention provides a method for managing the fixed value of the whole station protection based on NLP, the method comprising:
[0007] S1. Use sensors and monitoring equipment to collect protection fixed value data in real time, and pre-process the data to obtain pre-processed unlabeled data;
[0008] S2. Use a self-supervised learning algorithm to build a feature extraction model, use the feature extraction model to automatically learn the protection value features from unlabeled data, and then build a baseline model. Use the extracted protection value features to compare with the baseline model to detect abnormal changes and issue warnings in a timely manner;
[0009] S3. Build a hybrid deep prediction model based on the protection setting characteristics, use the time series prediction model to analyze the protection setting change trend of historical data, make predictions based on the analysis results, adjust the protection setting parameters in advance, and optimize the equipment operation status;
[0010] S4. Use a generative adversarial network to enhance the protection value data and generate high-quality simulated protection value data to enhance the training data set;
[0011] S5. Use the second variational autoencoder to perform dimension reduction and feature extraction on the protection constant value data after data enhancement to extract key features;
[0012] S6, using reinforcement learning algorithms to train intelligent agents to automatically adjust protection setting parameters based on environmental feedback;
[0013] The feature extraction model uses a one-dimensional convolutional neural network for feature extraction; the output of the convolutional layer of the one-dimensional convolutional neural network is expressed as follows:
[0014]
[0015] in, represents the output of the kth convolution kernel at position i, x i+j-1 Represents the value of the input data at position i+j-1, represents the weight of the kth convolution kernel, b (k) represents the bias, M represents the convolution kernel size;
[0016] Use autoencoders to encode the extracted features and generate low-dimensional feature representations;
[0017] Contrastive learning is introduced to design a loss function based on self-supervised learning, and the feature extraction model's ability to identify features is improved by constructing positive sample pairs and negative sample pairs; wherein the loss function L1 based on self-supervised learning is composed of reconstruction error, contrast loss, time consistency loss and sparsity loss;
[0018] The loss function L1 is expressed as follows:
[0019] L1=αL recon +βL contrastive +γL time +δL sparse
[0020] Among them, α, β, γ and δ are the weight coefficients of the loss term, which adjust the influence of reconstruction error, contrast loss, temporal consistency loss and sparsity loss respectively; L recon represents the reconstruction error, L contrastive represents the contrast loss, L time represents the temporal consistency loss, L sparse represents the sparsity loss;
[0021] The reconstruction error L recon It is expressed as follows:
[0022]
[0023] Among them, f i represents the i-th input feature, represents the i-th reconstruction feature, and n represents the number of features;
[0024] The contrast loss L contrastive It is expressed as follows:
[0025]
[0026] Among them, h i represents the low-dimensional feature representation of the i-th time step, k′ represents the offset of the randomly selected negative sample pair, δ represents the interval threshold, and h i+1 Represents the low-dimensional feature representation of the i+1th time step;
[0027] The temporal consistency loss L time It is expressed as follows:
[0028]
[0029] The sparsity loss L sparse It is expressed as follows:
[0030]
[0031] Furthermore, the preprocessing includes data cleaning, data normalization, data fusion and time alignment.
[0032] Furthermore, the baseline model trains the autoencoder through self-supervised learning, uses the low-dimensional feature representation generated by the autoencoder as the feature of the baseline model, and compares the features of the baseline model with the features at the current moment to calculate the anomaly score, which is expressed as follows:
[0033]
[0034] Among them, s t represents the anomaly score at the current time t, h t represents the feature representation at the current moment, h t-i represents the feature representation of the historical moment, N represents the historical window size, and ||·|| represents the Euclidean distance;
[0035] The anomaly score threshold is set according to the anomaly score at the current time t. When the anomaly score exceeds the threshold, an early warning is triggered, as shown below:
[0036] if s t >θ, then trigger alert
[0037] Among them, θ represents the preset anomaly score threshold, and then trigger alert means that the system will trigger an alert.
[0038] Further, the hybrid depth prediction model combines a long short-term memory network and a first variational autoencoder;
[0039] The calculation formula of the long short-term memory network is as follows:
[0040] f t =σ(W f ·[h t-1 , x t ]+b f )
[0041] i t =σ(W i ·[h t-1 , x t ]+b i )
[0042] o t =σ(W o ·[h t-1 , x t ]+b o )
[0043]
[0044] h t =o t tanh(C t )
[0045] Among them, f t represents the forget gate, i t represents the input gate, o t represents the output gate, represents the candidate memory unit, C t Represents a memory unit, h t represents the LSTM output, x t represents the input features, W f , W i , W o , W C represents the weight matrix, b f , b i , b o , b C represents the bias term, σ represents the Sigmoid activation function, and tanh represents the hyperbolic tangent function;
[0046] The reconstruction formula of the first variational autoencoder is expressed as follows:
[0047] z t ~q(z t |h t )
[0048]
[0049] Among them, z t represents a hidden variable, q(z t |h t ) represents the encoder, p(h t |z t ) represents the decoder, Represents the reconstruction feature.
[0050] Furthermore, an adaptive gating mechanism is introduced into the deep prediction model to adaptively fuse the outputs of the long short-term memory network and the variational autoencoder to adjust the weights of short-term and long-term dependencies; wherein the calculation formula of the adaptive gating mechanism is expressed as follows:
[0051]
[0052] Among them, g t represents the adaptive gating weight, h t represents the output of the long short-term memory network, represents the reconstructed features of the variational autoencoder, h represents the fused features, and W g represents the weight matrix, b grepresents the bias term;
[0053] The fused feature h′ t The final time series prediction value is generated through the fully connected layer, and the prediction error is calculated at the same time for deep prediction model training and optimization; wherein the final time series prediction value is expressed as follows:
[0054]
[0055] in, represents the final predicted value, W y represents the weight matrix of the fully connected layer, b y represents the bias term;
[0056] Dynamically adjust the protection setting parameters according to the prediction results; wherein the dynamic adjustment is expressed as follows:
[0057]
[0058] Among them, y t Indicates the actual value, then adjust parameters indicates adjusting the protection set value parameters.
[0059] Furthermore, the generative adversarial network introduces constraints of protection constants to design specific generator and discriminator structures; the generator adopts a multi-layer perceptron combined with a conditional generative adversarial network to generate data that meets the protection constant constraints through conditional input; the discriminator combines a convolutional neural network with conditional input to determine the authenticity of the data through conditional input.
