Energy abnormal data cleaning method and system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an energy abnormal data cleaning method and system based on artificial intelligence, and the method comprises the steps: obtaining initial monitoring data of an energy system, and carrying out the preprocessing; then dynamically adjusting a threshold interval of the range check according to a processing result, and generating primary processing data containing an abnormal mark; performing correlation detection on the primarily processed data, constructing a correlation mode under a normal working condition through an equipment topological graph, and scoring abnormal samples; performing data cleaning on the scoring result to obtain a preliminary cleaning strategy; and finally, performing quality evaluation on the cleaning result, and adjusting parameters of the preliminary cleaning strategy according to an evaluation result to obtain a target cleaning strategy. According to the invention, through dynamic adjustment and quality evaluation, the accuracy and adaptability of the cleaning strategy are ensured, the efficiency and quality of data cleaning are significantly improved, the operation and maintenance cost and the manual dependence are reduced, and the adaptability and expandability of the system are enhanced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based method and system for cleaning energy anomaly data. Background Technology

[0002] As a pillar industry of the nation, the accuracy and reliability of energy consumption data in the intelligent manufacturing sector are crucial to the implementation of the national energy security strategy and the advancement of the "dual carbon" target. Electricity consumption accounts for over 75% of total energy consumption, making energy data a key basis for operational efficiency assessment, energy-saving strategy formulation, and resource allocation optimization, playing a decisive role in the industry's green transformation.

[0003] The energy data acquisition process is affected by factors such as equipment characteristics, environmental interference, and transmission links, and generally exhibits three types of anomalies:

[0004] Sudden anomalies: such as converter failure causing instantaneous current jumps, a line once experienced a non-physical phenomenon where the current jumped from 1200A to 3800A within 3 seconds.

[0005] Trend-type anomalies: Energy consumption data is out of sync with actual operating conditions, such as the energy consumption curve of a workshop between 2 and 4 a.m. deviating from the historical baseline by 230%.

[0006] Noise-related anomalies: These are caused by high-frequency fluctuations due to sensor aging or electromagnetic interference, accounting for 8%-12% of the original data collected, which affects the accuracy of subsequent energy consumption analysis.

[0007] Traditional processing methods have significant limitations.

[0008] Low processing efficiency: When using manual rule configuration and offline batch processing mode, with more than 500 monitoring devices in a certain subway, manually annotating abnormal data takes 4.2 hours / day. The accuracy rate drops from 89% to 62% as the number of devices increases. The average delay in anomaly identification exceeds 45 minutes, resulting in a 32% loss of energy scheduling optimization window.

[0009] Single-dimensional detection has poor reliability: it relies on isolated threshold judgment and ignores the dynamic coupling relationship between subsystems. Experiments show that its recognition rate for composite anomalies is only 41%. It once caused the main substation to trip due to the failure to detect related anomalies, resulting in direct economic losses of several million yuan.

[0010] Low data reconstruction accuracy: When the data missing rate exceeds 8%, the traditional interpolation algorithm has an error of 19.3% for linear interpolation and 15.7% for spline interpolation, resulting in the actual effect of a certain line energy-saving renovation plan being 23% lower than expected.

[0011] The aforementioned problems have led to a "triple failure" in the energy management system: delayed abnormal response resulting in an annual energy waste of approximately 8.7%; a high false alarm rate leading to 30% of invalid maintenance work orders; and data quality defects causing carbon emission calculation errors of ±12%. There is an urgent need to introduce an intelligent algorithm system to solve these problems. Summary of the Invention

[0012] Therefore, the present invention provides an energy anomaly data cleaning method and system based on artificial intelligence to solve the aforementioned problems existing in the prior art.

[0013] To achieve the above objectives, on the one hand, the present invention provides an energy anomaly data cleaning method based on artificial intelligence, comprising:

[0014] Step S1: Obtain initial monitoring data of the energy system, and preprocess the initial monitoring data to obtain the processing result;

[0015] Step S2: Based on the processing results, dynamically adjust the threshold range of the range check according to the data distribution characteristics and equipment physical parameter constraints, and generate preliminary processing data with anomaly markers.

[0016] Step S3: Perform correlation detection on the preliminary processed data, construct the correlation pattern under normal operating conditions through the equipment topology map, and score the abnormal samples that deviate from the pattern to obtain the scoring results.

