Abnormity detection and automatic positioning method based on data curve

Abstract

The invention provides an anomaly detection and automatic positioning method based on a data curve. According to the method, historical data and image curve data of a system are collected, data preprocessing, feature extraction and selection are carried out, and an LSTM auto-encoder model is used to carry out time sequence feature learning and reconstruction error calculation, so that effective detection and positioning of anomalies are realized. Model training ensures accurate capture of a normal operation state by minimizing a reconstruction error, and performs anomaly detection and component positioning in combination with real-time data, so that the accuracy and efficiency of anomaly detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system operation and maintenance, and in particular to an abnormality detection and automatic positioning method based on a data curve. Background Art

[0002] In modern power operation and maintenance systems, monitoring of power loads and equipment operating status is one of the key links to ensure system stability and efficiency. The operating data in the power system is usually presented in the form of multiple curves, including a total curve and a load curve composed of multiple components. These curves reflect the real-time operating status of the power system as a whole and its components. However, with the continuous expansion of the scale and complexity of the power system, it has become a major challenge to detect curve anomalies in a timely and effective manner and determine the specific source of the anomalies.

[0003] The prior art invention patent with publication number CN111159961B discloses an anomaly detection method and system based on curve data, which specifically includes the following steps: step (1): first-level anomaly detection; input the curve data into a trained AE model for anomaly detection, obtain the k-dimensional data of the bottleneck layer and the actual output data, and use the loss function to calculate the loss value of the actual output data; step (2): second-level anomaly detection; input the k-dimensional data and loss value of the bottleneck layer obtained in step (1) as (k+1)-dimensional data into the second-level anomaly detector for anomaly detection, and output the anomaly value; when the anomaly value is greater than a preset threshold, the curve data to be detected is determined to be anomaly data; the second-level anomaly detector is implemented using a conventional anomaly detection algorithm. Summary of the invention

[0004] Traditional curve monitoring methods for power operation and maintenance systems mainly rely on comparing current values ​​with historical averages (such as the average of the previous five minutes), or manual judgment by operation and maintenance personnel based on experience. Although these methods are intuitive, they have the following defects in practical applications: 1. Time lag and accuracy limitation: The historical mean comparison method will produce time lag in the detection process, especially in complex load change scenarios, which can easily lead to delayed response to anomalies. In addition, this simple threshold comparison cannot effectively deal with the nonlinear and diversified operation modes existing in the power system, thereby reducing the accuracy of anomaly detection; 2. Complexity of multi-component curves: The total sum curve is usually composed of multiple components, and the interaction of these components affects the operating status of the entire system. Once the total sum curve is abnormal, it is difficult for traditional methods to quickly and accurately locate the specific abnormal component, resulting in operation and maintenance personnel spending a lot of time to conduct step-by-step troubleshooting when troubleshooting; 3. Challenges of multimodal data fusion: The operating status of the power system is a complex multi-dimensional problem, involving multiple modal information such as dispatch curve data and historical operation records. Traditional monitoring methods often only focus on single-dimensional data analysis and are difficult to effectively integrate and process these multimodal data, thus limiting the comprehensive understanding and analysis of curve anomalies. This single-dimensional analysis method cannot fully utilize the correlation between various types of data, reducing the accuracy and interpretability of anomaly detection.

[0005] To solve the above problems, the present invention provides a technical solution: a method for abnormality detection and automatic positioning based on data curves, comprising the following steps: S1, collect the historical data of the system total addition curve and its components, as well as the corresponding image curve data, and pre-process the data; S2, extract data features and build models; S3. Model training and optimization based on historical data; S4, real-time data is input into the model, the model reconstructs the data according to the time step and calculates the reconstruction error; S5. Set a reconstruction error threshold based on historical data, and perform anomaly detection and location based on the threshold.

[0006] Specifically, the data preprocessing in S1 includes data cleaning, removing outliers and missing values, using interpolation to fill missing data, normalizing all input data, and converting data into a standard distribution through standardization.

[0007] Specifically, extracting data features in S2 includes analyzing the data features of the total sum curve and each component, using the feature importance analysis method based on information gain to extract key features for modeling; and using statistical features and time series features to construct a representative feature set.

[0008] Specifically, the construction model in S2 includes constructing an LSTM autoencoder, and the input of the model is the value of each component at each time point and the value of the total sum curve; the encoder part of the model uses multi-layer LSTM units to process the input time series and gradually compress the time series into a low-dimensional representation; the decoder part of the model reconstructs all input data from the low-dimensional representation of the encoder, including each component and the total sum curve.

