A method for remote monitoring and control of power system data

Abstract

The present invention discloses a method for remote monitoring and control of power system data, which relates to the field of power system monitoring technology. The method comprises the following steps: collecting real-time data during the operation of the power system, encrypting and compressing the data, and then transmitting it; receiving the data at a data processing center, decompressing, decrypting, and cleaning the data; performing feature learning on current and voltage data using a convolutional neural network model to extract key feature data reflecting the operating status of the power system; clustering equipment temperature and mechanical vibration data using a clustering algorithm to identify abnormal feature data of the equipment operating status; performing fault diagnosis and using LSTM to predict the operating data of the power system for a period of time in the future; issuing a warning message when abnormal data is detected or a potential fault is predicted, and identifying the fault location and type. The method of the present invention can achieve comprehensive and accurate monitoring of the operating status of the power system, improving the safety and reliability of the power system operation.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring, and in particular to a method for remote monitoring and control of power system data. Background Art

[0002] As power systems continue to expand in size and become increasingly complex, ensuring their safe and stable operation has become a crucial task. During power system operation, data such as current, voltage, equipment temperature, and mechanical vibration contain a wealth of operational information. Effectively collecting, processing, and analyzing this data can promptly identify potential faults and prevent accidents. Furthermore, effectively processing data from power equipment operation can enhance circuit safety.

[0003] On the one hand, traditional single-source encryption methods struggle to ensure the security and integrity of power equipment operating data in the face of increasingly complex cyberattacks. Power equipment operating data contains critical parameters; if leaked or tampered with, it could cause power system failures and even major safety incidents. Furthermore, traditional data compression technologies are unable to effectively process the massive amounts of power equipment operating data, increasing the burden on data transmission and reducing the efficiency of data storage and processing.

[0004] On the other hand, traditional power system data monitoring methods mostly rely on analyzing a single data type or employ simple threshold judgment methods. For example, simply monitoring whether the current or voltage amplitude exceeds a set threshold to determine whether the system is faulty. This approach fails to fully tap into the potential information in the data and can easily lead to missed or misjudged faults. Moreover, traditional methods often lack effective analytical tools for equipment temperature and vibration data, making it difficult to detect abnormal equipment conditions in advance. Furthermore, traditional methods lack sophisticated handling of noise and outliers during data processing, affecting the accuracy and reliability of data analysis.

[0005] Therefore, remote monitoring and control of power system data needs further research to improve the safety and reliability of power equipment operation. Summary of the Invention

[0006] The problem to be solved by the present invention is to provide a method for remote monitoring and control of power system data, which comprehensively and accurately monitors the operating status of the power system by real-time collection of multi-source data, combining data processing and intelligent algorithms, timely discovers and predicts faults, and improves the safety and reliability of power system operation.

[0007] To solve the above technical problems, the present invention provides a technical solution: a method for remote monitoring and control of power system data, comprising the following steps:

[0008] S1. Real-time data collection of current, voltage, equipment temperature, and mechanical vibration during power system operation. Encrypting the preliminarily processed data using a hybrid encryption algorithm. Compressing the encrypted data using a deep learning-based compression algorithm and transmitting it to the data processing center.

[0009] S2. The data processing center decompresses and decrypts the received data, and cleans the data to remove noise, outliers, and duplicate data.

[0010] S3. Construct a convolutional neural network model to perform feature learning on current and voltage waveform data, extracting key feature data reflecting the operating status of the power system; perform cluster analysis on equipment temperature and mechanical vibration data using a clustering algorithm to identify abnormal feature data of the equipment operating status;

[0011] S4. Use the SVM (Support Vector Machine) algorithm to diagnose the current operating status of the power system based on the correspondence between feature data and known fault types, and use the LSTM (long short-term memory) network to predict the operating data of the power system in the future.

[0012] S5. Set warning thresholds at different levels. When the detected data exceeds the corresponding warning threshold or a potential fault is predicted, a warning message is issued based on the severity and impact range of the fault, and the fault location and type are marked on the monitoring interface.

[0013] Preferably, in step S1, real-time data is collected by deploying different types of sensors at key nodes in the power system, wherein the sensors include: current sensors, voltage sensors, temperature sensors, and vibration sensors.

[0014] Preferably, the data after preliminary processing is encrypted using a hybrid encryption algorithm. The data is initially encrypted using a symmetric encryption algorithm, and then the symmetric encryption key is encrypted using an asymmetric encryption algorithm. The specific method is as follows:

[0015] S101. Key generation:

[0016] Symmetric encryption key generation: Randomly generate a length of symmetric encryption key ;

[0017] Asymmetric encryption key pair generation: Use the RSA algorithm to generate public and private keys.

[0018] S102. Symmetric encryption of data:

[0019] The data collected and initially processed are , using AES algorithm to encode data Encryption is performed; the symmetric encryption function is expressed as , the encrypted data for: .

[0020] S103. Asymmetric encryption of symmetric encryption keys:

[0021] Public key using asymmetric encryption algorithm Symmetric encryption key Encryption; the asymmetric encryption function is expressed as , then the encrypted symmetric encryption key for: .

[0022] S104, final encrypted data:

[0023] The encrypted data ultimately used for transmission The encrypted data and the encrypted symmetric encryption key Composition, expressed as: .

[0024] Preferably, the encrypted data is compressed using a deep learning-based compression algorithm to remove redundant information in the data. The specific method is as follows:

[0025] S111, encrypted data Preprocess it and convert it into a format suitable for deep learning model input, that is, for the encrypted byte sequence, convert it into a numerical vector .

[0026] S112. Build a deep learning model: Use an autoencoder as a deep learning model to perform data compression; the autoencoder consists of two parts: an encoder and a decoder;

[0027] The encoder takes the input data Mapping to a low-dimensional latent space , realizing data compression; the encoder is composed of multiple neural network layers, with a total of layer, calculate the Output of the layer , outputs a vector in the latent space: ;

[0028] The decoder transforms the vector in the latent space Reconstructed to the same input data Similar output ; The decoder also consists of multiple neural network layers, with a total of layer, calculate the Output of the layer , output the reconstructed data: .

