Substation monitoring data anomaly detection method and system based on deep learning

Abstract

The invention discloses a substation monitoring data anomaly detection method and system based on deep learning, and mainly relates to the technical field of power data processing and monitoring. Comprising the following steps: carrying out preprocessing and feature extraction on original data collected in a transformer substation monitoring system; after feature extraction is completed, the extracted features are analyzed and processed by adopting a deep learning algorithm, and abnormal points or abnormal modes in the data are identified; performing hyper-parameter optimization on the constructed deep learning algorithm model; and outputting an anomaly detection result in a visual mode. The beneficial effects of the invention are that the method can achieve the quick and accurate anomaly detection of the monitoring data of the transformer substation, and provides a powerful guarantee for the safe and stable operation of the transformer substation.

Description

Technical Field

[0001] The present invention relates to the field of power data processing and monitoring technology, and specifically to a substation monitoring data anomaly detection method and system based on deep learning. Background Art

[0002] With the rapid development of smart grids, substations are the core components of power systems, and their safe and stable operation is crucial to ensuring the reliability and stability of power supply. Substation monitoring data, as an important information source reflecting the operating status of substations, contains a large amount of real-time and historical data. These data contain rich operating status information, and may also hide various abnormal or fault signals. Therefore, how to efficiently and accurately extract useful information from massive monitoring data and timely discover and handle potential anomalies has become an important challenge facing the current substation operation and maintenance management. Traditional substation monitoring data anomaly detection methods mostly rely on manually set thresholds or simple analysis methods based on statistics. These methods have many limitations when processing complex and changeable power system data, such as strong subjectivity in threshold setting, difficulty in adapting to system dynamic changes, and insufficient ability to identify potential abnormal patterns. In addition, with the rapid growth of substation monitoring data, traditional methods are also under tremendous pressure in terms of processing speed and accuracy.

[0003] In recent years, the rapid development of deep learning technology has provided new ideas for anomaly detection of substation monitoring data. With its powerful feature extraction and pattern recognition capabilities, deep learning has achieved remarkable results in image recognition, natural language processing and other fields. Applying deep learning to anomaly detection of substation monitoring data can automatically learn complex feature representations from data, effectively capture abnormal patterns in data, and improve the accuracy and efficiency of anomaly detection. However, there are still some challenges in directly applying existing deep learning models to anomaly detection of substation monitoring data. For example, substation monitoring data has time series characteristics and there are complex dependencies between data. Although traditional recurrent neural networks (RNNs) and their variants such as long short-term memory networks (LSTMs) and gated recurrent units (GRUs) can process time series data, they may face problems such as long training time and high consumption of computing resources when processing large-scale data sets. In addition, the characteristics of monitoring data from different substations are different, and the models need to be customized and optimized for specific application scenarios.

[0004] Therefore, there is an urgent need for a substation monitoring data anomaly detection method and system based on deep learning to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide a substation monitoring data anomaly detection method and system based on deep learning, which can realize fast and accurate anomaly detection of substation monitoring data and provide strong guarantee for the safe and stable operation of the substation.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0007] On the one hand, a method for detecting abnormality in substation monitoring data based on deep learning is provided, comprising the following steps:

[0008] S1: Preprocess and extract features of the raw data collected in the substation monitoring system;

[0009] S2: After feature extraction is completed, a deep learning algorithm is used to analyze and process the extracted features to identify abnormal points or abnormal patterns in the data;

[0010] S3: Optimize the hyperparameters of the built deep learning algorithm model;

[0011] S4: Output the anomaly detection results in a visual manner.

[0012] Preferably, the step S1 is specifically:

[0013] The preprocessing of the raw data includes data cleaning and data transformation, which is used to clean and transform the data, remove noise, process missing values, and convert data types;

[0014] The feature extraction of the original data is to extract key features useful for anomaly detection from the preprocessed data.