[0009] Specifically, the decoder part of the model reconstructs the input data by gradually outputting time series data through LSTM units; the model calculates the reconstruction error between the input and the reconstructed output based on the mean square error as a measurement indicator.

[0010] Specifically, the model training in step S3 specifically includes: training on historical normal data, so that the model can reconstruct all input data as accurately as possible, and minimize the gap between the input and the reconstructed value by optimizing the mean square error.

[0011] Specifically, the model optimization in step S3 specifically includes: adding Dropout regularization and early stopping strategy to avoid overfitting, and adjusting model hyperparameters through grid search method to optimize model performance.

[0012] Specifically, the calculation of the reconstruction error in step S4 specifically includes: inputting the real-time collected data, including all components and the total sum curve, into the trained LSTM autoencoder model, and the model attempts to reconstruct all input data according to the current time step to generate a reconstruction value corresponding to the input; for each time point, the mean square error is used to calculate the reconstruction error of each component and the total sum curve.

[0013] Specifically, the reconstruction error threshold in step S5 includes an overall abnormality detection threshold and a reconstruction error threshold set independently for each component. The threshold is determined by the mean and standard deviation of the reconstruction error of each component of the historical normal operation data and the total sum curve. The specific formula is as follows:

[0014] Among them, Th represents the reconstruction error threshold, μ represents the mean of the reconstruction error, σ represents the standard deviation of the reconstruction error, and k is an adjustment factor used to control the sensitivity of anomaly detection.

[0015] Specifically, in step S5, whether the overall reconstruction error of the total sum curve exceeds the overall anomaly detection threshold is used to determine whether the total sum curve has an anomaly. If an anomaly exists, the reconstruction error of each component is analyzed to determine whether it exceeds the independent reconstruction error threshold of each component, and their contribution to the anomaly of the total sum curve is evaluated. A step-by-step elimination method is used, starting from the largest reconstruction error, and components are gradually eliminated to observe changes in the reconstruction error of the total sum curve and confirm the specific source of the anomaly.

[0016] The beneficial effects of the present invention are: it can make full use of the historical operation data and image curve data of the power system, capture the complex time series characteristics of the total sum curve and its components, and when an abnormality occurs, it can not only quickly detect the abnormality of the total sum curve, but also automatically locate the specific abnormal component. This method should be able to process multi-modal inputs and adapt to complex nonlinear relationships to improve the system's abnormality detection and positioning capabilities and ensure the efficient operation and stability of the power operation and maintenance system. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a specific flow chart of the method of the present invention. DETAILED DESCRIPTION

[0018] The technical solutions in the disclosed embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings.

[0019] Embodiment 1: A method for abnormality detection and automatic positioning based on data curves, such as Figure 1 As shown, the following steps are included: S1, collect the historical data of the system total addition curve and its components, as well as the corresponding image curve data, and pre-process the data; S2, extract data features and build models; S3. Model training and optimization based on historical data; S4, real-time data is input into the model, the model reconstructs the data according to the time step and calculates the reconstruction error; S5. Set a reconstruction error threshold based on historical data, and perform anomaly detection and location based on the threshold.

[0020] Specifically, S1 collects historical data of the total sum curve and historical data of each component, as well as the corresponding image curve data. It collects historical data for one year to ensure that the data covers various operating scenarios. To ensure the comprehensiveness of the data, the complete historical data of all multiple components that constitute the total sum curve is collected synchronously. A certain collection frequency is maintained to ensure the time resolution of the data. It covers historical data in text format and curve data in image format. The bimodal characteristics of the data can effectively improve the generalization ability of the model and ensure the accuracy of anomaly detection.

[0021] The data preprocessing in S1 includes data cleaning, removing outliers and missing values, using interpolation to fill missing data, normalizing all input data, and converting data into a standard distribution through standardization.

[0022] Data preprocessing specifically includes: first cleaning the collected data and removing outliers and missing values ​​to ensure the integrity of the data.

[0023] Secondly, interpolation and mean are used to fill in all missing data.

[0024] Secondly, all input data are normalized to ensure that different feature dimensions have the same numerical range, which is conducive to the stable learning and training of the model. Normalization usually uses Min-Max normalization to scale the data to a specific range:

[0025] Among them, x is the original data, and are the minimum and maximum values ​​of the feature, respectively. is the normalized value. This method ensures that the values ​​of different features are in the same range, thus avoiding unstable model training due to different feature dimensions.