[0017] Step S4: Perform data cleaning on the scoring results to obtain a preliminary cleaning strategy;

[0018] Step S5: Perform a quality assessment on the cleaning results to obtain an assessment result, and adjust the parameters of the preliminary cleaning strategy according to the assessment result to obtain the target cleaning strategy.

[0019] Furthermore, the process of step S2 includes:

[0020] Calculate the historical data distribution characteristics of the initial monitoring data based on the processing results;

[0021] Based on the constraints of the equipment's physical parameters, the dynamic threshold ranges for each data dimension are determined. The constraints of the equipment's physical parameters include the allowable voltage fluctuation range, the rated current range, and the upper and lower limits of the equipment's operating status.

[0022] Based on the historical data distribution characteristics and equipment physical parameter constraints, the threshold range of the range check is dynamically adjusted.

[0023] Initial monitoring data that exceeds the dynamic threshold range is marked as abnormal data, and preliminary processing data containing abnormal tags is generated.

[0024] Furthermore, the process of dynamically adjusting the threshold range for the range check based on the historical data distribution characteristics and device physical parameter constraints includes:

[0025] The mean, standard deviation, and quantiles of each data dimension are calculated based on the historical data distribution characteristics, and an initial threshold range is determined based on these statistical indicators.

[0026] The initial threshold range is calibrated in conjunction with the constraints of the equipment's physical parameters to ensure that the threshold range conforms to the physical operating characteristics of the equipment;

[0027] The width and position of the threshold range are dynamically adjusted based on the real-time trend of data changes and the real-time operating status of the equipment.

[0028] The dynamically adjusted threshold range is applied to the range check, and the initial monitoring data that exceeds the threshold range is marked to generate preliminary processing data with anomaly markers.

[0029] Furthermore, the process of step S3 includes:

[0030] Based on the preliminary processed data, the correlation features between each data dimension are extracted, and the variable correlation matrix is ​​constructed by calculating the Pearson correlation coefficient.

[0031] Based on the variable correlation matrix and the equipment topology of the energy system, an equipment topology diagram is constructed, where each node represents a device or data acquisition point, and the edges represent the physical connection or data association between devices.

[0032] The graph neural network is used to analyze the topology of the equipment and learn the association patterns between the equipment under normal operating conditions to generate an association pattern model under normal operating conditions.

[0033] The abnormal samples in the preliminary processed data are compared with the association pattern model. By calculating the deviation between the abnormal samples and the normal pattern, the abnormal samples that deviate from the pattern are scored to obtain the scoring result.

[0034] Furthermore, the process of analyzing and learning the association patterns between devices under normal operating conditions using a graph neural network to generate an association pattern model under normal operating conditions includes:

[0035] Feature extraction is performed on each node in the device topology graph. The extracted features include the temporal features of device operating parameters, device status parameters, and interaction features between devices, in order to generate a node feature vector.

[0036] We choose a graph neural network structure and use a graph attention network as the basic architecture to automatically learn the important correlation weights between devices through the attention mechanism.

[0037] The node feature vectors are input into the graph attention network. Through the network's multi-layer propagation mechanism, the dynamic association patterns of each device in the device topology graph under normal operating conditions are learned to generate an association pattern model under normal operating conditions.

[0038] Furthermore, the process of step S4 includes:

[0039] Based on the scoring results, the abnormal samples are sorted according to their score values. Samples with score values ​​higher than a preset threshold are identified as high-confidence abnormal samples, and samples with score values ​​lower than a preset threshold are identified as low-confidence abnormal samples.

[0040] For high-confidence anomalous samples, a generative adversarial network is used to reconstruct the data, generating data that conforms to normal operating conditions to replace the anomalous data and obtain the first repair result.

[0041] For low-confidence abnormal samples, an adaptive interpolation method is used to repair the data by combining the equipment operating status and historical data to obtain a second repair result;

[0042] Based on the first and second repair results, a preliminary cleaning strategy is generated by combining the types of abnormal samples.

[0043] Furthermore, the process of generating a preliminary cleaning strategy based on the first repair result and the second repair result, combined with the type of the abnormal sample, includes:

[0044] Based on the first repair result and the second repair result, the repair quality of high-confidence outlier samples and low-confidence outlier samples is evaluated respectively.