[0029] S113, Model training: Use several power equipment operation data for training. The training goal is to minimize the reconstruction error. Use MSE as the loss function and use stochastic gradient descent to update the weights and biases of the encoder and decoder to minimize the loss function. .

[0030] S114, Data Compression: After the model training is completed, use the trained encoder to encrypt the data Compress and obtain the vector in the latent space .

[0031] Preferably, in step S2, the data processing center decompresses and decrypts the received data, and the specific method is as follows:

[0032] S201, data decompression:

[0033] At the receiving end, the trained decoder is used to decode the compressed data. Decompress and get the reconstructed data , the reconstructed data Perform inverse preprocessing to obtain decrypted data.

[0034] S202. Decryption of the symmetric encryption key:

[0035] The recipient uses the private key of the asymmetric encryption algorithm The encrypted symmetric encryption key Decryption; the asymmetric decryption function is expressed as , then the decrypted symmetric encryption key for: .

[0036] S203, data decryption:

[0037] The receiver uses the decrypted symmetric encryption key For encrypted data Decryption; the symmetric decryption function is expressed as , the decrypted data is: .

[0038] Preferably, in step S3, key features reflecting the operating status of the power system are extracted, and the specific method is as follows:

[0039] S301 , cutting the continuously collected current and voltage data into fixed time window time series segments of length T, converting them into a two-dimensional matrix, and performing normalization processing.

[0040] S302. Construct a convolutional neural network model:

[0041] The input layer includes: the first convolutional layer and the second convolutional layer. The activation function of both convolutional layers is ReLU. The first convolutional layer uses a small convolution kernel with a step size of 1, and the second convolutional layer uses a large convolution kernel.

[0042] Max pooling is used to reduce the dimensionality of the convolutional layer output, with a pooling window size of 2×1 and a stride of 2;

[0043] The pooled feature vector is flattened into a one-dimensional vector and input into the fully connected layer.

[0044] S303, Feature Learning:

[0045] The convolution kernel is used to slide over the time series segments and extract local patterns through a weight sharing mechanism. Small convolution kernels are used to capture high-frequency noise and transient disturbances, while large convolution kernels are used to capture low-frequency trends and steady-state features.

[0046] Shallow convolutional layers are used to learn basic features, including the waveform slope and harmonic amplitude within a single cycle. Convolutional layers are used through cross-layer connections to learn high-order features, including the distortion pattern of multi-cycle waveforms and the phase difference between current and voltage. Branches with different convolution kernel sizes are built in parallel to extract features at different time scales.

[0047] S304, Feature Screening and Fusion:

[0048] L1 regularization is used to calculate feature importance scores, screening out features that contribute most to fault diagnosis or cluster analysis. The feature maps of current and voltage are fused through channel splicing to generate composite features that include electrical quantity coupling relationships, thereby obtaining key feature vectors that reflect the operating status of the power system.

[0049] The key feature vectors reflect the degree of waveform distortion, the type of transient disturbance, the steady-state operating parameters and the timing characteristics of the abnormal mode.

[0050] Preferably, in step S301, the continuously collected current and voltage waveform data are converted into a two-dimensional matrix and normalized. The specific method is as follows:

[0051] S3011, the original time domain current signal , voltage signal Perform bandpass filtering:

[0052] ;

[0053] ;

[0054] in, is the impulse response of the bandpass filter.

[0055] S3012, the filtered current signal , voltage signal Perform sliding window splitting:

[0056] The filtered continuous real-time current and voltage data are cut into fixed time window time series segments of length T at intervals of one second. Each segment corresponds to a dual-channel input sample. ;

[0057] The window length is N, the step length is S, and the current window and voltage window signals of the kth window are:

[0058] ;

[0059] .

[0060] S3013. Extract time-frequency features of the current window signal and the voltage window signal:

[0061] ;

[0062] ;

[0063] in, For window boxes, is the time index, is the frequency index, It is the exponential form of Fourier transform, which represents the value of a complex sinusoidal signal with a frequency of f at discrete time n.

[0064] S3014. Calculate the spectrum amplitude matrix:

[0065] ;

[0066] .

[0067] S3015: Normalize the spectrum amplitude matrix.

[0068] S3016, Constructing a multi-channel matrix: Combining the two-dimensional matrix of current and voltage into a two-channel matrix .

[0069] Preferably, in step S3, the abnormal mode of the equipment operation state is identified, and the specific method is as follows:

[0070] Normalize temperature and vibration data to eliminate dimension effects;

[0071] Combine the temperature and vibration data for each device into a two-dimensional feature vector: ;in, For devices The temperature value, For devices Vibration value;

[0072] Set the number of clusters , randomly selected Initial cluster centers: .

[0073] Calculate the Euclidean distance of each data point to each cluster center and assign it to the cluster with the closest distance: ;in, For data points The cluster label, For the The first iteration cluster centers;

[0074] Recalculate the cluster center based on the current data points in the cluster: ;in, For the A set of data points in a cluster, is the number of data points in this cluster;

[0075] When the cluster center no longer changes significantly, , The iteration stops when the preset threshold is reached or the maximum number of iterations is reached.

[0076] After the iteration is completed, cluster labels are annotated based on business experience or historical data: if the temperature mean of a cluster is significantly higher than the normal threshold or the vibration amplitude exceeds the equipment operation standard, it is determined to be an abnormal mode cluster, and the other cluster is a normal mode cluster.

[0077] For the newly collected feature vector, calculate the cluster to which it belongs: ;like The value is the abnormal cluster label, which determines that the equipment operation status is abnormal and records the abnormal feature data collected when the equipment operation status is abnormal.