[0060] Furthermore, a protection setting constraint loss function is designed according to the protection setting, which is expressed as follows:
[0061]
[0062] Among them, C(G(z,h′ t )) represents the constraint loss, G i (z,h′ t ) represents the i-th feature of the generated data, range i represents the constraint range of the i-th feature, N represents the total number of features, G(z, h′ t ) represents the generated data sample, z represents the generator input noise vector, h′ t Represents the fused features;
[0063] The discriminator loss function L D The calculation formula is as follows:
[0064]
[0065] Among them, L D represents the discriminator loss function, p data (h′) represents the real data distribution, p z (z) represents the noise distribution, h′ represents the key feature input, D(h′, h′ t ) represents the output of the discriminator;
[0066] The generator loss function L G The calculation formula is as follows:
[0067]
[0068] Among them, L G represents the generator loss function, p z (z) represents the noise distribution, C(·) represents the loss function of the protection setting constraint condition, λ represents the weight coefficient, G(z, h′ t ) represents the generated data sample.
[0069] Furthermore, in S5, the second variational autoencoder is used for feature dimensionality reduction and extraction, and the input of the second variational autoencoder is a high-dimensional feature The output is the mean and variance μ and σ of the latent variable z. The input of the second variational self-decoder is the latent variable z, which reconstructs the original high-dimensional features. Among them, the loss function L2 of the second variational autoencoder is composed of the reconstruction error L recon ′, KL divergence L KL And the protection constant constraint loss function composition;
[0070] Among them, the loss function L2 of the second variational autoencoder is expressed as follows:
[0071]
[0072] Among them, a and b represent the weight coefficients of the loss term, which adjust the influence of reconstruction error, KL divergence and constraint loss respectively; Represents the hidden variable z under a given high-dimensional feature Expected value under
[0073] The reconstruction error L recon ′, which is expressed as follows:
[0074]
[0075] in, represents the input features, Represents reconstruction features;
[0076] The KL divergence L KLIt is expressed as follows:
[0077]
[0078] Among them, L KL represents KL divergence, μ j and They represent the mean and variance of the latent variables respectively, and d represents the dimension of the latent variable;
[0079] The protection value constraint loss function It is expressed as follows:
[0080]
[0081] in, represents the constraint loss, Represents the i-th feature of the reconstructed feature, range i ′ represents the constraint range of the i-th feature of the reconstructed feature.
[0082] Furthermore, the reinforcement learning algorithm is designed as follows:
[0083] The state is represented by the low-dimensional feature representation z after the second variational autoencoder dimensionality reduction processing t ; The action space is the set of parameters for adjusting the protection setting value Reward function formula:
[0084]
[0085] in, Indicates time The reward value, indicates performance reward, Indicates safety reward, represents efficiency reward, μ, v, ω represent weight coefficients;
[0086] The deep deterministic policy gradient algorithm is used in combination with the attention mechanism, so that the policy network can adaptively focus on key features and dynamically adjust the protection constant parameters;
[0087] The Actor-Critic method is used in combination with the deep deterministic policy gradient algorithm for training to update the parameters of the policy network and the value network.
[0088] According to the actions output by the strategy network, the protection setting parameters are dynamically adjusted to optimize the equipment operation status.
[0089] In a second aspect of the present invention, a whole-station protection setting management system based on NLP is provided, the system comprising:
[0090] The data collection module is used to collect protection setting data in real time using sensors and monitoring equipment, and preprocess the data to obtain preprocessed unlabeled data;
[0091] The feature extraction module is used to build a feature extraction model using a self-supervised learning algorithm, use the feature extraction model to automatically learn the protection value features from unlabeled data, and then build a baseline model. The extracted protection value features are compared with the baseline model to detect abnormal changes and issue warnings in a timely manner;
[0092] The fixed value prediction module is used to build a hybrid deep prediction model based on the protection fixed value characteristics, use the time series prediction model to analyze the protection fixed value change trend of historical data, make predictions based on the analysis results, adjust the protection fixed value parameters in advance, and optimize the equipment operation status;
[0093] A data enhancement module is used to enhance the protection value data using a generative adversarial network to generate high-quality simulated protection value data to enhance the training data set;
[0094] A data dimension reduction module is used to use the second variational autoencoder to perform dimension reduction and feature extraction on the protection constant value data after data enhancement, and extract key features;
[0095] Reinforcement learning module, which is used to train intelligent agents using reinforcement learning algorithms and automatically adjust protection setting parameters based on environmental feedback;
[0096] The feature extraction model uses a one-dimensional convolutional neural network for feature extraction; the output of the convolutional layer of the one-dimensional convolutional neural network is expressed as follows:
[0097]
[0098] in, represents the output of the kth convolution kernel at position i, x i+j-1 Represents the value of the input data at position i+j-1, represents the weight of the kth convolution kernel, b (k) represents the bias, M represents the convolution kernel size;
[0099] Use autoencoders to encode the extracted features and generate low-dimensional feature representations;
[0100] Contrastive learning is introduced to design a loss function based on self-supervised learning, and the feature extraction model's ability to identify features is improved by constructing positive sample pairs and negative sample pairs; wherein the loss function L1 based on self-supervised learning is composed of reconstruction error, contrast loss, time consistency loss and sparsity loss;
[0101] The loss function L1 is expressed as follows:
[0102] L1=αL recon +βL contrastive +γL time +δL sparse
[0103] Among them, α, β, γ and δ are the weight coefficients of the loss term, which adjust the influence of reconstruction error, contrast loss, temporal consistency loss and sparsity loss respectively; L recon represents the reconstruction error, L contrastive represents the contrast loss, L time represents the temporal consistency loss, L sparse represents the sparsity loss;
[0104] The reconstruction error L recon It is expressed as follows:
[0105]
[0106] Among them, f i represents the i-th input feature, represents the i-th reconstruction feature, and n represents the number of features;
[0107] The contrast loss L contrastive It is expressed as follows:
[0108]
[0109] Among them, h i represents the low-dimensional feature representation of the i-th time step, k represents the offset of the randomly selected negative sample pair, δ represents the interval threshold, and h i+1 Represents the low-dimensional feature representation of the i+1th time step;
[0110] The temporal consistency loss L time It is expressed as follows:
[0111]
[0112] The sparsity loss L sparse It is expressed as follows:
[0113]
[0114] The beneficial technical effects of the present invention are at least as follows:
[0115] (1) The present invention aims at the specific needs of the protection setting management of hydropower plants and ensures the efficiency and reliability of data acquisition and preprocessing. The design of the sensor installation and data acquisition system combines the real-time and accuracy of high-frequency data. The data cleaning, normalization and time alignment steps ensure the quality and consistency of the data. Data fusion solves the problem of multi-sensor data consistency. Data storage and management and data quality monitoring ensure the long-term preservation and real-time monitoring of the data. The entire step is organically combined, laying a solid foundation for subsequent self-supervised learning, time series analysis, GAN data enhancement, VAE feature extraction and reinforcement learning, and realizing the intelligent and automated goals of the protection setting management system.