[0026] The data is converted to a standard distribution through standardization to ensure that the feature data has a similar scale, which is convenient for subsequent feature extraction and model training. The formula for standardization is as follows:

[0027] Among them, μ is the mean of the data, σ is the standard deviation of the data, is the standardized value. Standardization helps improve the model's ability to learn features of different scales.

[0028] Specifically, the extraction of data features in S2 includes analyzing the data features of the total sum curve and each component, using the feature importance analysis method based on information gain to extract key features for modeling; and using statistical features and time series features to construct a representative feature set. Specifically including: Analysis and extraction: Analyze the data features of the total sum curve and each component, and extract key features for modeling. Use statistical features (such as mean, variance, maximum value, minimum value, etc.) and time series features (such as periodicity, trend, etc.) to construct a representative feature set.

[0029] Feature selection: Select the features that have the greatest impact on model prediction to reduce computational complexity. Use feature importance analysis (based on information gain) to select the most representative features to improve the accuracy and stability of the model. Information gain formula (information gain is used to measure the contribution of a feature to a classification task):

[0030] in It represents the information gain of feature X to target variable T, H(T) is the entropy of the target variable, and H(T|X) is the target entropy given feature X. The larger the information gain, the more important the feature is for prediction.

[0031] Specifically, the construction model in S2 includes constructing an LSTM autoencoder, and the input of the model is the value of each component at each time point and the value of the total sum curve; the encoder part of the model uses multi-layer LSTM units to process the input time series and gradually compress the time series into a low-dimensional representation; the decoder part of the model reconstructs all input data from the low-dimensional representation of the encoder, including each component and the total sum curve.

[0032] Specifically, the decoder part of the model reconstructs the input data by gradually outputting time series data through LSTM units; the model calculates the reconstruction error between the input and the reconstructed output based on the mean square error as a measurement indicator.

[0033] The design of LSTM autoencoder specifically includes: Model construction: construct an LSTM autoencoder, and the input of the model is the value of each component at each time point (a total of N components) and the value of the total added curve.

[0034] Encoder part: Use multi-layer LSTM units to process the input time series and gradually compress the time series into a low-dimensional representation.

[0035] The encoder part can capture the complex temporal dependencies between various components in the power system and compress the high-dimensional time series into low-dimensional feature representations.

[0036] Decoder part: The decoder tries to reconstruct all input data from the encoder's low-dimensional representation, including individual components and the total sum curve.

[0037] The process used by LSTM units for decoding is similar to the encoding process, by gradually outputting time series data for comparison with the original input data. The decoder outputs the reconstructed time series data.

[0038] Reconstruction error calculation formula: Calculate the difference between the input and the reconstructed output, and use the mean square error (MSE) as the measurement indicator. The decoder outputs the reconstructed time series data for subsequent calculation of the reconstruction error.

[0039] Specifically, the model training in step S3 specifically includes: training on historical normal data, so that the model can reconstruct all input data as accurately as possible, and minimize the gap between the input and the reconstructed value by optimizing the mean square error.

[0040] The training objectives and loss calculation process specifically include: Training goal: Train on historical normal data to enable the model to reconstruct all input data as accurately as possible, and minimize the gap between the input and the reconstructed value by optimizing the mean square error (MSE).

[0041]

[0042] in, is the original input, is the reconstructed output, and N is the number of samples.

[0043] Specifically, the model optimization in step S3 specifically includes: adding Dropout regularization and early stopping strategy to avoid overfitting, and adjusting model hyperparameters through grid search method to optimize model performance.

[0044] The specific process of model optimization includes: in order to improve the generalization ability of the model, adding regularization (such as Dropout layer) and early stopping strategy to avoid overfitting. At the same time, optimizing model performance by adjusting parameters to ensure the model's ability to distinguish between normal and abnormal situations.

[0045] Dropout regularization: By randomly discarding some neurons at each iteration, the risk of overfitting between neurons is reduced, and the experiment is adjusted according to the performance of the validation set. The method is as follows:

[0046] The Dropout probability is set to 0.5, that is, p=0.5, and x is a data point.

[0047] Early stopping strategy: During the training process, monitor the error change of the validation set. When the validation error does not show significant improvement within several consecutive epochs, stop training to prevent overfitting. Patience is set to 5 epochs.

[0048] Parameter adjustment process: The model hyperparameters such as the number of LSTM layers, number of hidden units, learning rate, etc. are adjusted by grid search method. The learning rate range is 0.0001 to 0.01 to find the most suitable setting for stable convergence of the model.