[0045] Based on the type of abnormal samples, abnormal samples are divided into mutation anomalies, trend anomalies, and noise anomalies, and different cleaning rules are formulated for different types of abnormal samples.

[0046] Based on the repair quality and the cleaning rules, a preliminary cleaning strategy is generated.

[0047] Furthermore, the process of step S5 includes:

[0048] Calculate the quality indicators of the cleaned data based on the cleaning results;

[0049] Based on the quality indicators, determine whether the cleaning results meet the preset quality threshold to obtain the judgment result;

[0050] For cleaning results that do not meet the quality threshold, analyze the deviation of the quality indicators, determine the cleaning strategy parameters that need to be adjusted, and obtain a parameter adjustment plan.

[0051] The initial cleaning strategy is optimized and adjusted according to the parameter adjustment scheme to obtain the target cleaning strategy.

[0052] Furthermore, the process of optimizing and adjusting the preliminary cleaning strategy according to the parameter adjustment scheme to obtain the target cleaning strategy includes:

[0053] The training parameters of the generator and discriminator of the generative adversarial network are adjusted according to the parameter adjustment scheme described above.

[0054] For the adaptive interpolation method, the interpolation parameters are dynamically adjusted according to the deviation of the quality index;

[0055] Based on the threshold parameters used in the data cleaning process, adjust the threshold range and sensitivity in range checks and correlation detection according to the deviation of quality indicators.

[0056] The adjusted parameters are applied to the initial cleaning strategy, and the optimized cleaning strategy is regenerated to obtain the target cleaning strategy.

[0057] On the other hand, the present invention also provides an energy anomaly data cleaning system based on artificial intelligence, comprising:

[0058] The data processing module is used to acquire initial monitoring data of the energy system and preprocess the initial monitoring data to obtain the processing result.

[0059] A threshold adjustment module, connected to the data processing module, is used to dynamically adjust the threshold range of the range check based on the processing results, data distribution characteristics, and equipment physical parameter constraints, and generate preliminary processed data containing anomaly markers.

[0060] The pattern analysis module, connected to the threshold adjustment module, is used to perform correlation detection on the preliminary processed data, construct the correlation pattern under normal operating conditions through the equipment topology map, and score abnormal samples that deviate from the pattern to obtain the scoring result.

[0061] The strategy generation module is connected to the pattern analysis module and is used to perform data cleaning on the scoring results to obtain a preliminary cleaning strategy.

[0062] The strategy optimization module, connected to the strategy generation module, is used to perform quality assessment on the cleaning results to obtain assessment results, and adjust the parameters of the initial cleaning strategy according to the assessment results to obtain the target cleaning strategy.

[0063] Compared with existing technologies, the advantages of this invention lie in its ability to achieve intelligent processing of the entire process from data acquisition to cleaning strategy optimization through a series of closely linked steps. After preprocessing, the initial monitoring data is dynamically adjusted in threshold ranges based on data distribution characteristics and equipment physical parameters to accurately identify anomalies. Correlation detection utilizes equipment topology maps to construct correlation patterns under normal operating conditions, scoring abnormal samples and providing a basis for subsequent cleaning. Data cleaning generates a preliminary strategy, while quality assessment judges the cleaning effect by calculating the integrity, accuracy, and consistency indicators of the cleaned data. Parameters are adjusted based on the assessment results to optimize the cleaning strategy, forming a closed-loop feedback mechanism. This entire method not only significantly improves the efficiency and quality of data cleaning but also effectively reduces operation and maintenance costs and reliance on manual labor, enhancing the system's adaptability and scalability. The dynamic adjustment and optimization of parameters between steps, based on real-time feedback from data characteristics and equipment operating status, ensures that the cleaning strategy always aligns with actual needs, thereby achieving efficient and accurate data cleaning and providing solid data support for the stable operation and optimized management of energy systems.

[0064] In particular, by dynamically adjusting the threshold range in real time, the system can adapt to seasonal fluctuations in data, equipment aging, and special operating conditions, avoiding misjudgments or missed detections caused by fixed thresholds. Combining historical data distribution characteristics with equipment physical parameter constraints ensures that the threshold setting conforms to both data statistical characteristics and the physical requirements of equipment operation, improving the accuracy of anomaly detection. The dynamic adjustment mechanism can flexibly adjust the threshold range based on seasonal fluctuations and equipment aging, enhancing the system's adaptability and robustness.