[0078] S317. The key feature data and the abnormal feature data are integrated into unified feature data to obtain a unified feature vector:

[0079] Based on key feature data and abnormal data Scale differences, key feature data Perform secondary normalization and feature concatenation to obtain integrated features:

[0080] ;

[0081] in, represents the dimension of Euclidean space, is the numerical code of the abnormal label. When normal, , abnormal, .

[0082] Preferably, in step S4, fault diagnosis is performed on the current operating state of the power system, and the specific method is as follows:

[0083] S401, dividing the feature data obtained in step S3 into a training set and a test set in a ratio of 7:3;

[0084] S402, constructing a nonlinear SVM model using a radial basis kernel function, setting the value ranges of hyperparameters C and γ, using 5-fold cross validation, using the test set accuracy as the evaluation indicator, and determining the optimal hyperparameters C and γ through grid search;

[0085] S403: Input the training set into the SVM model with determined hyperparameters for training. After the training is completed, input the test set into the trained SVM model to obtain a predicted fault type label for each test sample.

[0086] S404. For each test sample, the predicted fault type label is compared with the actual fault type label. If they are consistent, the sample is diagnosed correctly. If they are inconsistent, the sample is diagnosed incorrectly. The number of samples in the test set that are diagnosed correctly is counted, and the test set accuracy is calculated:

[0087] S405. Input the feature data collected and processed in real time into the trained SVM model. The model outputs the corresponding fault type label. If the output fault type label is one of the known fault types, it is determined that the power system currently has this type of fault; if the output fault type label is a normal operation label, it is determined that the power system is currently operating normally.

[0088] Preferably, in step S4, LSTM is used to predict the operation data of the power system in the future, and the specific method is as follows:

[0089] S411. The historical operation data of the power system is organized into a one-dimensional time series, and then the sliding window method is used to convert the series into input-output pairs according to the set window size and prediction step size; at the same time, the data is normalized;

[0090] S412. Build an LSTM network. The input layer dimension is set according to the data characteristics. The hidden layer consists of multiple layers of LSTM units. The output layer is a fully connected layer. The number of neurons corresponds to the prediction step size. The long-term dependencies of the data are captured through the forget gate, input gate, output gate, and cell state update mechanism.

[0091] S413, using mean square error as the loss function and Adam as the optimizer, the model is trained with training data, and the model parameters are adjusted through the back propagation algorithm;

[0092] S414: Input the real-time collected data into the trained model for prediction, and denormalize the prediction results to restore them to the original data scale.

[0093] Preferably, in step S5, the data processing center sends warning information to relevant personnel through one or more of text messages, emails, and sound and light alarms.

[0094] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0095] 1. The present invention collects multi-source data such as current, voltage, equipment temperature, mechanical vibration, and performs comprehensive processing and analysis to comprehensively and accurately reflect the operating status of the power system. Compared with traditional analysis methods based on a single data type, it greatly improves the accuracy and reliability of monitoring.

[0096] 2. The method of the present invention uses convolutional neural networks, clustering algorithms, SVM algorithms and LSTM to deeply mine and analyze data. The convolutional neural network can effectively extract the key features of current and voltage waveforms, the clustering algorithm can accurately identify abnormal modes of equipment operation status, the SVM algorithm can realize accurate fault diagnosis, and the LSTM can reliably predict the future operation data of the power system, thereby improving the accuracy of fault diagnosis and prediction.

[0097] 3. By setting the early warning threshold, the present invention can issue early warning information in time before or when a fault occurs, and clearly mark the fault location and type on the monitoring interface, so that operation and maintenance personnel can quickly locate the fault and take corresponding measures, reduce the impact of the fault on the operation of the power system, and improve the safety and stability of the power system.

[0098] 4. The method of the present invention can automatically complete tasks such as data collection, processing, analysis, diagnosis and prediction, thereby reducing manual intervention, improving operation and maintenance efficiency, and reducing operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 This is a flow chart of the method for remote monitoring and control of power system data of the present invention;

[0100] Figure 2 It is a block diagram of the power system data remote monitoring and control system of the present invention. DETAILED DESCRIPTION

[0101] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the application are further elaborated in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in this field on this embodiment fall within the scope of protection of the present invention. At the same time, the step numbers in the embodiments of the present invention are only set for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0102] In one embodiment of the present invention, a method for remote monitoring and control of power system data is provided. Figure 1 As shown, the following steps are included:

[0103] Step 1: Real-time collection and transmission of multi-source data

[0104] First, deploy current sensors, voltage sensors, temperature sensors, and vibration sensors at key nodes in the power system (such as transformers, circuit breakers, and transmission line interfaces).

[0105] Current monitoring: A non-invasive current sensor based on the Rogowski coil principle is used to achieve contactless measurement of AC current, avoiding damage to the original circuit structure.

[0106] Voltage monitoring: The voltage sensor based on capacitive voltage division technology obtains high-precision voltage signals in real time and is suitable for high-voltage and low-voltage scenarios.

[0107] Temperature monitoring: A Class A precision Pt100 thermal resistor temperature sensor is selected with a temperature measurement range of -200°C~850°C to meet the temperature rise monitoring needs of power equipment.

[0108] Vibration monitoring: Equipped with a three-axis accelerometer vibration sensor with a sampling frequency of up to 10kHz, it can capture the time domain and frequency domain characteristics of the equipment's mechanical vibration.

[0109] Then, each sensor collects data in real time at a synchronous sampling frequency of 1000 Hz, uses a hybrid encryption algorithm to encrypt the collected and preliminarily processed data, and uses a deep learning-based compression algorithm to compress the encrypted data.

[0110] Specifically, in this embodiment, the hybrid encryption algorithm first uses a symmetric encryption algorithm to preliminarily encrypt data, and then uses an asymmetric encryption algorithm to encrypt the symmetric encryption key.