[0116] (2) The present invention automatically extracts the features of the protection setting data by designing a convolutional neural network and an autoencoder, and constructs a baseline model for anomaly detection. The innovative contrastive learning loss function combines reconstruction error, contrast loss, time consistency and sparsity, and improves the accuracy and robustness of feature extraction and anomaly detection based on the characteristics of the protection setting data of the hydropower plant. The entire step is organically combined to serve the entire patent solution, ensuring that the intelligent and automated goals of the protection setting management system are achieved.
[0117] (3) The present invention combines the LSTM and VAE models and introduces an adaptive gating mechanism (AGM) to design an innovative time series prediction method that can simultaneously capture the short-term and long-term dependencies of protection setting data. The input data is a low-dimensional feature representation extracted by self-supervised learning. The long-term and generative features are extracted through the LSTM-VAE layer, and then adaptively fused through the AGM mechanism to generate accurate time series prediction values. The protection setting parameters are dynamically adjusted according to the prediction results to optimize the equipment operation status. The entire step is organically combined to serve the entire patent solution, ensuring that the intelligent and automated goals of the protection setting management system are achieved.
[0118] (4) The present invention designs a generative adversarial network (GAN) and adopts a structure in which a generator and a discriminator compete with each other to generate data that is realistic and meets the constraints of the protection settings of the hydropower plant. The input data is a low-dimensional feature representation of time series analysis and prediction. The generator generates enhanced data samples, the discriminator distinguishes between true and false data, and an innovative loss function is used to ensure that the generated data meets the constraints of the protection settings of the hydropower plant. The trained generator generates a large amount of protection setting data, which is combined with the original data to enhance the diversity of the training data set and solve the problem of insufficient data. The entire step is organically combined to serve the entire patent solution, ensuring that the intelligent and automated goals of the protection setting management system are achieved.
[0119] (5) The present invention extracts low-dimensional feature representations of the protection setting data of the hydropower plant by designing a variational autoencoder (VAE). The input data is a data set enhanced by a generative adversarial network (GAN). The encoder maps the high-dimensional features to the latent variable space, and the decoder reconstructs the latent variables into high-dimensional features. The innovative loss function combines the reconstruction error, KL divergence, and protection setting constraints to ensure that the generated data meets the protection setting constraints. The entire step is organically combined to serve the entire patent solution, ensuring that the intelligent and automated goals of the protection setting management system are achieved.
[0120] (6) The present invention realizes automatic adjustment of protection constants by designing a reinforcement learning algorithm. The input data is a low-dimensional feature representation after VAE dimensionality reduction processing. The reward function comprehensively considers system performance, security and efficiency to ensure that the optimized protection constant parameters can improve the stability and security of the system. The policy network is combined with the attention mechanism to enable the model to adaptively focus on key features, dynamically adjust the protection constant parameters, and optimize the operating status of the equipment. The entire step is organically combined to serve the entire patent solution, ensuring that the intelligent and automated goals of the protection constant management system are achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0121] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.
[0122] Figure 1 This is a flow chart of the whole-station protection constant value management method based on NLP of the present invention.
[0123] Figure 2 This is a framework diagram of the NLP-based full-station protection setting management system of the present invention. DETAILED DESCRIPTION
[0124] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.
[0125] like Figure 1 As shown, the NLP-based full-station protection setting management method provided by the embodiment of the present invention includes the following steps S1-S6:
[0126] S1. Use sensors and monitoring equipment to collect protection constant value data in real time, and preprocess the data to obtain preprocessed unlabeled data.
[0127] Specifically, high-precision sensors are installed on key equipment and nodes (such as generators, transformers and transmission lines) of hydropower plants to collect real-time protection parameters such as current, voltage, power, frequency and temperature. Fiber optic sensors and wireless sensor networks (WSN) are used to ensure the stability and anti-interference ability of data transmission.
[0128] Build a distributed data acquisition system based on edge computing, deploy edge computing nodes on site, perform preliminary data processing and caching, and reduce data transmission delays. Set the data acquisition frequency to 50Hz to ensure sufficient timeliness and accuracy to adapt to high-frequency power system data. Use network clock synchronization protocols (such as PTP: Precision Time Protocol) to achieve precise synchronization of sensor data and ensure time consistency of different sensor data.
[0129] Furthermore, the preprocessing includes:
[0130] Data cleaning: Apply a bandpass filter to remove high-frequency noise. The formula is as follows:
[0131]
[0132] Among them, y(t) is the filtered data, x(t) is the original data, and a i and b j are the filter coefficients, and N and M are the filter orders.
[0133] Outlier detection: Detect and remove outliers using statistical analysis-based methods and calculate the z-score for each data point:
[0134]
[0135] Where x is the data point value, p is the data mean, and σ is the data standard deviation. If |z|>3, it is considered an outlier.
[0136] The current, voltage and other data are standardized and the dimensions are unified. The formula is as follows:
[0137]
[0138] Among them, x is the original data, p is the mean, σ is the standard deviation, and x′ is the standardized data.
[0139] Perform spatial and temporal fusion of different sensor data to ensure data consistency. Kalman filtering is used for data fusion, and the formula is as follows:
[0140]
[0141] in, is the estimated value after fusion, z k is the observation value, H is the observation matrix, K k is the Kalman gain.
[0142] For unsynchronized data points, linear interpolation method is used for time alignment, the formula is as follows:
[0143]
[0144] Where t is the time of the interpolation point, t1 and t2 are the times of adjacent known data points, and x(t1) and x(t2) are the values of adjacent known data points.
[0145] Furthermore, the preprocessed data is stored in a distributed database in a time series format, and each record contains parameters such as timestamp, current, voltage, power, frequency, and temperature. The database uses a time series database (such as InfluxDB) to support large-scale data storage and fast query.
[0146] Back up data regularly to prevent data loss, adopt an incremental backup strategy, and back up new data every day. Configure a disaster recovery mechanism to ensure that data can be quickly restored in the event of system failure or data corruption.
[0147] Furthermore, the quality of data collection and preprocessing is monitored in real time, including indicators such as data packet loss rate, data noise level, and data alignment accuracy:
[0148] Data packet loss rate monitoring: Count the number of lost data points in each time window and calculate the packet loss rate.
[0149] Data noise level monitoring: Calculate the standard deviation of the data and assess the noise level of the data.