[0049] Specifically, the calculation of the reconstruction error in step S4 specifically includes: inputting the real-time collected data, including all components and the total sum curve, into the trained LSTM autoencoder model, and the model attempts to reconstruct all input data according to the current time step to generate a reconstruction value corresponding to the input; for each time point, the mean square error is used to calculate the reconstruction error of each component and the total sum curve.

[0050] The specific steps include: inputting the real-time collected data (including all components and the total sum curve) into the trained LSTM autoencoder model. The model attempts to reconstruct all input data according to the current time step and generates a reconstructed value corresponding to the input.

[0051] Reconstruction error calculation: For each time point, the reconstruction error of each component and the total sum curve is calculated. The reconstruction error is calculated using the mean square error (MSE).

[0052] Specifically, the reconstruction error threshold in step S5 includes an overall abnormality detection threshold and a reconstruction error threshold set independently for each component, and the threshold is determined by the mean and standard deviation of the reconstruction error of each component and the total sum curve of the historical normal operation data.

[0053] The specific steps of setting the threshold include: Historical data statistics: based on historical normal operation data, the mean and standard deviation of the reconstruction error of each component and the total sum curve are calculated.

[0054] The reconstruction error threshold of each component is determined by historical statistical data. The following formula can usually be used to set the threshold so that it can reasonably distinguish normal and abnormal data:

[0055] Among them, Th represents the reconstruction error threshold, μ represents the mean of the reconstruction error, σ represents the standard deviation of the reconstruction error, and k is an adjustment factor used to control the sensitivity of anomaly detection. It is usually set to 2 or 3, indicating the multiple of the threshold deviation from the mean. This method is based on the "68-95-99.7 rule" in statistics, that is, in a normal distribution, the probability that the data is within the range of the mean plus or minus two standard deviations is about 95%. Different values ​​of the adjustment factor k can control the sensitivity of anomaly detection. The larger the k, the looser the threshold and the fewer anomalies are detected. The smaller the k, the more anomalies are detected.

[0056] Overall threshold and independent component threshold setting: Set an overall anomaly detection threshold based on historical normal operation data to determine the anomaly of the total sum curve. Set an independent reconstruction error threshold for each component so that the specific abnormal component can be further located when the total sum curve is abnormal.

[0057] Specifically, in step S5, whether the total sum curve is abnormal is determined by whether the overall reconstruction error of the total sum curve exceeds the overall abnormality detection threshold. When the reconstruction error of the total sum curve exceeds the set overall threshold, it is determined that the total sum curve is abnormal. The time point when the abnormality occurs and the size of the reconstruction error are recorded for subsequent analysis.

[0058] When anomalies exist, analyze whether the reconstruction error of each component exceeds the independent reconstruction error threshold of each component, and evaluate their contribution to the anomaly of the total sum curve. Use the step-by-step elimination method, starting from the largest reconstruction error, and gradually eliminate components to observe the changes in the reconstruction error of the total sum curve to confirm the specific source of the anomaly.

[0059] The specific steps of analyzing the abnormal location include: analyzing the reconstruction error of each component to determine whether it exceeds its independent threshold. If the reconstruction error of a component exceeds its set threshold, it is considered that the component may be the main source of the total sum curve abnormality.

[0060] Sorting and contribution analysis: Sort the reconstruction errors of all components and find the components with the largest reconstruction errors exceeding the threshold. Mark these components as possible anomaly sources and evaluate their contribution to the total anomaly of the curve.

[0061] Confirmation of abnormal source: If the reconstruction error of multiple components exceeds the threshold, further logical judgment and analysis are used to confirm the main abnormal source. The step-by-step elimination method is used, that is, starting from the largest reconstruction error, gradually eliminating components and observing the changes in the reconstruction error of the total sum curve to further confirm the specific source of the abnormality.

[0062] After detecting an anomaly, this embodiment has an alarm mechanism and reports the production and operation of the anomaly.

[0063] Alarm triggering: When the reconstruction error of the total sum curve or some components exceeds the set threshold, the alarm is triggered immediately. The system generates an alarm signal, which can prompt the operator in two ways: visually (such as a red indicator light on the control panel) and sound (such as a siren).

[0064] Alarm type classification: The system divides alarms into different levels (such as low, medium, and high) according to the severity of the abnormality. For example, if only a few components are abnormal and the deviation is not large, a low-level alarm is triggered; if multiple components are abnormal and the reconstruction error is large, a high-level alarm is triggered.

[0065] Notification and log recording: The alarm information is notified to the relevant operation and maintenance personnel through SMS or email to ensure timely response. The alarm information includes the time when the abnormality occurred, the abnormal component, and the size of the reconstruction error. At the same time, the alarm record is stored in the log database for subsequent analysis and review.