[0111] The collected and preliminarily processed data is encrypted. The specific method is as follows:

[0112] S101. Key generation:

[0113] Symmetric encryption key generation: Randomly generate a length of symmetric encryption key ;

[0114] Asymmetric encryption key pair generation: Use the RSA algorithm to generate public and private keys.

[0115] S102. Symmetric encryption of data:

[0116] The data collected and initially processed are , using AES algorithm to encode data Encryption is performed; the symmetric encryption function is expressed as , the encrypted data for: .

[0117] S103. Asymmetric encryption of symmetric encryption keys:

[0118] Public key using asymmetric encryption algorithm Symmetric encryption key Encryption; the asymmetric encryption function is expressed as , then the encrypted symmetric encryption key for: .

[0119] S104, final encrypted data:

[0120] The encrypted data ultimately used for transmission The encrypted data and the encrypted symmetric encryption key Composition, expressed as: .

[0121] Specifically, in this embodiment, a compression algorithm based on deep learning is used to compress the encrypted data to remove redundant information in the data. The specific method is as follows:

[0122] S111, encrypted data Preprocess it and convert it into a format suitable for deep learning model input, that is, for the encrypted byte sequence, convert it into a numerical vector ;

[0123] in, is a preprocessing function that converts a byte sequence into a numeric vector.

[0124] S112. Build a deep learning model:

[0125] An autoencoder is used as a deep learning model for data compression. The autoencoder consists of two parts: an encoder and a decoder.

[0126] The encoder takes the input data Mapping to a low-dimensional latent space , realizing data compression; the encoder is composed of multiple neural network layers, with a total of Layer, Output of the layer The calculation method is as follows:

[0127] ;

[0128] in, It is The weight matrix of the layer, It is The bias vector of the layer, is the activation function;

[0129] Finally, the encoder output is a vector in the latent space: .

[0130] The decoder transforms the vector in the latent space Reconstructed to the same input data Similar output ; The decoder also consists of multiple neural network layers, with a total of Layer, Output of the layer The calculation method is as follows:

[0131] ;

[0132] in, It is The weight matrix of the layer, It is The bias vector of the layer, is the activation function;

[0133] Finally, the decoder output Here is the reconstructed data: .

[0134] S113, Model Training:

[0135] The training is performed using several power equipment operation data. The goal of the training is to minimize the reconstruction error and use MSE as the loss function:

[0136] ;

[0137] in, is the amount of training data, It is Input data, It is Reconstructed data;

[0138] Use stochastic gradient descent to update the weights and biases of the encoder and decoder to minimize the loss function .

[0139] S114, Data Compression:

[0140] After the model training is completed, use the trained encoder to encrypt the data Compress and obtain the vector in the latent space :

[0141] ;

[0142] in, is the trained encoder.

[0143] Then, the data is transmitted to the data processing center via the 5G communication satellite network or the LoRaWAN wireless communication protocol, ensuring low power consumption and stability of data transmission.

[0144] Specifically, in this embodiment, when the 5G signal in the area where the power equipment is located is good, the compressed and encrypted data will be preferentially transmitted to the data processing center through the 5G communication network; if the 5G signal is poor or interrupted, it will automatically switch to satellite communication to ensure the continuity of data transmission; during the data transmission process, an adaptive transmission rate adjustment strategy is adopted to dynamically adjust the data transmission rate according to the network bandwidth and signal quality to ensure stable data transmission.

[0145] Step 2: Data cleaning and preprocessing

[0146] The data processing center decrypts and decompresses the received data, and cleans the data to remove noise, outliers and duplicate data.

[0147] Among them, noise filtering can adopt the adaptive Wiener filtering algorithm, outlier detection and elimination can use the interquartile range (IQR) method, and duplicate data removal can be achieved through the hash value comparison algorithm to calculate the hash value of each data record. If the hash value is repeated, the earliest collected record will be retained to ensure data uniqueness.

[0148] Specifically, in this embodiment, the data is decompressed and decrypted to restore the original data. The specific method is as follows:

[0149] S201, data decompression:

[0150] At the receiving end, the trained decoder is used to decode the compressed data. Decompress and get the reconstructed data: ;in, is the trained decoder;

[0151] The reconstructed data Perform inverse preprocessing to obtain the decrypted data: ;

[0152] in, is the inverse preprocessing function, which converts a numeric vector into a byte sequence.

[0153] S202. Decryption of the symmetric encryption key:

[0154] The recipient uses the private key of the asymmetric encryption algorithm The encrypted symmetric encryption key Decryption; the asymmetric decryption function is expressed as , then the decrypted symmetric encryption key for: .

[0155] S203, data decryption:

[0156] The receiver uses the decrypted symmetric encryption key For encrypted data Decryption; the symmetric decryption function is expressed as , the decrypted data is: .

[0157] Step 3: Feature extraction and pattern recognition

[0158] First, extract the current and voltage waveform features: the continuously collected current and voltage waveform data are cut into fixed time window time series segments of length T at intervals of one second. Each segment corresponds to a dual-channel input to form a dual-channel input sample. , and normalize each time series segment.

[0159] Current and voltage waveform feature extraction: the original time domain current signal , voltage signal Perform bandpass filtering:

[0160] ;

[0161] ;

[0162] in, is the impulse response of the bandpass filter;

[0163] The filtered current signal and voltage signal are segmented by sliding windows: the window length is N, the step length is S, and the current window signal of the kth window is: ; Similarly, the voltage window signal of the kth window is obtained .

[0164] Extract time-frequency features of current window signal and voltage window signal:

[0165] ;

[0166] ;

[0167] in: is a window function (such as Hanning window), is the time index, is the frequency index.

[0168] Compute the spectral magnitude matrix:

[0169] ;

[0170] .

[0171] Matrix normalization:

[0172] ;

[0173] .