[0015] Preferably, the extracting of key features useful for anomaly detection specifically includes:

[0016] S11: Set a threshold τ1 to filter out features whose variance is less than τ1 or whose absolute value of correlation coefficient with the target variable is less than τ1. The obtained feature set is recorded as F1.

[0017] S12: Train a model on F1 that supports feature importance evaluation, denoted as M, and M outputs the importance score of each feature, denoted as importance i , where i represents the feature index;

[0018] S13: For each pair of features in F1 (f i , f j ), calculate its correlation and get the correlation matrix C ij , where C uj Represents the feature f i and f j The correlation between

[0019] S14: Calculate the weighted score S of each feature based on the importance score of the feature and the correlation between the features i :

[0020]

[0021] Among them, |F1| represents the number of features in F1;

[0022] S15: Set a threshold τ2 and select a weighted score S i Features greater than τ2 are used as the final feature subset F find .

[0023] Preferably, the step S2 specifically includes the following process:

[0024] According to the deep learning algorithm, the activation function in the neural network is calculated to determine the activation mode and strength of the neuron:

[0025] z t =σ(W z ·[h t-1 , x t ])

[0026] r t =σ(W r ·[h t-1 , x t ])

[0027] h t =(1-z t )*h t-1 +z t *h t

[0028]

[0029] Among them, z t and r t represent the update gate and reset gate respectively, W z and W r represents the parameter training matrix, represents the candidate cell state at the current moment, h t and h t-1 Respectively represent the output of the hidden neuron at the current moment and the previous moment, σ and tanh represent the activation function;

[0030] Use the sigmoid function to define a smooth switching function s(x), and the switching function s(x) increases from 0 to 1 as x increases. The TanhReLU function can be defined as:

[0031] TanhReLU(x)=s(x)·ReLU(x)+(1-s(x))·Tanh(x)

[0032]

[0033] Among them, β is used to control the smoothness of switching, and θ is the threshold for switching to occur;

[0034] Substituting the definition of s(x) into the TanhReLU function, we get:

[0035]

[0036] Among them, β and θ are hyperparameters;

[0037] In the deep learning algorithm model, the Smooth ElasticNet regularization function expression is as follows:

[0038]

[0039] Among them, w is the parameter vector of the model, w i is the i-th parameter, n is the number of parameters, λ1, λ2, λ3 are regularization coefficients used to control the importance of different regularization terms, and ε is a small positive number used to avoid i = 0, a division by zero error occurs, and the function is i Closer to 0, it is smoother.

[0040] Preferably, in step S3, specifically: optimizing the hyperparameters in the model using an improved particle swarm optimization algorithm, including:

[0041] The basic algorithm of particle evolution QPSO is as follows:

[0042]

[0043] Among them, m best The average optimal position of the particle swarm; P ij is the optimal position of the ith particle in the jth dimension; P gj is the optimal position of the particle in the jth dimension; YesP ij and P gj A random position between ; M is the size of the particle group; M and u are random numbers in [0, 1]; α is the shrinkage diffusion coefficient.

[0044] Preferably, the step S4 comprises:

[0045] In the visualization interface, the detected abnormal data points or abnormal patterns are highlighted to identify abnormal data.

[0046] On the other hand, a detection system based on the above-mentioned substation monitoring data anomaly detection method based on deep learning is provided, comprising:

[0047] The data preprocessing module is used to preprocess and extract features of the raw data collected in the substation monitoring system;

[0048] The model building and analysis module is used to: after feature extraction is completed, use the deep learning algorithm to analyze and process the extracted features, identify abnormal points or abnormal patterns in the data, and optimize the hyperparameters of the constructed deep learning algorithm model;