[0065] In particular, by combining the temporal characteristics of equipment topology and operating parameters, a dynamic correlation pattern model is generated, which can adapt to changes in the operating state of the energy system under different operating conditions, improving the model's robustness and adaptability. By calculating and scoring the deviation between abnormal samples and normal patterns, abnormal samples can be sorted and located, providing maintenance personnel with clear directions for anomaly investigation and significantly shortening fault diagnosis time. Furthermore, by integrating equipment operating parameters, status parameters, and interaction features, node feature vectors are generated, enriching the model's input information and improving the accuracy of anomaly detection.

[0066] In particular, data reconstruction using generative adversarial networks (GANs) can generate high-quality normal operating condition data to replace abnormal data, ensuring data integrity and accuracy. Adaptive interpolation repair, combining equipment operating status and historical data, can effectively handle missing data and low-confidence anomalies, reducing false alarms and missed alarms. Selecting appropriate interpolation methods based on the contextual information of abnormal samples improves the flexibility and accuracy of data repair.

[0067] In particular, through quality assessment and parameter adjustment, the cleaned data is ensured to meet preset quality thresholds, significantly improving data integrity and accuracy. Based on deviations in quality indicators, the parameters of the cleaning strategy are dynamically adjusted to ensure its adaptability and optimization. Adjusting the parameters of range checks and correlation detection enhances the flexibility of the cleaning strategy, enabling it to better adapt to data from different equipment and operating conditions. Optimizing the cleaning strategy reduces false alarms and missed alarms caused by data anomalies, improving system reliability and stability. Attached Figure Description

[0068] Figure 1 A flowchart illustrating the energy anomaly data cleaning method based on artificial intelligence provided by this invention;

[0069] Figure 2 This is a flowchart illustrating step S2 in the AI-based energy anomaly data cleaning method provided by the present invention.

[0070] Figure 3 This is a flowchart illustrating step S3 in the AI-based energy anomaly data cleaning method provided by the present invention.

[0071] Figure 4 A schematic diagram of the structure of the energy anomaly data cleaning method based on artificial intelligence provided by the present invention. Detailed Implementation

[0072] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0073] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0074] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0075] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0076] Please see Figure 1 As shown, this invention provides an artificial intelligence-based method for cleaning energy anomaly data, comprising:

[0077] Step S1: Obtain initial monitoring data of the energy system, and preprocess the initial monitoring data to obtain the processing result;

[0078] Specifically, initial monitoring data is acquired from the energy system. This data includes, but is not limited to, key parameters such as voltage, current, power, and equipment status. The data acquisition frequency is set according to the equipment type and application scenario; for example, the sampling frequency for traction substations is 1Hz-10Hz, while for ordinary equipment such as station lighting, the sampling frequency is 15 minutes / time. 5G time synchronization technology (synchronization accuracy ±1ms) is used to unify data from different sources (such as traction substations, overhead contact line monitoring points, and train ATP data) onto the same timeline, ensuring data consistency. The acquired data undergoes an integrity check. Missing data due to signal interruption or sensor failure is marked with "NaN," covering all key energy parameters (such as voltage, current, and power). Based on the 3σ principle (calculating standard deviation using a 24-hour sliding window), single-dimensional data is initially screened, and data exceeding the normal range is marked as outliers. The device parameter database is called to verify whether the data is within the physical measurement range. For example, voltage data is checked against allowable fluctuations (such as ±5% for underground line voltage fluctuations); data exceeding this range is marked as "RANGEERR". By combining SCADA system alarm records (such as voltage sag events on a subway line), known abnormal data is labeled to build a training sample library for subsequent model training and validation. Basic features are constructed, including time features (such as hourly segments and weekday / weekend identifiers), energy consumption features (such as real-time power and energy consumption change rate), and operational features (such as passenger flow and train operation density). Time-frequency features of voltage sags are extracted using wavelet transform to generate high-dimensional feature vectors.

[0079] Step S2: Based on the processing results, dynamically adjust the threshold range of the range check according to the data distribution characteristics and equipment physical parameter constraints, and generate preliminary processing data with anomaly markers.