[0174] Construct a multi-channel matrix: Combine the two-dimensional matrices of current and voltage into a two-channel matrix:

[0175] .

[0176] Then, a convolutional neural network (CNN) model is constructed, whose input layer dimension is The multi-layer convolutional layer includes the first convolutional layer with a small 3×1 convolution kernel and a stride of 1, and the second convolutional layer with a large 7×1 convolution kernel. The activation function of both convolutional layers is ReLU. Maximum pooling is used to reduce the dimensionality of the convolutional layer output, with a pooling window size of 2×1 and a stride of 2. The pooled feature vector is flattened into a one-dimensional vector, and the features are further fused through a fully connected layer. The extracted features are filtered through L1 regularization, and the feature maps of current and voltage are fused through channel splicing.

[0177] Next, the convolution kernel is used to slide over the time series segments, and the local pattern is extracted with the help of the weight sharing mechanism. Among them, the small convolution kernel captures high-frequency noise and transient disturbances, while the large convolution kernel captures low-frequency trends and steady-state features.

[0178] Shallow convolutional layer: learns basic features, including waveform slope and harmonic amplitude within a single cycle;

[0179] Deep convolutional layers: Through cross-layer connections, they learn high-level features, including the distortion pattern of multi-cycle waveforms and the phase difference between current and voltage;

[0180] Build branches with different convolution kernel sizes in parallel to extract features at different time scales, and then fuse them by splicing;

[0181] L1 regularization is used to calculate the feature importance score and screen out features that have relatively high contribution to subsequent fault diagnosis or cluster analysis.

[0182] In particular, for the dual-channel data of current and voltage, the feature maps of the two are fused through channel splicing after the convolution layer to generate a composite feature containing the coupling relationship of electrical quantities; finally, the key feature vector that can reflect the operating status of the power system is obtained. This key feature vector embodies the degree of waveform distortion, transient disturbance type, steady-state operating parameters and the timing characteristics of abnormal mode.

[0183] Next, cluster analysis is performed on the equipment temperature and vibration data: the temperature data and vibration data are normalized to eliminate the dimension effect:

[0184] ;

[0185] in, is the original data, is the normalized data;

[0186] Combine the temperature and vibration data for each device into a two-dimensional feature vector:

[0187] ;

[0188] in, For equipment The temperature value, For equipment Vibration value.

[0189] In this embodiment, the number of clusters is set , randomly selected Initial cluster centers: ;

[0190] Calculate the Euclidean distance of each data point to each cluster center and assign it to the cluster with the closest distance:

[0191] ;

[0192] in, For data points The cluster label, For the The first iteration cluster centers;

[0193] Recalculate the cluster center based on the current data points in the cluster: ;in, For the A set of data points in a cluster, is the number of data points in this cluster;

[0194] When the cluster center no longer changes significantly, , The iteration stops when the preset threshold is reached or the maximum number of iterations is reached.

[0195] Finally, after the iteration is complete, cluster labels are added based on business experience or historical data: if the temperature mean of a cluster is significantly higher than the normal threshold or the vibration amplitude exceeds the equipment operation standard, it is determined to be an abnormal mode cluster; the other cluster is a normal mode cluster;

[0196] For the newly collected feature vector, calculate the cluster to which it belongs: ;like If the value is equal to the abnormal cluster label, the device operation status is determined to be abnormal, and the abnormal feature data collected when the device operation status is abnormal is recorded.

[0197] The key feature data and the abnormal feature data are integrated into a unified feature data, that is, a unified feature vector:

[0198] (1) If the scale difference between key feature data and abnormal integral data is significant, the key feature data is normalized twice;

[0199] (2) Perform feature splicing to obtain integrated features: ;in, is the numerical code of the abnormal label. When normal, , abnormal, .

[0200] Step 4: Fault diagnosis and prediction

[0201] First, perform fault diagnosis as follows:

[0202] The feature data obtained in step 3 are divided into training set and test set in a ratio of 7:3. The nonlinear SVM model is constructed using the radial basis kernel function, and the hyperparameters C and γ are set.

[0203] In this embodiment, the value range of C is [0.1, 100], the value range of γ is [0.001, 10], and 5-fold cross validation is used. The test set accuracy is used as the evaluation indicator, and the optimal hyperparameters C and γ are determined through grid search.

[0204] The training set is input into the SVM model with determined hyperparameters for training. After the training is completed, the test set is input into the trained SVM model to obtain the predicted fault type label of each test sample.

[0205] For each test sample, the predicted fault type label is compared with the actual fault type label. If they are consistent, the sample is diagnosed correctly. If they are inconsistent, the sample is diagnosed incorrectly. The number of samples diagnosed correctly in the test set is counted to calculate the test set accuracy:

[0206] Test set accuracy = (number of samples diagnosed correctly / total number of test set samples) × 100%.

[0207] The feature data collected and processed in real time is input into the trained SVM model, and the model outputs the corresponding fault type label. If the output fault type label is one of the known fault types, it is determined that the power system currently has this type of fault; if the output fault type label is a normal operation label, it is determined that the power system is currently operating normally.

[0208] Then, run the data prediction as follows:

[0209] First, the historical operation data of the power system is organized into a one-dimensional time series. Then, the sliding window method is used to convert the series into input-output pairs according to the set window size and prediction step size. At the same time, the data is normalized.

[0210] Build an LSTM network. The input layer dimension is set according to the data characteristics. The hidden layer consists of multiple layers of LSTM units. The output layer is a fully connected layer. The number of neurons corresponds to the prediction step size. The long-term dependency of the data is captured through the forget gate, input gate, output gate and cell state update mechanism.

[0211] Using mean square error as the loss function and Adam as the optimizer, the model is trained with training data, and the model parameters are adjusted through the back propagation algorithm; the real-time collected data is input into the trained model for prediction, and the prediction results are denormalized and restored to the original data scale.