[0049] The result output module is used to output the anomaly detection results in a visual manner.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] 1. The WIFS weighted integrated feature screening method combines the advantages of multiple feature selection methods such as filtering and embedding. The filtering method reduces the amount of computational effort for subsequent processing by preliminarily screening low-variance or irrelevant features; while the embedding method further screens out features that have a significant impact on the model's prediction performance through feature importance evaluation during model training. This comprehensive approach can more comprehensively consider the effectiveness and importance of features. In addition, WIFS introduces a weighting strategy that not only considers the feature importance scores given by the model, but also the correlation between features. This weighting method helps avoid selecting highly relevant but redundant features, thereby improving the representativeness and effectiveness of feature subsets. At the same time, the weighting strategy also allows weights to be adjusted according to specific application scenarios, increasing the flexibility and adaptability of the method;

[0052] 2. TanhReLU activation function helps to maintain the stability of gradients in the early stages of training and avoid training problems caused by gradients that are too large or too small. In addition, the TanhReLU function introduces a learnable switching mechanism that enables the function to automatically adjust its behavior according to different input values, which helps the model better fit complex data distributions;

[0053] 3. Smooth ElasticNet regularization function helps reduce overfitting and improve the stability of the model, and can flexibly control the relative importance of sparsity, smoothness, and weight change rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flow chart of the method of the present invention;

[0055] Figure 2 It is the overall model structure diagram of the present invention;

[0056] Figure 3 It is a flow chart of the improved particle swarm optimization algorithm of the present invention;

[0057] Figure 4 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0059] In the present invention, terms such as "upper", "lower", "left", "right", "front", "back", "vertical", "horizontal", "side", "bottom", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings. They are relational words determined only for the convenience of describing the structural relationships of the various parts or elements of the present invention, and do not specifically refer to any part or element in the present invention and should not be understood as limitations on the present invention.

[0060] In the present invention, terms such as "fixed connection", "connected", "connection", etc. should be understood in a broad sense, indicating that it can be fixedly connected, integrally connected or detachably connected; it can be directly connected or indirectly connected through an intermediate medium. Relevant scientific research or technical personnel in this field can determine the specific meanings of the above terms in the present invention according to specific circumstances, and they should not be understood as limiting the present invention.

[0061] Example:

[0062] like Figure 1 As shown, this embodiment provides a substation monitoring data anomaly detection method based on deep learning, comprising the following steps:

[0063] S1: Preprocess and extract features of the raw data collected in the substation monitoring system;

[0064] S2: After feature extraction is completed, a deep learning algorithm is used to analyze and process the extracted features to identify abnormal points or abnormal patterns in the data;

[0065] S3: Optimize the hyperparameters of the built deep learning algorithm model;

[0066] S4: Output the anomaly detection results in a visual manner.

[0067] The overall structural model is as follows Figure 2 shown.

[0068] Among them, step S1 is specifically to preprocess and extract features of the raw data collected in the substation monitoring system: preprocessing includes data cleaning (noise removal, missing value processing, etc.) and data transformation (normalization, standardization, etc.); feature extraction is to use statistical methods or special feature engineering techniques to extract key features useful for anomaly detection from the preprocessed data. These features should be able to fully characterize the difference between the normal state and the abnormal state of the data;

[0069] Feature screening is another important part of feature engineering. Its purpose is to select the most important features for model prediction from the original features to improve the accuracy and efficiency of the model. In order to screen out feature information useful for the model, the present invention proposes a weighted integrated feature screening method WIFS (Weighted Integrated Feature Selection). The specific steps are as follows:

[0070] 1. Preliminary screening by filtering method: Use the variance selection method or correlation coefficient method to set a threshold τ1 to filter out features with a variance less than τ1 or an absolute value of the correlation coefficient with the target variable less than τ1. The feature set obtained in this step is recorded as F1:

[0071] 2. Embedding method training basic model: Train a model that supports feature importance evaluation (such as random forest, gradient boosting tree, etc.) on F1, and record the model as M. Model M will output the importance score of each feature, recorded as importance i , where i represents the feature index;