[0080] Specifically, such as Figure 2As shown, the process of step S2 includes:

[0081] Step S21: Calculate the historical data distribution characteristics of the initial monitoring data based on the processing results;

[0082] Specifically, based on the processing results, statistical analysis is performed on the initial monitoring data to calculate the historical data distribution characteristics for each data dimension, including statistical indicators such as mean, standard deviation, and quantiles. The calculation scope covers 23 categories of key energy parameters, including but not limited to:

[0083] Voltage-related: contact wire end voltage (DC 1000~1800V), 110kV incoming line voltage deviation (±10%), 35kV bus voltage fluctuation (≤5% of rated value).

[0084] Current-related: traction converter output current (0~3000A), rail return current imbalance (>5%).

[0085] Equipment status categories: rectifier transformer oil temperature (≤130℃ alarm), GIS equipment SF6 gas pressure (0.4~0.6MPa), cable joint temperature (>90℃ warning).

[0086] Power quality: Total harmonic distortion of current (THDi≤8%), power factor (≥0.9).

[0087] Step S22: Combine the physical parameter constraints of the equipment to determine the dynamic threshold range of each data dimension. The physical parameter constraints of the equipment include the allowable voltage fluctuation range, the rated current range, and the upper and lower limits of the equipment operating status.

[0088] Specifically, the rated operating parameters and physical constraints of each device are extracted from the device parameter database, including: Permissible voltage fluctuation range: ±5% for underground lines (corresponding to 1425V-1575V), ±8% for ground / elevated lines. Rated current range: e.g., traction converter rated current 1500A, permissible short-term inrush current up to 300% of rated value (lasting 1 minute). Upper limit of equipment temperature: e.g., transformer oil temperature alarm threshold 130℃, warning threshold 90℃. Gas pressure range: e.g., SF6 gas pressure 0.4~0.6MPa for GIS equipment. Based on the physical parameter constraints of the equipment, initial threshold ranges are set for each data dimension: Voltage threshold: ±5% or ±8% of rated value. Current threshold: ±20% of rated value (based on 95% confidence interval of actual subway measurement data). Temperature threshold: 80% of the equipment alarm temperature is used as the warning value.

[0089] Step S23: Dynamically adjust the threshold range of the range check based on the distribution characteristics of the historical data and the constraints of the equipment physical parameters;

[0090] Specifically, the process of dynamically adjusting the threshold range of the range check based on the historical data distribution characteristics and device physical parameter constraints includes:

[0091] The mean, standard deviation, and quantiles of each data dimension are calculated based on the historical data distribution characteristics, and an initial threshold range is determined based on these statistical indicators.

[0092] Specifically, based on the 3σ principle, the initial threshold interval is set to [μ-3σ, μ+3σ]. If the data distribution is skewed, the quantile method is used, and the threshold interval is set to [Q5%, Q95%].

[0093] The initial threshold range is calibrated in conjunction with the constraints of the equipment's physical parameters to ensure that the threshold range conforms to the physical operating characteristics of the equipment;

[0094] Specifically, if the calculated threshold range exceeds the physical range allowed by the device, it will be limited to the physical range. For example, if the voltage threshold range is calculated to be ±15%, but the device's physical constraint is ±10%, it will be adjusted to ±10%. If the calculated threshold range is too strict (e.g., ±2%), causing normal data to be misjudged as abnormal, it will be relaxed to a reasonable range of the device's physical constraints (e.g., ±5%).

[0095] The width and position of the threshold range are dynamically adjusted based on the real-time trend of data changes and the real-time operating status of the equipment.

[0096] Specifically, for data exhibiting seasonal fluctuations (such as summer air conditioning load), the threshold range is adjusted based on historical data from the same period. For example, the summer voltage fluctuation threshold can be relaxed from ±5% to ±6%.

[0097] Equipment Aging Adjustment: To address performance degradation caused by equipment aging, thresholds are dynamically adjusted based on the equipment's operating years. For example, after a transformer has been in operation for more than 10 years, the temperature threshold is adjusted from the 85% alarm value to 80%. Under special operating conditions (such as morning and evening rush hours on rail transit), the dynamic threshold range is automatically expanded. For example, during the morning rush hour (7:00-9:00), the voltage threshold is widened from ±10% of the rated value to ±12%, and the GNN network's recognition weight for instantaneous impact loads is enhanced. When the false detection rate exceeds 5% or the missed detection rate exceeds 3%, the system automatically triggers the threshold adjustment mechanism.