[0066] Operation and maintenance report generation: The system automatically generates a detailed abnormality report based on the test results. The report includes: the total sum curve and the reconstruction error information of the component. The time point and duration of the abnormality. The specific description of the abnormal component and its evaluation of the overall impact on the system.

[0067] Embodiment 2: The present invention provides a method for curve anomaly detection and component automatic positioning of an electric power operation and maintenance system based on LSTM autoencoder. A provincial-level local dispatching center is responsible for the operation monitoring of the 500kV backbone power grid, 220kV transmission network and substations at all levels within its jurisdiction. The area contains multiple large industrial loads and important users. The local dispatching monitors the entire power grid in real time through the SCADA system. In order to improve the reliability of power grid operation, the local dispatching uses the method of this embodiment to establish an anomaly detection system based on the LSTM autoencoder. The specific implementation process of the method of the present invention is illustrated below through the actual application case of the power dispatching center (local dispatching).

[0068] The specific steps include: 1. Data collection and preprocessing In this step, first collect the historical data of the total plus curve and its components in the power operation and maintenance system, specifically collect one year of historical data to ensure that the data covers various operation scenarios. To ensure the comprehensiveness of the data, synchronously collect the complete historical data of all the more than 80 components that constitute the total plus curve. Sampling frequency: collect once every 5 minutes to ensure the time resolution of the data. The collected data includes the historical data of the total plus curve (text format) and the image curve data, as well as the corresponding historical data of all component points. In order to ensure the adequacy and representativeness of the data, a full year of operation data was collected as a training sample.

[0069] In specific application scenarios, the system collects the following data in real time: - SCADA system data: * Power of each section of 500kV main grid * 220kV bus voltage of each substation * Power of important users’ incoming lines * Power factor of each substation - PMU data: * Key node phase angle difference * Frequency and frequency change rate - Equipment operation data: * Main transformer operating parameters * Circuit breaker status * Protection device status - Environmental information: * Weather data * Lightning warning information - Scheduling information: * Day-ahead power generation plan * Maintenance plan * Load forecast curve After data collection is completed, the data is cleaned to remove outliers and missing values ​​to ensure data integrity and consistency. Missing values ​​are filled using interpolation or mean, and then data entries with completely repeated timestamps are deleted and all data are standardized and normalized to eliminate the impact of different feature scales, thereby ensuring the stability and effectiveness of the model training process.

[0070] Specifically, in this embodiment, the collected data is preprocessed, including removing data entries with duplicate timestamps through data cleaning and using linear interpolation to fill in missing data. Then the data is normalized, all data is mapped to the interval [0,1], and the scale differences between different features are eliminated. Taking the data collected at 14:25 on xx / xx / 2024 as an example, the original data is first cleaned and normalized to ensure data quality.

[0071] Secondly, extract features from text data and image curves. For text data, extract statistical features (such as mean, variance, etc.) and time series features; for image curves, extract curve morphology features and change features. Use the information gain method to select the most representative features to improve the efficiency of the model. In the above case, a complete feature representation is obtained through feature extraction.

[0072] 2. Feature Engineering and Model Architecture Design The collected multimodal data is subjected to feature extraction and selection, and representative key features are extracted through statistical features and time series features. For example, statistical features such as mean and variance are extracted to analyze the changing trends of various components in the power system. At the same time, the information gain method is used to select the features that have the greatest impact on the detection results to reduce the computational complexity.

[0073] The construction of the LSTM autoencoder model is a key part in this embodiment. The encoder part uses multi-layer LSTM units to process the input time series and compress the high-dimensional time series into a low-dimensional feature representation to capture the complex temporal dependencies between components in the system. The decoder part reconstructs based on the low-dimensional features and generates outputs similar to the original input, thereby calculating the reconstruction error and providing a basis for subsequent anomaly detection.

[0074] In terms of model design, this embodiment constructs a two-layer LSTM autoencoder structure. The encoder part uses a two-layer LSTM network, the first layer contains 128 hidden units, and the second layer contains 64 hidden units, which compresses the input data into a 32-dimensional feature vector. The decoder part adopts a symmetrical structure and reconstructs the input data through a two-layer LSTM network.

[0075] 3. Model training and optimization The model training is conducted using historical normal data, with the goal of minimizing the mean square error (MSE) between the input data and the reconstructed value, to ensure that the model can accurately capture the characteristics of normal operating conditions. To avoid overfitting, Dropout regularization is added to the model training process, and an early stopping strategy is adopted to stop training when the verification error no longer improves significantly. At the same time, hyperparameters such as learning rate and number of hidden units are adjusted through grid search to ensure the optimal performance of the model.