[0072] 3. Correlation evaluation: For each pair of features in F1 (f i , f j ), calculate the correlation between them (such as Pearson correlation coefficient, mutual information, etc.), and obtain the correlation matrix C ij , where C ij Represents feature f i and f j The correlation between

[0073] 4. Weighted integration: Calculate the weighted score S of each feature based on the importance score of the feature and the correlation between the features i :

[0074]

[0075] Among them, |F1| represents the number of features in F1. This expression takes into account the importance of the feature and the average absolute value of its correlation with other features. If a feature is highly correlated with other features, its weighted score will be reduced due to the existence of correlation;

[0076] 5. Select the final feature: set a threshold τ2 and select the weighted score S i Features greater than τ2 are used as the final feature subset F find .

[0077] Step S2, anomaly detection based on deep learning algorithm, includes:

[0078] As a variant of recurrent neural network, Gated Recurrent Unit (GRU) has unique advantages in extracting feature information of data. On the one hand, GRU effectively controls the flow of information by introducing two gating mechanisms, Update Gate and Reset Gate. This mechanism enables GRU to selectively retain or forget historical information, thereby better capturing long-term dependencies in sequence data. Compared with traditional RNN, GRU's gating mechanism helps alleviate the gradient vanishing problem, allowing the model to learn long-distance dependencies more stably during training.

[0079] Compared with another RNN variant, the Long Short-Term Memory (LSTM), GRU has fewer parameters while maintaining similar performance, which requires less computing resources during training and is faster to train. The GRU model structure is relatively simple and easy to implement and debug. The basic algorithm is as follows:

[0080] z t =σ(W z ·[h t-1 , x t ])

[0081] r t =σ(W r ·[h t-1 , x t ])

[0082] h t =(1-z t )*h t-1 +z t *h t

[0083]

[0084] Among them, z t and r t represent the update gate and reset gate respectively, W z and W r represents the parameter training matrix, represents the candidate cell state at the current moment, ht and h t-1 Respectively represent the output of the hidden neuron at the current moment and the previous moment, σ and tanh represent the activation function;

[0085] The activation function in the neural network is very important, which determines the activation mode and strength of the neuron. The present invention discloses a TanhReLU activation function. The purpose of the TanhReLU function is to use the smoothness of the Tanh function to avoid neuron "death" when the input value is small, and to use the linearity of the ReLU function to keep the gradient stable when the input value is large. However, directly combining these two functions may not be the best choice because they are different in output range and gradient characteristics. Therefore, a transition mechanism needs to be designed to smoothly connect these two areas:

[0086] Assume that there is a smooth switching function s(x) that gradually increases from 0 to 1 as x increases. The sigmoid function or a similar smooth function can be used to define s(x). Then, the TanhReLU function can be defined as:

[0087] TanhReLU(x)=s(x)·ReLU(x)+(1-s(x))·Tanh(x)

[0088]

[0089] Among them, β is used to control the smoothness of the switching (the larger β is, the steeper the switching is), and θ is the threshold for the switching to occur (i.e., when x = θ, s(x) = 0.5);

[0090] Substituting the definition of s(x) into the TanhReLU function, we can infer:

[0091]

[0092] Among them, β and θ are hyperparameters, which are adjusted according to the specific task;

[0093] The TanhReLU function has a smooth curve and good nonlinear characteristics, which helps to maintain the stability of the gradient in the early stage of training and avoid training problems caused by excessive or small gradients. In addition, the TanhReLU function introduces a learnable switching mechanism, which enables the function to automatically adjust its behavior according to different input values, helping the model to better fit complex data distributions.

[0094] In deep learning models, regularization functions have advantages such as preventing overfitting, improving generalization ability, and reducing model complexity. This embodiment proposes a new regularization function that combines the characteristics of L1 and L2 regularization, called "ElasticNet-like" regularization, and makes some improvements to it to make it unique.