[0098] The dynamically adjusted threshold range is applied to the range check, and the initial monitoring data that exceeds the threshold range is marked to generate preliminary processing data with anomaly markers.

[0099] Step S24: Mark the initial monitoring data that exceeds the dynamic threshold range as abnormal data and generate preliminary processing data containing abnormal markings.

[0100] Specifically, data exceeding the physical range of the device is marked as "RANGEERR". Data exceeding the dynamic threshold but not exceeding the range is marked as "ABNORMAL". Known abnormal events (such as SCADA system alarm records) are marked as "CONFIRMED_ABN". By adjusting the dynamic threshold, the proportion of abnormal data is controlled between 5% and 15%, achieving data slimming and reducing the computational load of subsequent modules. A preliminary processing dataset containing anomaly tags is generated. Each data record includes: timestamp, parameter value, anomaly tag, threshold range, and device ID.

[0101] Specifically, by dynamically adjusting the threshold range in real time, the system can adapt to seasonal fluctuations in data, equipment aging, and special operating conditions, avoiding false positives or false negatives caused by fixed thresholds. By combining historical data distribution characteristics with equipment physical parameter constraints, the system ensures that the threshold setting conforms to both data statistical characteristics and the physical requirements of equipment operation, thus improving the accuracy of anomaly detection. The dynamic adjustment mechanism can flexibly adjust the threshold range based on seasonal fluctuations and equipment aging, enhancing the system's adaptability and robustness.

[0102] Step S3: Perform correlation detection on the preliminary processed data, construct the correlation pattern under normal operating conditions through the equipment topology map, and score the abnormal samples that deviate from the pattern to obtain the scoring results.

[0103] Specifically, such as Figure 3 As shown, the process of step S3 includes:

[0104] Step S31: Based on the preliminary processed data, extract the correlation features between each data dimension, and construct the variable correlation matrix by calculating the Pearson correlation coefficient;

[0105] Specifically, key data dimensions are selected from the initially processed data, including but not limited to parameters such as voltage, current, power, and equipment status. These parameters reflect the operating status of the energy system. A Pearson correlation coefficient is calculated for each pair of data dimensions to measure the linear correlation between the two variables. The correlation coefficient ranges from -1 to 1, where 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no correlation. The calculated correlation coefficients are then organized into a matrix to form a variable association matrix. The rows and columns of the matrix correspond to different data dimensions, and each element in the matrix represents the correlation between the corresponding dimensions.

[0106] Step S32: Based on the variable correlation matrix and combined with the equipment topology of the energy system, construct an equipment topology diagram, where each node represents a device or data acquisition point, and the edges represent physical connections or data associations between devices.

[0107] Specifically, each node represents a device or data acquisition point. Node attributes include device type (e.g., transformer, converter, sensor), device location (e.g., line number, site location), and device operating parameters (e.g., voltage, current, power). Edges represent physical connections or data associations between devices. The connection relationships between nodes are determined based on the actual connections between devices (e.g., power line connections, data transmission links) and the correlation in the variable correlation matrix. Strongly correlated devices are connected by edges, and the edge weights can be determined by the correlation coefficient. All nodes and edges are combined into a graph structure, forming a device topology graph. This reflects the physical and data associations between devices in the energy system.

[0108] Step S33: Analyze the device topology graph using a graph neural network to learn the association patterns between devices under normal operating conditions in order to generate an association pattern model under normal operating conditions.

[0109] Specifically, the process of analyzing and learning the association patterns between devices under normal operating conditions using a graph neural network to generate an association pattern model under normal operating conditions includes:

[0110] Feature extraction is performed on each node in the device topology graph. The extracted features include the temporal features of device operating parameters, device status parameters, and interaction features between devices, in order to generate a node feature vector.