[0076] The model training in this embodiment uses mean square error (MSE) as the loss function, uses Adam optimizer for optimization, and the initial learning rate is set to 0.001. To prevent overfitting, the Dropout strategy (dropout rate is 0.5) and early stopping mechanism (patience is set to 5 epochs) are used. The training data set contains one year of historical data, 80% of which is used for training and 20% for validation.

[0077] 4. Real-time data processing, anomaly detection and component positioning After the model training is completed, the real-time collected power data is input into the trained LSTM autoencoder. The model first attempts to reconstruct all input data and calculates the reconstruction error of each component and the total sum curve. Through statistical analysis of historical data, the anomaly detection threshold of each component and the total sum curve is set. If the reconstruction error exceeds the threshold, it is judged that the component is abnormal.

[0078] To locate the source of the anomaly, the system sorts the reconstruction errors of all components, finds the component with the largest error and exceeds the threshold, and marks it as the source of the anomaly. Further analysis is performed through the step-by-step elimination method to confirm the component with the greatest impact on the total sum curve anomaly, and finally locate the specific source of the anomaly.

[0079] In actual applications, the system receives new data every 5 minutes, performs the same preprocessing on the data, and then inputs it into the trained model. In the example at 14:25 on xx / xx / 2024, the system analyzes the preprocessed data and finds that the reconstruction error reaches 0.185, significantly exceeding the normal threshold of 0.132 for this component (based on the statistical characteristics of the reconstruction error of normal samples, taking the mean plus two standard deviations). The system considers that there is an anomaly and immediately triggers a low-level alarm, and begins anomaly location analysis.

[0080] By analyzing the reconstruction errors of all 80 components, the system found that the error of the component data of the No. 3 power distribution circuit ranked first, with an abnormal contribution of 76%, while the reconstruction errors of other components were all within the normal range. This sorting and step-by-step elimination method based on the size of the reconstruction error accurately located the source of the abnormality.

[0081] The system has a three-level alarm mechanism. In this case, since only a single component was abnormal and the error was relatively small, the system triggered a low-level alarm. The alarm information was promptly sent to the relevant operation and maintenance personnel via SMS and email, and was also displayed on the system control panel. After receiving the alarm, the operation and maintenance personnel immediately conducted an on-site inspection and found that a certain electrical equipment in the circuit was indeed faulty. Timely handling avoided greater losses.

[0082] Each time an anomaly is detected, the system automatically generates an operation and maintenance report, which includes the time when the anomaly occurred, a list of anomaly components, reconstruction error value (0.185), anomaly severity assessment, etc. This information is stored in the system database for subsequent analysis and system optimization.

[0083] In specific application scenarios, the following abnormal situations occurred: Operation background: During the peak period of electricity consumption in winter, the total regional load is about 12GW, of which heavy industrial load accounts for 45%.

[0084] At 19:30, the system detected an abnormal situation: SCADA real-time data: - A 500kV substation: * A bus voltage: 515kV (normal range) * Bus voltage B: 485kV (lower than normal) * #1 main transformer load: 82% (normal range) * #2 main transformer load: 89% (normal range) PMU data: - Phase angle difference abnormality: * Normal value range: 15°-20° * Current value: 23.5° (outside the normal range) Section power data: - South section: 2.8GW (close to the limit) - North section: 1.2GW (normal range) The LSTM autoencoder analyzes the anomaly as follows: The system analyzes the collected multi-dimensional data: 1) Overall state reconstruction error: 0.185 (exceeding the warning threshold of 0.15) 2) Component reconstruction error: - 500kV B bus voltage: 0.268 (seriously exceeded the standard) - Sectional power: 0.142 (normal range) - Phase angle difference: 0.195 (exceeds threshold) 3) Correlation analysis: - Equipment status: B bus PT related signal is abnormal - Protection device: Not activated - Load status: Continuous upward trend 5. Alarm mechanism and operation and maintenance report generation When an abnormality is detected, the system triggers an alarm mechanism, which is divided into low, medium and high levels according to the severity of the abnormality. The alarm information is sent to the relevant operation and maintenance personnel via SMS or email, and is displayed on the system control panel at the same time, prompting the operator to handle it in time.