[0150] Data alignment accuracy monitoring: Compare the timestamps of different sensors to evaluate the data alignment error.
[0151] At the same time, a data quality alarm mechanism is set up. When the data quality index exceeds the preset threshold, an alarm is automatically triggered and a log is recorded. If the packet loss rate exceeds 1%, the noise level exceeds twice the standard deviation, and the alignment error exceeds 1ms, an alarm is triggered.
[0152] S2. Use a self-supervised learning algorithm to build a feature extraction model, use the feature extraction model to automatically learn the protection value features from unlabeled data, and then build a baseline model. Use the extracted protection value features to compare with the baseline model to detect abnormal changes and issue timely warnings.
[0153] Specifically, the input data of the feature extraction model is preprocessed time series data such as current, voltage, power, frequency and temperature. A one-dimensional convolutional neural network (CNN) is used for feature extraction to capture local patterns and global trends of the data. CNN contains multiple convolutional layers, pooling layers and fully connected layers. Among them, the convolutional layer output calculation formula is:
[0154]
[0155] in, is the output of the kth convolution kernel at position i, x i+j-1 is the value of the input data at position i+j-1, is the weight of the kth convolution kernel, b (k) is the bias, and M is the convolution kernel size.
[0156] Then, an autoencoder is used to encode the extracted features to generate a low-dimensional feature representation. The autoencoder consists of an encoder and a decoder.
[0157] Encoding process formula:
[0158] h=σ(W e ·f+b e ) (7)
[0159] Among them, h is the low-dimensional feature representation, f is the input feature, and W e is the encoder weight matrix, b e is the bias, and σ is the activation function (such as ReLU).
[0160] The decoding process formula of the autoencoder is:
[0161]
[0162] in, is the reconstructed high-dimensional feature, W d is the decoder weight matrix, b d For bias.
[0163] Furthermore, in order to better adapt to the characteristics of hydropower plant protection setting data, a special one is designed to improve the accuracy of anomaly detection. The contrastive learning method is introduced to improve the model's ability to identify features by constructing positive sample pairs and negative sample pairs.
[0164] The feature extraction model loss function consists of reconstruction error, contrast loss, temporal consistency loss, and sparsity loss. Among them, reconstruction error:
[0165]
[0166] Among them, fi is the i-th input feature, is the i-th reconstruction feature, and n is the number of features.
[0167] Contrastive loss: positive pairs (adjacent time steps) should be more similar, and negative pairs (random time steps) should be more different.
[0168]
[0169] Among them, h i is the low-dimensional feature representation of the i-th time step, k is the offset of the randomly selected negative sample pair, and δ is the interval threshold.
[0170] Temporal consistency loss: ensures the smoothness and continuity of feature representation in time.
[0171]
[0172] Among them, h i is the low-dimensional feature representation of the i-th time step.
[0173] Sparsity loss: encourages the sparsity of feature representation and improves the robustness of the model.
[0174]
[0175] Total loss function:
[0176] L1=αL recon +βL contrastive +γL time +δL sparse (13)
[0177] Among them, α, β, γ and δ are the weight coefficients of the loss term, which adjust the influence of reconstruction error, contrast loss, temporal consistency loss and sparsity loss respectively.
[0178] Furthermore, the construction of the baseline model: the autoencoder is trained through self-supervised learning, and the low-dimensional feature representation generated by the encoder is used as the feature of the baseline model. The features of the baseline model are compared with the features at the current moment to calculate the anomaly score. The anomaly score formula is:
[0179]
[0180] Among them, s t is the anomaly score at the current time t, h t is the feature representation of the current moment, h t-i is the feature representation of the historical moment, N is the historical window size, and ||·|| is the Euclidean distance.
[0181] Furthermore, an abnormality score threshold is set, and when the abnormality score exceeds the threshold, an early warning is triggered. Anomaly detection early warning mechanism:
[0182] if s t >θ,then trigger alert (15)
[0183] Here, θ is the preset anomaly score threshold, and then trigger alert means that the system will trigger an alert. This usually means that the system will perform certain predefined actions in response to the detected anomaly. Specific actions may include but are not limited to:
[0184] Send alert notifications: The system may send emails, text messages, or other forms of notifications to relevant personnel to inform them that an abnormality has been detected.
[0185] Log: The system may record this event in a log file for future review and analysis.
[0186] Execute automated actions: The system may take automated response actions, such as starting a backup system, adjusting system parameters, or executing failure recovery steps.
[0187] Generate report: The system may generate a detailed exception report containing the time, characteristics and other relevant information of the exception.
[0188] Triggering other system events: The system may interact with other monitoring or management systems, triggering a series of response events.
[0189] S3. Construct a hybrid deep prediction model based on the characteristics of protection settings, use the time series prediction model to analyze the trend of protection setting changes on historical data, make predictions based on the analysis results, adjust the protection setting parameters in advance, and optimize the equipment operation status.
[0190] Specifically, the input data of the hybrid depth prediction model is the low-dimensional feature representation h extracted through self-supervised learning. t , including current, voltage, power, frequency and temperature. The key features are selected as the input of the time series prediction model to capture the changing trend of the protection setting.
[0191] An improved hybrid deep prediction model (Hybrid Deep Prediction Model) is adopted, combining the long short-term memory network (LSTM) and the variational autoencoder (VAE), and a new adaptive gating mechanism (Adaptive Gating Mechanism, AGM) is designed to simultaneously capture the short-term and long-term dependencies of the data and solve the problems that the traditional LSTM cannot fully cope with.
[0192] Among them, a multi-layer LSTM network is constructed to extract long-term dependencies in the time series, and VAE is combined for feature reconstruction to enhance the generation ability and robustness of the model. LSTM unit calculation formula:
[0193]
[0194] Among them, f t For the forget gate, i t is the input gate, o t is the output gate, is a candidate memory unit, C t is the memory unit, h t is the LSTM output, x t is the input feature, W f , W i , W o , W C is the weight matrix, b f , b i , b o , b C is the bias term, σ is the Sigmoid activation function, and tanh is the hyperbolic tangent function.
[0195] VAE reconstruction formula:
[0196]
[0197] Among them, z t is a hidden variable, q(z t |h t ) is the encoder, p(h t |z t ) is a decoder, To reconstruct features.
[0198] Furthermore, an adaptive gating mechanism (AGM) is introduced to adaptively fuse the outputs of LSTM and VAE and adjust the weights of short-term and long-term dependencies. AGM calculation formula:
[0199]
[0200] Among them, gt is the adaptive gating weight, h t is the LSTM output, Reconstruct features for VAE, h′ t is the fused feature, W g is the weight matrix, b g is the bias term.