[0111] Specifically, historical operating parameters for each device are extracted from the initial processed data, including key parameters such as voltage, current, and power. A fixed-length sliding window (e.g., 30 minutes, 1 hour) is used to calculate statistical indicators such as the mean, standard deviation, maximum, and minimum values ​​of the parameters within the window. Short-term rates of change and long-term trends of the parameters are calculated; for example, the 5-minute rate of change and 24-hour trend of voltage. Periodic features of the parameters, such as daily and weekly cycles, are extracted using Fourier transform or wavelet transform to extract periodic components. The above statistical indicators, rates of change, and periodic features are combined into a time-series feature vector. Operating status parameters of the devices are extracted, such as temperature, pressure, device operating time, and alarm status. These status parameters are standardized or normalized to generate a status feature vector. For example, temperature parameters can be standardized using Z-score. Based on the device topology, the physical connections between devices are identified, such as current transmission, power distribution, and signal interaction. The amount and direction of current transmission between devices are calculated. The power distribution ratio between devices is calculated. The signal transmission delay and frequency between devices are calculated. Interaction features are standardized to generate an interaction feature vector. Temporal features, state features, and interaction features are fused to generate a comprehensive feature vector for each node. Depending on the model requirements, the feature vectors are either reduced or increased in dimensionality to ensure that the feature vectors input to the graph neural network have consistent dimensions.

[0112] We choose a graph neural network structure and use a graph attention network as the basic architecture to automatically learn the important correlation weights between devices through the attention mechanism.

[0113] Specifically, a Graph Attention Network (GAT) is used as the basic architecture. GAT automatically learns the important association weights between devices through an attention mechanism, making it suitable for handling complex relationships in the device topology graph. A multi-layer (e.g., 2- or 3-layer) GAT network is designed, with each layer responsible for capturing device relationships at a different level. Within each GAT layer, the association weights between device nodes are calculated using an attention mechanism. These attention weights represent the importance of devices at the current layer. A multi-head attention mechanism is employed to learn the relationships between devices from multiple perspectives, improving the model's expressive power.

[0114] The node feature vectors are input into the graph attention network. Through the network's multi-layer propagation mechanism, the dynamic association patterns of each device in the device topology graph under normal operating conditions are learned to generate an association pattern model under normal operating conditions.

[0115] Specifically, the feature vector of each node is input into the first layer of the GAT network. The network weights are initialized using random initialization or pre-trained weights. In the first layer of GAT, the association weights between nodes are calculated through an attention mechanism, updating the node's feature representation. In subsequent layers, the updated feature vectors are propagated to further learn the complex relationships between devices. In each layer, the feature information of neighboring nodes is aggregated, updating the feature representation of the current node. Through multi-layer propagation, the dynamic association patterns of devices in the device topology graph under normal operating conditions are learned. The model is trained using data from normal operating conditions, optimizing the model parameters to ensure that the model can accurately describe the normal operating relationships between devices. After training, a relationship pattern model under normal operating conditions is generated, which reflects the normal operating relationships between devices.

[0116] Step S34: Compare the abnormal samples in the preliminary processed data with the association pattern model, and score the abnormal samples that deviate from the pattern by calculating the deviation between the abnormal samples and the normal pattern to obtain the score result.

[0117] Specifically, samples marked as anomalous are extracted from the initially processed data; these samples were already marked as exceeding the dynamic threshold range in step S2. The anomalous samples are compared with the association pattern model of normal operating conditions to calculate the deviation between the anomalous samples and the normal pattern. The deviation can be calculated in several ways, such as: calculating the Euclidean distance between the anomalous samples and the normal pattern in the feature space; calculating the cosine similarity between the anomalous samples and the normal pattern to measure their directional difference; or considering the covariance structure of the data to calculate the Mahalanobis distance between the anomalous samples and the normal pattern. Based on the calculated deviation, the anomalous samples are scored. The score reflects the degree to which the anomalous samples deviate from the normal pattern; a higher score indicates a greater degree of deviation.

[0118] Specifically, by combining the temporal characteristics of equipment topology and operating parameters, a dynamic correlation pattern model is generated, which can adapt to changes in the operating state of the energy system under different operating conditions, improving the model's robustness and adaptability. By calculating and scoring the deviation between abnormal samples and normal patterns, abnormal samples can be sorted and located, providing maintenance personnel with clear directions for anomaly investigation and significantly shortening fault diagnosis time. Furthermore, by integrating equipment operating parameters, status parameters, and interaction features, node feature vectors are generated, enriching the model's input information and improving the accuracy of anomaly detection.