[0085] The alarm information in the specific application scenario is as follows: Time: 2024-xx-xx 19:30 Warning level: Level II Warning area: 500kV substation Abnormal characteristics: - B bus voltage low: 485kV - Phase angle difference abnormal: 23.5° - Reconstruction error: 0.268 (exceeding the threshold by 78.7%) Real-time impact: 1. Impact range: May affect the transport capacity of the southward section 2. Potential risk: may cause the section power to exceed the limit Scheduling suggestions: 1. Immediately reduce power transmission in the south section 2. Start the backup power supply 3. Notify the operation and maintenance personnel to check the B busbar equipment 4. Prepare a load transfer plan The operator handles the warning as follows: 19:30 - System issued an alert 19:31 - Dispatcher confirms warning information 19:32 - Dispatch instructions issued: - Reduce power transmission in the south section - Activate emergency backup power supply 19:35 - Operation and maintenance personnel arrived at the scene 19:38 - Found a fault in the secondary circuit of the B bus PT 19:45 - Troubleshooting completed, system restored to normal The system automatically generates an operation and maintenance report based on the test results. The report content includes: the reconstruction error information of the total sum curve and each component, the time and duration of the abnormality, the description of the abnormal component and its impact assessment on the overall system. These reports provide an important basis for subsequent abnormality review and system optimization.

[0086] The operation and maintenance reports in specific application scenarios are as follows Event Overview: - Time of occurrence: 2024-xx-xx 19:30 - Anomalous Nature: Equipment Malfunction - Impact scope: 500kV substation B busbar system Processing sequence: 1. Anomaly Detection: 19:30 2. Warning issued: 19:30 3. Dispatch response: 19:31 4. Order issued: 19:32 5. Arrival at the site: 19:35 6. Troubleshooting: 19:38-19:45 System Response Assessment: 1. Timeliness of testing: - Detect voltage abnormalities at an early stage - 3-5 minutes ahead of traditional alarm thresholds 2. Early warning accuracy: - Accurately locate the B bus PT system - The reconstruction error index is significant (0.268) 3. Processing efficiency: - Response time: 1 minute - Total processing time: 15 minutes - Avoids possible cross-section power limit This case verifies the advantages of the method of the present invention in power dispatching and operation and maintenance: 1. Early warning: Detect anomalies before the traditional alarm threshold is triggered, saving valuable time for scheduling decisions; 2. Accurate positioning: Combine multi-source data for analysis to quickly locate specific faulty equipment and types; 3. Scheduling guidance: Provide professional scheduling suggestions and support quick decision-making; 4. Experience accumulation: Form a complete fault feature library and continuously improve abnormality identification capabilities The actual application in this substation shows that this method can accurately detect the abnormal situation of the total addition curve and quickly locate the specific abnormal component. Compared with the traditional threshold comparison and manual judgment method, this method has higher detection accuracy and faster response speed, which significantly improves the abnormal detection efficiency of the power operation and maintenance system. At the same time, this method has a high degree of automation, reduces manual intervention, and improves the efficiency of operation and maintenance work. For example, in the above case, from the occurrence of the abnormality to the system alarm, locating the source of the fault, and then to the completion of the operation and maintenance personnel, the entire process took only about 30 minutes, which greatly improved the efficiency of fault handling.

[0087] The present invention proposes a method for detecting anomalies in curves and locating components of power operation and maintenance systems based on LSTM autoencoders. First, the method achieves all-round system monitoring by fusing multimodal data, including equipment operating status, environmental factors, and various sensor data and historical records. This multi-dimensional data fusion not only improves the comprehensiveness of detection, but also significantly enhances the generalization ability of the model, making anomaly detection more accurate and reliable.

[0088] In terms of data processing and feature engineering, this method uses an intelligent processing mechanism that can automatically extract statistical features and time series features, and select features through information gain. This automated feature processing method greatly reduces the difficulty of manual feature engineering and improves the efficiency of system operation. At the same time, the end-to-end anomaly detection framework constructed by the LSTM autoencoder can automatically learn complex time series dependencies without manually setting cumbersome rules, and has strong adaptive capabilities.

[0089] In terms of anomaly detection and location, this method demonstrates excellent precision positioning capabilities. By analyzing reconstruction errors and combining them with the step-by-step elimination method, the system can not only quickly identify anomalies, but also accurately locate specific anomaly components and support collaborative analysis of multiple anomaly components. This precise positioning capability greatly improves the efficiency of problem handling and reduces the troubleshooting time of operation and maintenance personnel.

[0090] The method also establishes a flexible alarm mechanism, implements graded alarms according to the degree of abnormality, and supports multiple alarm methods to ensure that relevant personnel can obtain abnormal information in a timely manner. At the same time, the detailed operation and maintenance report automatically generated by the system provides an important basis for subsequent analysis and optimization. In terms of model performance, by adopting Dropout regularization and early stopping strategy, combined with grid search to optimize hyperparameters, the model is ensured to have strong generalization performance and stability.