[0201] The fused feature h′ t The final time series prediction value is generated through the fully connected layer. The final prediction value calculation formula is:
[0202]
[0203] in, is the final predicted value, W y is the weight matrix of the fully connected layer, b y is the bias term.
[0204] Calculate the prediction error for model training and optimization. Prediction error formula:
[0205]
[0206] Among them, Lpred is the prediction error, yt is the true value, is the predicted value, and T is the time step.
[0207] Dynamically adjust the protection setting parameters according to the prediction results to optimize the equipment operation status. Dynamic adjustment formula:
[0208]
[0209] in, is the predicted value, yt is the actual value, and the parameter adjustment strategy is set according to the specific protection setting requirements. Thenadjust parameters means adjusting the protection setting parameters. The specific adjustment process may include:
[0210] Threshold adjustment: Dynamically adjust the trigger threshold of the protection device based on the prediction error. For example, in the power system, adjust the current or voltage trigger threshold of the relay protection device to improve the sensitivity and accuracy of fault detection.
[0211] Delay parameter adjustment: Adjust the operating delay parameters of the protection device to ensure timely response when an abnormality is detected to avoid false or delayed operation. For example, adjust the operating delay parameters of the circuit breaker so that it can cut off the fault line more quickly.
[0212] Sensitivity adjustment: According to the actual situation, adjust the sensitivity parameters of the protection device to more accurately detect and respond to abnormal conditions. This can include adjusting the sensitivity of the sensor or adjusting the parameters of the signal processing algorithm.
[0213] Adjustment of reclosing times and intervals: In some protection systems, the number of reclosing operations and time intervals are adjusted to improve the system's self-healing ability and stability.
[0214] Adjustment of other protection parameters: According to the specific application scenario, adjust other protection-related parameters, such as the setting of protection area, configuration of protection logic, etc., to optimize the protection performance of the system.
[0215] S4. Use a generative adversarial network to perform data enhancement on the protection value data and generate high-quality simulated protection value data to enhance the training data set.
[0216] Specifically, the input data of the generative adversarial network is the low-dimensional feature representation h′ predicted by time series analysis t , including features such as current, voltage, power, frequency and temperature. Key features are selected as the input of the generative adversarial network to enhance the data diversity of the protection setting.
[0217] The generative adversarial network adopts a structure of mutual game between the generator and the discriminator to generate realistic protection setting data. In order to address the particularity of the protection setting of hydropower plants, a specific generator and discriminator structure is designed, and the constraints of the protection setting are introduced.
[0218] The design is as follows:
[0219] Generator input noise vector And combined with the key feature input h′ t , generate enhanced data samples G(z, h′ t ).
[0220] The generator structure combines a multi-layer perceptron (MLP) with a conditional generative adversarial network (ConditionalGAN, cGAN) to generate data that meets the protection value constraints through conditional input. The output formula of the generator is:
[0221] G(z,h′ t )=σ(W g ·[z,h′ t ]+b g ) (twenty two)
[0222] Among them, G(z, h′ t ) is the generated data sample, [z, h′ t ] is the combination of noise vector and key features, W g is the weight matrix, b g is the bias term, and σ is the activation function (such as ReLU).
[0223] The discriminator inputs the real data h′ t and generate data G(z,h′ t ), output the true or false probability D(h′) t ) and D(G(z,h′ t )).
[0224] The discriminator structure combines convolutional neural network (CNN) with conditional input to determine the authenticity of data through conditional input. The output formula of the discriminator is:
[0225] D(x,h′ t )=σ(W d ·[x,h′ t ]+b d ) (twenty three)
[0226] Where D(x, h′ t ) is the probability of the discriminator output, [x, h′ t ] is the combination of input data and key features, W d is the weight matrix, b d is the bias term, and σ is the Sigmoid activation function.
[0227] Furthermore, in view of the particularity of the protection setting data of the hydropower plant, an innovative loss function including the protection setting constraint is designed. A constraint loss function C(·) is designed to ensure that the generated data meets the protection setting constraint of the hydropower plant. The protection setting constraint loss function formula is:
[0228]
[0229] Where c(G(z, h′ t )) is the constraint loss, G i (z,h′ t ) is the i-th feature of the generated data, range i is the constraint range of the feature.
[0230] Furthermore, the discriminator loss function L D Calculation formula:
[0231]
[0232] Among them, L D is the discriminator loss, p data (h′) is the real data distribution, p z (z) is the noise distribution.
[0233] Generator loss function L G Calculation formula:
[0234]
[0235] Among them, L G is the generator loss, p z (z) is the noise distribution, C(·) is the loss function of the protection constant constraint condition, and λ is the weight coefficient.
[0236] The generator and the discriminator are trained alternately to enhance each other's capabilities and generate high-quality protection value data.
[0237] Furthermore, the discriminator update formula is:
[0238]
[0239] Among them, θ d is the discriminator parameter, η is the learning rate, is the gradient of the discriminator loss function.
[0240] Generator update formula:
[0241]
[0242] Among them, θ g is the generator parameter, η is the learning rate, is the gradient of the generator loss function.
[0243] Use the trained generator to generate a large amount of protection value data and combine it with the original data to enhance the diversity of the training data set. Data enhancement formula:
[0244]
[0245] in, is the enhanced data, h′ is the original data, G(z, h′ t ) is the generated data, and λ is the fusion weight coefficient.
[0246] S5. Use the second variational autoencoder to perform dimensionality reduction and feature extraction on the protection constant value data after data enhancement to extract key features.
[0247] Specifically, the input data of the second variational autoencoder is a dataset enhanced by a generative adversarial network (GAN) It contains features such as current, voltage, power, frequency and temperature. The key features are selected as the input of the variational autoencoder to extract the low-dimensional feature representation of the protection constant.
[0248] The second variational autoencoder inputs high-dimensional features Output the mean and variance p and σ of the latent variable z to generate the latent variable z. The second variational autoencoder output formula is:
[0249]
[0250] Among them, p is the mean of the latent variable, logσ 2 is the logarithm of the variance of the latent variable, z is the latent variable, W e is the encoder weight matrix, b e is the bias term, f μ and f σ is the activation function.
[0251] The second variational autoencoder inputs the latent variable z and reconstructs the original high-dimensional features Decoder output formula:
[0252]
[0253] in, is the reconstructed high-dimensional feature, W d is the decoder weight matrix, b d is the bias term, f d is the activation function.
[0254] Furthermore, in view of the particularity of the protection setting data of hydropower plants, an innovative loss function including the protection setting constraint is designed. The loss function consists of reconstruction error, KL divergence and protection setting constraint.