[0212] Step 5: Warning processing

[0213] Set different levels of warning thresholds. When the detected data exceeds the corresponding threshold or a potential fault is predicted, warning information will be sent to relevant personnel through one or more methods such as SMS, email, and sound and light alarms based on the severity and scope of the fault. The fault location and type will be marked with different colors and icons on the monitoring interface so that staff can take timely measures.

[0214] For example, set the following warning levels:

[0215] Level 1 warning (serious fault): Triggered by SVM diagnosis of a short circuit / overload fault, or temperature > 100°C and vibration > 8m / s². Notifications are sent via audio and visual alarm (flashing red + buzzer), SMS (response within 10 seconds), and email.

[0216] Level 2 warning (abnormal status): The trigger condition is that the cluster analysis determines that the pattern cluster is abnormal, or the LSTM prediction error is greater than 5%. The system will display a yellow mark on the monitoring interface and send an email notification.

[0217] Furthermore, the method of the present invention is used to remotely monitor and control data of a 10kV transmission line in a certain substation.

[0218] 1. Data Collection

[0219] The following sensors are deployed at transmission line nodes (sampling frequency 1000 Hz, continuously collecting data for 1 second):

[0220] Current sensor: Rogowski coil, measured current waveform data (partial sampling points):

[0221] (effective value is about 50A);

[0222] Voltage sensor: Capacitor voltage division, measured voltage waveform data:

[0223] (RMS value is about 10kV);

[0224] Temperature sensor: Pt100, measured temperature: 35°C;

[0225] Vibration sensor: triaxial accelerometer, average vibration amplitude: 2.5m / s²;

[0226] Data transmission: 1,000 sets of current / voltage data, temperature, and vibration values are transmitted to the data processing center within 1 second via the LoRaWAN protocol.

[0227] 2. Data Cleaning and Preprocessing

[0228] 1. Current / voltage waveform cleaning

[0229] Noise filtering: Adaptive Wiener filtering is applied to the current waveform to remove high-frequency noise (a certain abnormal point was corrected from 52.1A to 51.0A).

[0230] Outlier detection: Using the interquartile range (IQR) method, the lower quartile of the current data is set to , the upper quartile is The interquartile range is then: ;

[0231] Calculate the current data .

[0232] The outlier threshold range is: .

[0233] Remove outliers (if any) outside this range.

[0234] Deduplication: Through hash value comparison, there are no duplicate records.

[0235] 2. Temperature / vibration data normalization

[0236] Assume the original data is , the normalized data is , then the normalization formula is: ;

[0237] Original temperature value: 35℃, historical temperature range is , after normalization: ;

[0238] Vibration original value: 2.5m / s 2 , the historical vibration range is , after normalization: ;

[0239] Combined into a two-dimensional feature vector: .

[0240] 3. Feature Extraction and Pattern Recognition

[0241] 1. Current / voltage waveform feature extraction (CNN model)

[0242] Time window: intercept 1 second of data as 1 time series segment, the segment length is , dual channel input (current + voltage), forming .

[0243] Normalization: Normalize the current / voltage value to interval.

[0244] Convolutional layer calculation:

[0245] The first convolution layer uses a 3×1 convolution kernel with a step size of 1 and an output channel number of , the output dimension is:

[0246] ;

[0247] Used to extract high-frequency transient features (such as harmonics).

[0248] The second convolution layer uses a 7×1 convolution kernel with a step size of 1 and an output channel number of , the output dimension is:

[0249] ;

[0250] Used to extract low-frequency steady-state features (such as fundamental wave trends).

[0251] Pooling layer: max pooling (window 2×1, stride 2), dimension after dimensionality reduction: .

[0252] Feature fusion: After flattening to a one-dimensional vector, a fully connected layer is used to generate key feature vectors (including 50-dimensional features such as waveform distortion rate and phase difference).

[0253] 2. Temperature / vibration cluster analysis

[0254] The cluster center is obtained by historical data training:

[0255] Normal cluster center: (corresponding to temperature 30°C, vibration 2m / s 2 );

[0256] Abnormal cluster center: (corresponding to temperature 50°C, vibration 4m / s 2 );

[0257] Calculate the Euclidean distance for the current data, assuming the data point is , the cluster center is , the Euclidean distance formula is:

[0258] ;

[0259] in is the feature vector dimension, where .

[0260] Calculate the current data point arrive The Euclidean distance of:

[0261] ;

[0262] Calculated to The Euclidean distance of:

[0263] ;

[0264] Assigned to the nearest cluster: Normal cluster (label 1), the device status is determined to be normal.

[0265] 4. Fault Diagnosis and Prediction

[0266] 1. Fault diagnosis (SVM model)

[0267] Trained SVM model parameters:

[0268] Kernel function: radial basis kernel (RBF), optimal hyperparameters: C=10, γ=0.1;

[0269] Input features: current / voltage CNN feature vector (50 dimensions).

[0270] The current feature vector is input into the SVM model, and the output label is normal operation (no fault type matching), which means that the system is operating normally.

[0271] 2. Run data prediction (LSTM model)

[0272] Historical data: Current effective value series for the past hour (unit: A):

[0273] (60 points in total, 1 point per minute);

[0274] Sliding window: Set the window size to w=5 and the prediction step size to p=1 to generate input-output pairs (e.g. input [48,50,52,49,51] output 50).

[0275] Normalization: Data is normalized to [0,1].

[0276] LSTM model prediction: input current 5 points of data , the predicted current value for the next minute is 50.5A (after inverse normalization).

[0277] Prediction error: Compared with the actual value (50.8A), the error calculation formula is:

[0278] The calculated error is: , no warning was triggered.

[0279] 5. Early Warning Processing

[0280] Current status: Temperature 35℃<100℃, Vibration 2.5m / s 2 <8m / s 2 , SVM diagnosis was normal, and LSTM prediction error was 0.59% < 5%.