[0091] In addition, the method has excellent real-time processing capabilities and interpretability. The system can monitor and analyze data in real time, respond quickly to abnormal situations, and clearly display the degree and source of abnormalities through quantified reconstruction errors, so that operation and maintenance personnel can intuitively understand and deal with problems. This efficient and intelligent anomaly detection scheme not only improves the operational stability of the power system, but also significantly reduces the operation and maintenance costs, providing a strong guarantee for the safe and reliable operation of the power system. In general, this method overcomes the limitations of traditional anomaly detection methods and provides a more advanced and reliable solution.

[0092] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for anomaly detection and automatic positioning based on data curves, characterized in that: The following steps are involved: S1, collect the historical data of the system total addition curve and its components, as well as the corresponding image curve data, and pre-process the data; S2, extract data features and build models; S3. Model training and optimization based on historical data; S4, real-time data is input into the model, the model reconstructs the data according to the time step and calculates the reconstruction error; S5. Set a reconstruction error threshold based on historical data, and perform anomaly detection and location based on the threshold.

2. The method for abnormality detection and automatic positioning based on data curve according to claim 1, characterized in that: The data preprocessing in S1 includes data cleaning, removing outliers and missing values, using interpolation to fill missing data, normalizing all input data, and converting data into a standard distribution through standardization.

3. The method for abnormality detection and automatic positioning based on data curve according to claim 1, characterized in that: Extracting data features in S2 includes analyzing the data features of the total sum curve and each component, using a feature importance analysis method based on information gain to extract key features for modeling; and using statistical features and time series features to construct a representative feature set.

4. The method for abnormality detection and automatic positioning based on data curve according to claim 1, characterized in that: The construction model in S2 includes constructing an LSTM autoencoder, the input of the model is the value of each component at each time point and the value of the total addition curve; the encoder part of the model uses a multi-layer LSTM unit to process the input time series and gradually compress the time series into a low-dimensional representation; the decoder part of the model reconstructs all input data from the low-dimensional representation of the encoder, including each component and the total addition curve.

5. The method for abnormality detection and automatic positioning based on data curve according to claim 4, characterized in that: The decoder part of the model reconstructs the input data by gradually outputting time series data through LSTM units; the model calculates the reconstruction error between the input and the reconstructed output based on the mean square error as a measurement indicator.

6. The method for abnormality detection and automatic positioning based on data curve according to claim 1, characterized in that: The model training in step S3 specifically includes: training on historical normal data so that the model can accurately reconstruct the input data and minimize the gap between the input and the reconstructed value by optimizing the mean square error.

7. The method for abnormality detection and automatic positioning based on data curve according to claim 1, characterized in that: The model optimization in step S3 specifically includes: adding Dropout regularization and early stopping strategy to avoid overfitting, and adjusting model hyperparameters through grid search method to optimize model performance.

8. The method for abnormality detection and automatic positioning based on data curve according to claim 1, characterized in that: The calculation of the reconstruction error in step S4 specifically includes: inputting the real-time collected data, including all components and the total sum curve, into the trained LSTM autoencoder model, and the model attempts to reconstruct all input data according to the current time step to generate a reconstruction value corresponding to the input; for each time point, the reconstruction error of each component and the total sum curve is calculated using the mean square error.

9. The method for abnormality detection and automatic positioning based on data curve according to claim 1, characterized in that: The reconstruction error threshold in step S5 includes the overall abnormality detection threshold and the reconstruction error threshold set independently for each component. The threshold is determined by the mean and standard deviation of the reconstruction error of each component and the total sum curve of the historical normal operation data. The specific formula is as follows: Among them, Th represents the reconstruction error threshold, μ represents the mean of the reconstruction error, σ represents the standard deviation of the reconstruction error, and k is an adjustment factor used to control the sensitivity of anomaly detection.

10. The method for abnormality detection and automatic positioning based on data curve according to claim 1 or 9, characterized in that: In step S5, whether the overall reconstruction error of the total sum curve exceeds the overall anomaly detection threshold is used to determine whether the total sum curve has an anomaly. If an anomaly exists, the reconstruction error of each component is analyzed to see whether it exceeds the independent reconstruction error threshold of each component, and their contribution to the anomaly of the total sum curve is evaluated. A step-by-step elimination method is used, starting from the largest reconstruction error, and components are gradually eliminated to observe the changes in the reconstruction error of the total sum curve and confirm the specific source of the anomaly.

Images