[0095] The regularization function expression of Smooth ElasticNet is as follows:

[0096]

[0097] Among them, w is the parameter vector of the model, w i is the i-th parameter, n is the number of parameters, λ1, λ2, λ3 are regularization coefficients used to control the importance of different regularization terms, and ε is a small positive number used to avoid i = 0, a division by zero error occurs, and the function is i It is smoother when it is close to 0. The third term is a smoothing term, which encourages the difference between adjacent parameters to be as small as possible, thus helping to produce smoother weight changes;

[0098] The Smooth ElasticNet regularization function proposed in the present invention has the following characteristics:

[0099] Sparsity: Pass Item, Smooth ElasticNet regularization encourages the model to produce sparse weights, similar to L1 regularization;

[0100] Smoothness: The L2 regularization term and the additional smoothing term work together to make the weight values ​​smoother, which helps reduce overfitting and improve the stability of the model;

[0101] Flexibility: By adjusting the values ​​of λ1, λ2, and λ3, the relative importance of sparsity, smoothness, and weight change rate can be flexibly controlled;

[0102] Therefore, when constructing the GRU model, the original activation function and regularization function were improved to further improve the overall performance of the model.

[0103] Step S3, using the improved particle swarm optimization algorithm to optimize the hyperparameters in the model, the process is as follows Figure 3 As shown, including:

[0104] Hyperparameters are the "knobs" used to control the learning process in machine learning algorithms, such as learning rate, regularization coefficient, number of hidden layers, number of neurons, etc. By systematically adjusting these parameters, the optimal configuration that best suits the current data set and task can be found, thereby significantly improving the performance of the model. The quantum-behaved particle swarm optimization (QPSO) algorithm is an optimization algorithm that simulates the group behavior of bird flocks or fish schools in nature. It introduces the concept of quantum dynamic kinematics based on the classical particle swarm optimization (PSO) algorithm, and enhances the randomness and global search ability of particles through quantum behavior. The algorithm can search for the global optimal solution in the entire feasible solution space. In model hyperparameter optimization, the QPSO algorithm can be used to find the best model configuration, which can achieve the best performance of the model on a given data set. The basic algorithm of particle evolution QPSO is as follows:

[0105]

[0106] Among them, m best The average optimal position of the particle swarm; P ij is the optimal position of the ith particle in the jth dimension; P gj is the optimal position of the particle in the jth dimension; YesP ij and P gj A random position between ; M is the size of the particle group; M and u are random numbers in [0, 1]; α is the shrinkage diffusion coefficient.

[0107] Step S4, visualizing the result output, including:

[0108] In the visual interface, the detected abnormal data points or abnormal patterns are highlighted in a special way (such as different colors, shapes, sizes or marks) so that monitoring personnel can quickly identify which data is abnormal. This intuitive method helps to quickly locate the problem.

[0109] On the other hand, this embodiment also provides a detection system based on the above-mentioned substation monitoring data anomaly detection method based on deep learning, including:

[0110] The data preprocessing module is used to preprocess and extract features of the raw data collected in the substation monitoring system;

[0111] The model building and analysis module is used to: after feature extraction is completed, use the deep learning algorithm to analyze and process the extracted features, identify abnormal points or abnormal patterns in the data, and optimize the hyperparameters of the constructed deep learning algorithm model;

[0112] The result output module is used to output the anomaly detection results in a visual manner.

[0113] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the described embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for detecting abnormality in substation monitoring data based on deep learning, characterized in that: The following steps are involved: S1: Preprocess and extract features of the raw data collected in the substation monitoring system; S2: After feature extraction is completed, a deep learning algorithm is used to analyze and process the extracted features to identify abnormal points or abnormal patterns in the data; S3: Optimize the hyperparameters of the built deep learning algorithm model; S4: Output the anomaly detection results in a visual manner.