[0281] No warning is triggered, and the monitoring interface displays a green icon (normal operation).

[0282] The embodiment of the present invention also provides a power system data remote monitoring and control system for implementing the above method, such as Figure 2 Shown, including:

[0283] Data acquisition module: This includes current sensors, voltage sensors, temperature sensors, and vibration sensors deployed at key nodes in the power system. It is used to collect multi-source data during the operation of the power system in real time and transmit the data to the data processing module through the wireless communication module.

[0284] Data processing module: Receives data transmitted by the data acquisition module, cleans, extracts features, and analyzes the data. This includes convolutional neural network feature extraction of current and voltage waveforms, and cluster analysis of equipment temperature and vibration data, providing processed data for subsequent fault diagnosis and prediction.

[0285] Fault diagnosis and prediction module: Uses the SVM algorithm to diagnose faults in the current operating status of the power system and uses LSTM to predict future operating data of the power system;

[0286] Early warning module: used to set early warning thresholds at different levels. When data anomalies are detected or potential faults are predicted, early warning information is issued to relevant personnel in various ways, and the fault location and type are intuitively displayed on the monitoring interface.

[0287] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A method for remote monitoring and control of power system data, characterized in that: The following steps are involved: S1. Real-time data collection of current, voltage, equipment temperature, and mechanical vibration during power system operation. Encrypting the preliminarily processed data using a hybrid encryption algorithm. Compressing the encrypted data using a deep learning-based compression algorithm and transmitting it to the data processing center. S2. The data processing center decompresses and decrypts the received data, and cleans the data to remove noise, outliers, and duplicate data. S3. Construct a convolutional neural network model to perform feature learning on current and voltage waveform data, extracting key feature data reflecting the operating status of the power system; perform cluster analysis on equipment temperature and mechanical vibration data using a clustering algorithm to identify abnormal feature data of the equipment operating status; S4. Use the SVM algorithm to diagnose the current operating status of the power system based on the correspondence between feature data and known fault types, and use LSTM to predict the operating data of the power system in the future; S5. Set warning thresholds at different levels. When the detected data exceeds the corresponding warning threshold or a potential fault is predicted, a warning message is issued based on the severity and impact range of the fault, and the fault location and type are marked on the monitoring interface.

2. The method for remote monitoring and control of power system data according to claim 1, characterized in that: In step S1, the data after preliminary processing is encrypted using a hybrid encryption algorithm. The specific method is as follows: S101. Key generation: Symmetric encryption key generation: Randomly generate a length of symmetric encryption key , expressed as: ;in, Represents the generation of a Random number of bits; Asymmetric encryption key pair generation: Use RSA algorithm to generate public key and private key of asymmetric encryption algorithm; S102. Symmetric encryption of data: The data collected and initially processed are , using AES algorithm to encode data Encryption is performed; the symmetric encryption function is expressed as , the encrypted data for: ; S103. Asymmetric encryption of symmetric encryption keys: Public key using asymmetric encryption algorithm Symmetric encryption key Encryption; the asymmetric encryption function is expressed as , then the encrypted symmetric encryption key for: ; S104, final encrypted data: The encrypted data ultimately used for transmission The encrypted data and the encrypted symmetric encryption key Composition, expressed as: .

3. The method for remote monitoring and control of power system data according to claim 2, characterized in that: In step S1, the encrypted data is compressed using a deep learning-based compression algorithm. The specific method is as follows: S111, the encrypted data Convert to a numerical vector suitable for deep learning model input : ； in, is a preprocessing function used to convert a byte sequence into a numerical vector; S112. Build a deep learning model: Use an autoencoder as a deep learning model for data compression, including an encoder and a decoder; The encoder takes the input data Vectors mapped to a low-dimensional latent space , realizing data compression; the decoder converts the vector in the latent space Reconstructed to the same input data Similar output ; S113. Model training: Use some power equipment operation data for training. The goal is to minimize the reconstruction error. Use MSE as the loss function: ； in, is the amount of training data, It is Input data, It is Reconstructed data; Use stochastic gradient descent to update the weights and biases of the encoder and decoder to minimize the loss function ; S114, Data Compression: Use the trained encoder to compress the encrypted data Compress and obtain the vector in the latent space : ； in, is the trained encoder.

4. The method for remote monitoring and control of power system data according to claim 3, characterized in that: In step S112, the encoder and decoder are both composed of multiple neural network layers; Encoder total Layer, Output of the layer The calculation method is as follows: ； The encoder outputs a vector in latent space: ; in, It is The weight matrix of the layer, It is The bias vector of the layer, is the activation function; Decoder total Layer, Output of the layer The calculation method is as follows: ； Decoder output Here is the reconstructed data: ; in, It is The weight matrix of the layer, It is The bias vector of the layer.

5. The method for remote monitoring and control of power system data according to claim 3, characterized in that: In step S2, the data processing center decompresses and decrypts the received data. The specific method is as follows: S201, data decompression: At the receiving end, the trained decoder is used to decode the compressed data. Decompress and get the reconstructed data: ； in, is the trained decoder; The reconstructed data Perform inverse preprocessing to obtain decrypted data : ； in, is the inverse preprocessing function, used to convert a numeric vector into a byte sequence; S202. Decryption of the symmetric encryption key: The recipient uses the private key of the asymmetric encryption algorithm The encrypted symmetric encryption key Decryption; the asymmetric decryption function is expressed as , then the decrypted symmetric encryption key for: ; S203, data decryption: The receiver uses the decrypted symmetric encryption key For encrypted data Decryption; the symmetric decryption function is expressed as , the decrypted data is: .