[0255] Reconstruction error formula:
[0256]
[0257] Among them, L recon ′ is the reconstruction error, is the input feature, To reconstruct features.
[0258] KL divergence formula:
[0259]
[0260] Among them, L KL is the KL divergence, μ j and are the mean and variance of the latent variables respectively, and d is the dimension of the latent variable.
[0261] Design the protection constant value constraint condition loss function C(·) to ensure that the generated data meets the protection constant value constraint. The protection constant value constraint loss function formula is:
[0262]
[0263] in, is the constraint loss, is the i-th value of the reconstructed feature, range i is the constraint range of the feature.
[0264] Total loss function formula: The loss function L2 of the second variational autoencoder is expressed as follows:
[0265]
[0266] Among them, a and b represent the weight coefficients of the loss term, which adjust the influence of reconstruction error, KL divergence and constraint loss respectively; Represents the hidden variable z under a given high-dimensional feature Expected value under
[0267] Furthermore, the parameters of the encoder and decoder are optimized by minimizing the total loss function.
[0268] Encoder and decoder update formula:
[0269]
[0270] Among them, θ e and θ d are the parameters of the encoder and decoder respectively, η is the learning rate, and are the gradients of the total loss function with respect to the encoder and decoder parameters, respectively.
[0271] S6. Use reinforcement learning algorithm to train intelligent agents and automatically adjust protection setting parameters according to environmental feedback.
[0272] Specifically, the state of the reinforcement learning algorithm is represented by the low-dimensional feature representation z after VAE dimensionality reduction processing t , including current, voltage, power, frequency and temperature. The action space is the set of parameters for adjusting the protection setting. Including adjusting current constant value, voltage constant value, etc.
[0273] Furthermore, a reward function that comprehensively considers system performance and safety is designed to ensure that the optimized protection setting parameters can improve the stability and safety of the system. The innovation of the reward function lies in combining the unique operating parameters and constraints of the hydropower plant, such as water level, flow, power generation, etc. Reward function formula:
[0274]
[0275] in, Indicates time The reward value, indicates performance reward, Indicates safety reward, represents efficiency reward, μ, v, ω represent weight coefficients.
[0276] Furthermore, the Deep Deterministic Policy Gradient (DDPG) algorithm is used in combination with the attention mechanism to enable the policy network to adaptively focus on key features and dynamically adjust the protection constant parameters. The policy network input state s t , output action a t The probability distribution of π(a|s;θ π), introduce the attention layer, adaptively select key features, and the strategy network formula is:
[0277] a t =π(s t θ π , A t ) (39)
[0278] Among them, a t The action chosen for time t, s t is the state, θ π is the network parameter, A t is the attention weight.
[0279] Furthermore, the attention mechanism adaptively adjusts the feature weights so that the model can better capture key features. The attention weight calculation formula is:
[0280] A t =softmax(W a ·s t +b a ) (40)
[0281] Among them, A t is the attention weight, W a is the weight matrix, b a is the bias term.
[0282] Preferably, the Actor-Critic method is used in combination with the DDPG algorithm for training to update the parameters of the policy network and the value network. Policy network parameters θ π Update formula:
[0283]
[0284] Among them, η π is the learning rate, δ t is the timing difference error at time t.
[0285] Value network parameter θ v Update formula:
[0286]
[0287] Among them, η v is the learning rate and γ is the discount factor.
[0288] According to the actions output by the strategy network, the protection setting parameters are dynamically adjusted to optimize the equipment operation status.
[0289] Dynamic adjustment formula:
[0290] a t ~π(a|s t θπ ) (43)
[0291] Among them, a t The action chosen for time t, π(a|s t θ π ) is the action probability distribution output by the policy network.
[0292] like Figure 2 As shown, another embodiment of the present invention provides a full-station protection setting management system based on NLP, the system comprising:
[0293] The data collection module 101 is used to collect protection fixed value data in real time using sensors and monitoring equipment, and pre-process the data to obtain pre-processed unlabeled data;
[0294] The feature extraction module 102 is used to construct a feature extraction model using a self-supervised learning algorithm, automatically learn the protection value features from the unlabeled data using the feature extraction model, and then construct a baseline model, and compare the extracted protection value features with the baseline model to detect abnormal changes and issue warnings in a timely manner;
[0295] The fixed value prediction module 103 is used to construct a hybrid depth prediction model according to the protection fixed value characteristics, use the time series prediction model to analyze the protection fixed value change trend of historical data, make predictions based on the analysis results, adjust the protection fixed value parameters in advance, and optimize the equipment operation status;
[0296] A data enhancement module 104 is used to enhance the protection value data using a generative adversarial network to generate high-quality simulated protection value data to enhance the training data set;
[0297] A data dimension reduction module 105 is used to use a second variational autoencoder to perform dimension reduction and feature extraction on the protection constant value data after data enhancement, and extract key features;
[0298] A reinforcement learning module 106, for training the intelligent agent using a reinforcement learning algorithm to automatically adjust the protection setting parameters according to environmental feedback;
[0299] The feature extraction model uses a one-dimensional convolutional neural network for feature extraction; the output of the convolutional layer of the one-dimensional convolutional neural network is expressed as follows:
[0300]
[0301] in, represents the output of the kth convolution kernel at position i, x i+j-1 Represents the value of the input data at position i+j-1, represents the weight of the kth convolution kernel, b(k) represents the bias, M represents the convolution kernel size;
[0302] Use autoencoders to encode the extracted features and generate low-dimensional feature representations;
[0303] Contrastive learning is introduced to design a loss function based on self-supervised learning, and the feature extraction model's ability to identify features is improved by constructing positive sample pairs and negative sample pairs; wherein the loss function L1 based on self-supervised learning is composed of reconstruction error, contrast loss, time consistency loss and sparsity loss;
[0304] The loss function L1 is expressed as follows:
[0305] L1=αL recon +βL contrastive +γL time +δL sparse
[0306] Among them, α, β, γ and δ are the weight coefficients of the loss term, which adjust the influence of reconstruction error, contrast loss, temporal consistency loss and sparsity loss respectively; L recon represents the reconstruction error, L contrastive represents the contrast loss, L time represents the temporal consistency loss, L sparse represents the sparsity loss;
[0307] The reconstruction error L recon It is expressed as follows:
[0308]
[0309] Among them, fi represents the i-th input feature, represents the i-th reconstruction feature, and n represents the number of features;
[0310] The contrast loss L contrastive It is expressed as follows:
[0311]
[0312] Among them, h i represents the low-dimensional feature representation of the i-th time step, k represents the offset of the randomly selected negative sample pair, δ represents the interval threshold, and h i+1 Represents the low-dimensional feature representation of the i+1th time step;
[0313] The temporal consistency loss L time It is expressed as follows:
[0314]
[0315] The sparsity loss L sparseIt is expressed as follows:
[0316]
[0317] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of them according to actual needs to implement g of the scheme of this embodiment, and there is no limitation here.