2. According to the method for detecting abnormality of substation monitoring data based on deep learning in claim 1, it is characterized in that: The step S1 is specifically: The preprocessing of the raw data includes data cleaning and data transformation, which is used to clean and transform the data, remove noise, fill in missing values, and convert data types; The feature extraction of the original data is to extract key features useful for anomaly detection from the preprocessed data.

3. According to the method for detecting abnormality of substation monitoring data based on deep learning in claim 2, it is characterized in that: The key features extracted for anomaly detection specifically include: S11: Set a threshold τ1 to filter out features whose variance is less than τ1 or whose absolute value of correlation coefficient with the target variable is less than τ1. The obtained feature set is recorded as F1. S12: Train a model on F1 that supports feature importance evaluation, denoted as M, and M outputs the importance score of each feature, denoted as importance i , where i represents the feature index; S13: For each pair of features in F1 (f i , f j ), calculate its correlation and get the correlation matrix C ij , where C ij Represents the feature f i and f j The correlation between S14: Calculate the weighted score S of each feature based on the importance score of the feature and the correlation between the features i : Among them, |F1| represents the number of features in F1; S15: Set a threshold τ2 and select a weighted score S i Features greater than τ2 are used as the final feature subset F find .

4. According to the method for detecting abnormality of substation monitoring data based on deep learning in claim 3, it is characterized in that: The step S2 specifically includes the following process: According to the deep learning algorithm, the activation function in the neural network is calculated to determine the activation mode and strength of the neuron: z t =σ(W z ·[h t-1 ,x t ]) r t =σ(W r ·[h t-1 ,x t ]) h t =(1-z t )*h t-1 +z t *h t Among them, z t and r t Represent the update gate and reset gate respectively, W z and W r represents the parameter training matrix, represents the candidate cell state at the current moment, h t and h t-1 Respectively represent the output of the hidden layer neurons at the current moment and the previous moment, σ and tanh represent the activation function; Use the sigmoid function to define a smooth switching function s(x), and the switching function s(x) increases from 0 to 1 as x increases. The TanhReLU function can be defined as: TanhReLU(x)=s(x)·ReLU(x)+(1-s(x))·Tanh(x) Among them, β is used to control the smoothness of switching, and θ is the threshold for switching to occur; Substituting the definition of s(x) into the TanhReLU function, we get: Among them, β and θ are hyperparameters; In the deep learning algorithm model, the Smooth ElasticNet regularization function expression is as follows: Among them, w is the parameter vector of the model, w i is the i-th parameter, n is the number of parameters, λ1, λ2, λ3 are regularization coefficients used to control the importance of different regularization terms, and ε is a small positive number used to avoid i = 0, a division by zero error occurs, and the function is i Closer to 0, it is smoother.

5. According to the method for detecting abnormality of substation monitoring data based on deep learning according to claim 4, it is characterized in that: In the step S3, specifically, the hyperparameters in the model are optimized using the improved particle swarm optimization algorithm, including: The basic algorithm of particle evolution QPSO is as follows: Among them, m best The average optimal position of the particle swarm; P ij is the optimal position of the ith particle in the jth dimension; P gj is the optimal position of the particle in the jth dimension; YesP ij and P gj A random position between ; M is the size of the particle group; M and u are random numbers in [0, 1]; α is the shrinkage diffusion coefficient.

6. According to the method for detecting abnormality in substation monitoring data based on deep learning according to claim 5, it is characterized in that: The step S4 comprises: In the visualization interface, the detected abnormal data points or abnormal patterns are highlighted to identify abnormal data.

7. A detection system based on the substation monitoring data anomaly detection method based on deep learning as described in claim 6, characterized in that: include: The data preprocessing module is used to preprocess and extract features of the raw data collected in the substation monitoring system; The model building and analysis module is used to: after feature extraction is completed, use the deep learning algorithm to analyze and process the extracted features, identify abnormal points or abnormal patterns in the data, and optimize the hyperparameters of the constructed deep learning algorithm model; The result output module is used to output the anomaly detection results in a visual manner.

Images