6. The method for remote monitoring and control of power system data according to claim 1, characterized in that: In step S3, key characteristic data reflecting the operating status of the power system is extracted. The specific method is as follows: S301, cutting the continuously collected current and voltage data into fixed time window time series segments of length T, converting them into a two-dimensional matrix, and performing normalization processing; S302. Construct a convolutional neural network model: The input layer includes: the first convolutional layer and the second convolutional layer. The activation function of both convolutional layers is ReLU. The first convolutional layer uses a small convolution kernel with a step size of 1, and the second convolutional layer uses a large convolution kernel. Max pooling is used to reduce the dimensionality of the convolutional layer output, with a pooling window size of 2×1 and a stride of 2; Flatten the pooled feature vector into a one-dimensional vector and input it into the fully connected layer; S303, Feature Learning: The convolution kernel is used to slide over the time series segments and extract local patterns through a weight sharing mechanism. Small convolution kernels are used to capture high-frequency noise and transient disturbances, while large convolution kernels are used to capture low-frequency trends and steady-state features. Shallow convolutional layers are used to learn basic features, including waveform slope and harmonic amplitude within a single cycle. Cross-layer connections are used to learn high-order features, including the distortion pattern of multi-cycle waveforms and the phase difference between current and voltage. Branches with different convolution kernel sizes are built in parallel to extract features at different time scales. S304, Feature Screening and Fusion: The feature importance score is calculated through L1 regularization to screen out features that contribute highly to fault diagnosis or cluster analysis. The feature maps of current and voltage are fused through channel splicing to generate composite features containing the coupling relationship of electrical quantities, and obtain the key feature vector reflecting the operating status of the power system.

7. The method for remote monitoring and control of power system data according to claim 6, characterized in that: In step S301, the continuously collected current and voltage waveform data are converted into a two-dimensional matrix and normalized. The specific method is as follows: S3011, the original time domain current signal , voltage signal Perform bandpass filtering: ； ； in, is the impulse response of the bandpass filter; S3012, the filtered current signal , voltage signal Perform sliding window splitting: The filtered continuous real-time current and voltage data are cut into fixed time window time series segments of length T at intervals of one second. Each segment corresponds to a dual-channel input sample. ; The window length is N, the step length is S, and the current window and voltage window signals of the kth window are: ； ； S3013. Extract time-frequency features of the current window signal and the voltage window signal: ； ； in, For window boxes, is the time index, is the frequency index, is the Fourier transform exponential form; S3014. Calculate the spectrum amplitude matrix: ； ； S3015, spectrum amplitude matrix normalization processing: ； ； S3016, Construct a multi-channel matrix: Combine the two-dimensional matrices of current and voltage into a two-channel matrix: 。 8. The method for remote monitoring and control of power system data according to claim 7, characterized in that: In step S3, the abnormal characteristic data of the equipment operation status is identified, and the specific method is as follows: S311. Normalize the temperature data and vibration data to eliminate the dimension effect; S312. Combine the temperature data and vibration data of each device into a two-dimensional feature vector: ； in, For equipment The temperature value, For equipment Vibration value; S313. Set the number of clusters , randomly selected Initial cluster centers: ; Calculate the Euclidean distance of each data point to each cluster center and assign it to the cluster with the closest distance: ； in, For data points The cluster label, For the The first iteration cluster centers; Recalculate the cluster center based on the current data points in the cluster: ； in, For the A set of data points in a cluster, is the number of data points in this cluster; S314: When the cluster center change is less than the preset threshold, , When the preset threshold or the maximum number of iterations is reached, the iteration is stopped; S315. Label the clusters based on business experience or historical data: if the temperature mean of a cluster is higher than the normal threshold or the vibration amplitude exceeds the equipment operation standard, it is determined to be an abnormal mode cluster; the other cluster is a normal mode cluster; S316: For the newly collected feature vector , calculate the cluster : ； in, For the cluster centers, if The value is an abnormal cluster label, which determines that the equipment operation status is abnormal and records the abnormal feature data collected when the equipment operation status is abnormal; S317. The key feature data and the abnormal feature data are integrated into unified feature data to obtain a unified feature vector: Based on key feature data and abnormal data Scale differences, key feature data Perform secondary normalization and feature concatenation to obtain integrated features: ； in, represents the dimension of Euclidean space, Numeric encoding for the anomaly label.

9. The method for remote monitoring and control of power system data according to claim 8, characterized in that: In step S4, a fault diagnosis is performed on the current operating state of the power system. The specific method is as follows: S401, dividing the feature data obtained in step S3 into a training set and a test set in a ratio of 7:3; S402. Build a nonlinear SVM model using a radial basis kernel function, set a hyperparameter value range, use 5-fold cross validation, use the test set accuracy as the evaluation indicator, and determine the optimal hyperparameters through grid search; S403: Input the training set into the SVM model with determined hyperparameters for training. After the training is completed, input the test set into the trained SVM model to obtain a predicted fault type label for each test sample. S404: For each test sample, the predicted fault type label is compared with the actual fault type label. If they are consistent, the sample diagnosis is correct; if they are inconsistent, the sample diagnosis is incorrect. Count the number of samples diagnosed correctly in the test set and calculate the accuracy of the test set: S405: Input the processed feature data into the trained SVM model and output the corresponding fault type label.

10. The method for remote monitoring and control of power system data according to claim 1, characterized in that: In step S4, LSTM is used to predict the operation data of the power system in the future. The specific method is as follows: S411. Treat the historical operation data of the power system as a one-dimensional time series, convert the series into input-output pairs according to the set window size and prediction step size through a sliding window method, and normalize the data; S412. Build an LSTM network. The input layer dimension is set according to the data characteristics. The hidden layer consists of multiple layers of LSTM units. The output layer is a fully connected layer. The number of neurons corresponds to the prediction step size. The long-term dependencies of the data are captured through the forget gate, input gate, output gate, and cell state update mechanism. S413, using mean square error as the loss function and Adam as the optimizer, the model is trained with training data, and the model parameters are adjusted through the back propagation algorithm; S414: Input the real-time collected data into the trained model for prediction, and denormalize the prediction results to restore them to the original data scale.

Images