[0318] In addition, for technical details that are not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, and will not be repeated here.
[0319] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.
[0320] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.
[0321] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.
[0322] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. The NLP-based full-station protection setting management method is characterized by: The method comprises: S1. Use sensors and monitoring equipment to collect protection fixed value data in real time, and pre-process the data to obtain pre-processed unlabeled data; S2. Use a self-supervised learning algorithm to build a feature extraction model, use the feature extraction model to automatically learn the protection value features from unlabeled data, and then build a baseline model. Use the extracted protection value features to compare with the baseline model to detect abnormal changes and issue warnings in a timely manner; S3. Build a hybrid deep prediction model based on the protection setting characteristics, use the time series prediction model to analyze the protection setting change trend of historical data, make predictions based on the analysis results, adjust the protection setting parameters in advance, and optimize the equipment operation status; S4. Use a generative adversarial network to enhance the protection value data and generate high-quality simulated protection value data to enhance the training data set; S5. Use the second variational autoencoder to perform dimension reduction and feature extraction on the protection constant value data after data enhancement to extract key features; S6. Use a reinforcement learning algorithm to train the intelligent agent to automatically adjust the protection setting parameters according to environmental feedback; the reinforcement learning algorithm is designed as follows: The state is represented by the low-dimensional feature representation z after the second variational autoencoder dimensionality reduction processing t ; The action space is the set of parameters for adjusting the protection setting value Reward function formula: in, Indicates time The reward value, indicates performance reward, Indicates safety reward, represents efficiency reward, μ, ν, ω represent weight coefficients; The deep deterministic policy gradient algorithm is used in combination with the attention mechanism, so that the policy network can adaptively focus on key features and dynamically adjust the protection constant parameters; The Actor-Critic method is used in combination with the deep deterministic policy gradient algorithm for training to update the parameters of the policy network and the value network. According to the actions output by the strategy network, dynamically adjust the protection setting parameters to optimize the equipment operation status; The feature extraction model uses a one-dimensional convolutional neural network for feature extraction; the output of the convolutional layer of the one-dimensional convolutional neural network is expressed as follows: in, represents the output of the kth convolution kernel at position i, x i+j-1 Represents the value of the input data at position i+j-1, represents the weight of the kth convolution kernel, b (k) represents the bias, M represents the convolution kernel size; Use autoencoders to encode the extracted features and generate low-dimensional feature representations; Contrastive learning is introduced to design a loss function based on self-supervised learning, and the feature extraction model's ability to identify features is improved by constructing positive sample pairs and negative sample pairs; wherein the loss function L1 based on self-supervised learning is composed of reconstruction error, contrast loss, time consistency loss and sparsity loss; The loss function L1 is expressed as follows: L1=αL recon +βL contrastive +γL time +δL sparse ; Among them, α, β, γ and δ are the weight coefficients of the loss term, which adjust the influence of reconstruction error, contrast loss, temporal consistency loss and sparsity loss respectively; L recon represents the reconstruction error, L contrastive represents the contrast loss, L time represents the temporal consistency loss, L sparse represents the sparsity loss; The reconstruction error L recon It is expressed as follows: Among them, f i represents the i-th input feature, represents the i-th reconstruction feature, and n represents the number of features; The contrast loss L contrastive It is expressed as follows: Among them, h i represents the low-dimensional feature representation of the i-th time step, k′ represents the offset of the randomly selected negative sample pair, δ represents the interval threshold, and h i+1 Represents the low-dimensional feature representation of the i+1th time step; The temporal consistency loss L time It is expressed as follows: The sparsity loss L sparse It is expressed as follows:
2. The NLP-based full-station protection setting management method according to claim 1 is characterized in that: The preprocessing includes data cleaning, data normalization, data fusion and time alignment.
3. The NLP-based full-station protection setting management method according to claim 1 is characterized in that: The baseline model trains the autoencoder through self-supervised learning, uses the low-dimensional feature representation generated by the autoencoder as the feature of the baseline model, and compares the features of the baseline model with the features at the current moment to calculate the anomaly score, which is expressed as follows: Among them, s t represents the anomaly score at the current time t, h t represents the feature representation at the current moment, h t-i represents the feature representation of the historical moment, N represents the historical window size, and ||·|| represents the Euclidean distance; The anomaly score threshold is set according to the anomaly score at the current time t. When the anomaly score exceeds the threshold, an alert is triggered, as shown below: if s t >θ,then trigger alert; Among them, θ represents the preset anomaly score threshold, and then trigger alert means that the system will trigger an alert.
4. The NLP-based full-station protection setting management method according to claim 1 is characterized in that: The hybrid deep prediction model combines a long short-term memory network and a first variational autoencoder; The calculation formula of the long short-term memory network is as follows: Among them, f t represents the forget gate, i t represents the input gate, o t represents the output gate, represents the candidate memory unit, C t Represents a memory unit, h t represents the LSTM output, x t represents the input features, W f ,W i ,W o ,W C represents the weight matrix, b f ,b i ,b o ,b C represents the bias term, σ represents the Sigmoid activation function, and tanh represents the hyperbolic tangent function; The reconstruction formula of the first variational autoencoder is expressed as follows: Among them, z t represents a hidden variable, q(z t |h t ) represents the encoder, p(h t |z t ) represents the decoder, Represents the reconstruction feature.
5. The NLP-based full-station protection setting management method according to claim 4 is characterized in that: An adaptive gating mechanism is introduced into the deep prediction model to adaptively fuse the outputs of the long short-term memory network and the variational autoencoder to adjust the weights of short-term and long-term dependencies. The calculation formula of the adaptive gating mechanism is expressed as follows: Among them, g t represents the adaptive gating weight, h t represents the output of the long short-term memory network, represents the reconstruction features of the variational autoencoder, W g represents the weight matrix, b g represents the bias term of the adaptive gating mechanism, h′ t Represents the fused features; The fused feature h′ t The final time series prediction value is generated through the fully connected layer, and the prediction error is calculated at the same time for deep prediction model training and optimization; wherein the final time series prediction value is expressed as follows: in, represents the final predicted value, W y represents the weight matrix of the fully connected layer, b y represents the bias term of the fully connected layer; Dynamically adjust the protection setting parameters according to the prediction results; wherein the dynamic adjustment is expressed as follows: Among them, y t Indicates the actual value, then adjust parameters indicates adjusting the protection set value parameters.
6. The NLP-based full-station protection setting management method according to claim 1 is characterized in that: The generative adversarial network introduces the constraint conditions of the protection constant to design a specific generator and discriminator structure; the generator adopts a multi-layer perceptron combined with a conditional generative adversarial network to generate data that meets the protection constant constraint through conditional input; the discriminator combines the convolutional neural network with the conditional input to judge the authenticity of the data through the conditional input.
7. The NLP-based full-station protection setting management method according to claim 6 is characterized in that: The protection setting constraint loss function is designed according to the protection setting, which is expressed as follows: Among them, C(G(z,h′ t )) represents the constraint loss, G i (z,h′ t ) represents the i-th feature of the generated data, range o represents the constraint range of the i-th feature, N represents the total number of features, G(z,h′ t ) represents the generated data sample, z represents the generator input noise vector, h′ t Represents the fused features; Discriminator loss function L D The calculation formula is as follows: Among them, L D represents the discriminator loss function, p data (h′) represents the real data distribution, p z (z) represents the noise distribution, h′ represents the key feature input, D(h′,h′ t ) represents the output of the discriminator; Generator loss function L G The calculation formula is as follows: Among them, L G represents the generator loss function, p z (z) represents the noise distribution, C(·) represents the loss function of the protection setting constraint condition, λ represents the weight coefficient, G(z,h′ t ) represents the generated data sample.
8. The NLP-based full-station protection setting management method according to claim 1 is characterized in that: In S5, the second variational autoencoder is used for feature dimensionality reduction and extraction, and the input of the second variational autoencoder is a high-dimensional feature. The output is the mean and variance μ and σ of the latent variable z. The input of the second variational self-decoder is the latent variable z, which reconstructs the original high-dimensional features. Among them, the loss function L2 of the second variational autoencoder is composed of the reconstruction error L recon ′, KL divergence L KL And the protection constant constraint loss function composition; Among them, the loss function L2 of the second variational autoencoder is expressed as follows: Among them, a and b represent the weight coefficients of the loss term, which adjust the influence of reconstruction error, KL divergence and constraint loss respectively; Represents the hidden variable z under a given high-dimensional feature Expected value under The reconstruction error L recon ′, which is expressed as follows: in, represents the input features, Represents reconstruction features; The KL divergence L KL It is expressed as follows: Among them, L KL represents KL divergence, μ j and They represent the mean and variance of the latent variables respectively, and d represents the dimension of the latent variable; The protection value constraint loss function It is expressed as follows: in, represents the constraint loss, Represents the i-th feature of the reconstructed feature, range i ′ represents the constraint range of the i-th feature of the reconstructed feature.
9. The NLP-based full-station protection setting management system is characterized by: The system comprises: The data collection module is used to collect protection setting data in real time using sensors and monitoring equipment, and preprocess the data to obtain preprocessed unlabeled data; The feature extraction module is used to build a feature extraction model using a self-supervised learning algorithm, use the feature extraction model to automatically learn the protection value features from unlabeled data, and then build a baseline model. The extracted protection value features are compared with the baseline model to detect abnormal changes and issue warnings in a timely manner; The fixed value prediction module is used to build a hybrid deep prediction model based on the protection fixed value characteristics, use the time series prediction model to analyze the protection fixed value change trend of historical data, make predictions based on the analysis results, adjust the protection fixed value parameters in advance, and optimize the equipment operation status; A data enhancement module is used to enhance the protection value data using a generative adversarial network to generate high-quality simulated protection value data to enhance the training data set; A data dimension reduction module is used to use the second variational autoencoder to perform dimension reduction and feature extraction on the protection constant value data after data enhancement, and extract key features; The reinforcement learning module is used to train the intelligent agent using the reinforcement learning algorithm to automatically adjust the protection setting parameters according to environmental feedback; the reinforcement learning algorithm is designed as follows: The state is represented by the low-dimensional feature representation z after the second variational autoencoder dimensionality reduction processing t ; The action space is the set of parameters for adjusting the protection setting value Reward function formula: in, Indicates time The reward value, indicates performance reward, Indicates safety reward, represents efficiency reward, μ, ν, ω represent weight coefficients; The deep deterministic policy gradient algorithm is used in combination with the attention mechanism, so that the policy network can adaptively focus on key features and dynamically adjust the protection constant parameters; The Actor-Critic method is used in combination with the deep deterministic policy gradient algorithm for training to update the parameters of the policy network and the value network. According to the actions output by the strategy network, dynamically adjust the protection setting parameters to optimize the equipment operation status; The feature extraction model uses a one-dimensional convolutional neural network for feature extraction; the output of the convolutional layer of the one-dimensional convolutional neural network is expressed as follows: in, represents the output of the kth convolution kernel at position i, x i+j-1 Represents the value of the input data at position i+j-1, represents the weight of the kth convolution kernel, b (k) represents the bias, M represents the convolution kernel size; Use autoencoders to encode the extracted features and generate low-dimensional feature representations; Contrastive learning is introduced to design a loss function based on self-supervised learning, and the feature extraction model's ability to identify features is improved by constructing positive sample pairs and negative sample pairs; wherein the loss function L1 based on self-supervised learning is composed of reconstruction error, contrast loss, time consistency loss and sparsity loss; The loss function L1 is expressed as follows: L1=αL recon +βL contrastive +γL time +δL sparse ; Among them, α, β, γ and δ are the weight coefficients of the loss term, which adjust the influence of reconstruction error, contrast loss, temporal consistency loss and sparsity loss respectively; L recon represents the reconstruction error, L contrastive represents the contrast loss, L time represents the temporal consistency loss, L sparse represents the sparsity loss; The reconstruction error L recon It is expressed as follows: Among them, f i represents the i-th input feature, represents the i-th reconstruction feature, and n represents the number of features; The contrast loss L contrastive It is expressed as follows: Among them, h i represents the low-dimensional feature representation of the i-th time step, k′ represents the offset of the randomly selected negative sample pair, δ represents the interval threshold, and h i+1 Represents the low-dimensional feature representation of the i+1th time step; The temporal consistency loss L time It is expressed as follows: The sparsity loss L sparse It is expressed as follows:
Citation Information
Patent Citations
Real-time control system and method for power generation efficiency of power plant
CN117477655A
Unsupervised reconstruction network anomaly detection method and system based on pseudo tag
CN118053047A
Substation communication network equipment monitoring and fault early warning method and system
CN118200118A
Network security early warning method and system based on deep learning
CN118353667A