[0119] Step S4: Perform data cleaning on the scoring results to obtain a preliminary cleaning strategy;

[0120] Specifically, step S4 includes the following process:

[0121] Based on the scoring results, the abnormal samples are sorted according to their score values. Samples with score values ​​higher than a preset threshold are identified as high-confidence abnormal samples, and samples with score values ​​lower than a preset threshold are identified as low-confidence abnormal samples.

[0122] Specifically, outlier samples are sorted according to their score. The score reflects the degree to which a sample deviates from normal operating conditions. A preset threshold is used to identify samples with scores higher than the threshold as high-confidence outliers, and samples with scores lower than the threshold as low-confidence outliers. For example, the threshold can be set to 0.85, meaning that samples with scores greater than 0.85 are high-confidence outliers, and those less than 0.85 are low-confidence outliers.

[0123] For high-confidence anomalous samples, a generative adversarial network is used to reconstruct the data, generating data that conforms to normal operating conditions to replace the anomalous data and obtain the first repair result.

[0124] Specifically, high-confidence outlier samples are normalized to meet the input requirements of Generative Adversarial Networks (GANs). An improved GAN is used, where the generator incorporates prior knowledge of energy consumption characteristics (such as daily cycle patterns and weekly regularity factors) to guide the generated data to fit the business scenario. The generator learns the data distribution under normal operating conditions to generate data that conforms to normal operating conditions. A multi-scale discriminator is designed and trained adversarially from three levels: overall distribution, local temporal fluctuations, and extreme value probability distribution. The discriminator is used to distinguish between generated data and real data to ensure the quality of the generated data. The generated normal operating condition data replaces the original high-confidence outlier data to obtain the first repair result.

[0125] For low-confidence abnormal samples, an adaptive interpolation method is used to repair the data by combining the equipment operating status and historical data to obtain a second repair result;

[0126] Specifically, the contextual information of low-confidence outlier samples is analyzed by combining equipment operating status (such as load rate, uptime, etc.) and historical data. An appropriate interpolation method is then selected based on this contextual information. For example, cubic spline interpolation is used for short-term missing data (<5 sampling points); while LSTM interpolation based on an attention mechanism is used for long-term missing data (≥5 sampling points).

[0127] Based on the first and second repair results, a preliminary cleaning strategy is generated by combining the types of abnormal samples.

[0128] Specifically, the process of generating a preliminary cleaning strategy based on the first repair result and the second repair result, combined with the type of the abnormal sample, includes:

[0129] Based on the first repair result and the second repair result, the repair quality of high-confidence outlier samples and low-confidence outlier samples is evaluated respectively.

[0130] Specifically, the similarity (e.g., JS divergence) between the first repair result and the real data is calculated to evaluate the quality of the generated data. For example, a JS divergence of 0.92 indicates that the generated data is highly similar to the distribution of the real data. The repair error (e.g., MAE) of the second repair result is calculated to evaluate the quality of the interpolation repair. For example, an MAE controlled within 15% of the standard deviation of the original data indicates good repair quality.

[0131] Based on the type of abnormal samples, abnormal samples are divided into mutation anomalies, trend anomalies, and noise anomalies, and different cleaning rules are formulated for different types of abnormal samples.

[0132] Specifically, abrupt anomalies refer to significant changes in data values ​​within a short period, typically caused by equipment failure, sensor malfunction, or other sudden events. For example, a current jump from 1200A to 3800A within 3 seconds. Abrupt anomalies are identified by calculating the difference or rate of change between adjacent data points. If the difference or rate of change exceeds a preset threshold (e.g., a current change rate exceeding 300%), the data point is considered an abrupt anomaly. Trend anomalies refer to data values ​​that consistently deviate from the normal trend over a period of time, typically caused by equipment aging, changes in operating conditions, or other long-term factors. For example, the energy consumption curve of a workshop between 2-4 AM deviates from the historical baseline by 230%. Trend anomalies are identified by calculating the deviation between the data value and the historical baseline or predicted trend. If the deviation exceeds a preset threshold (e.g., an energy consumption deviation exceeding 20%), the data point is considered a trend anomaly. Noise anomalies refer to the presence of high-frequency fluctuations or random noise in the data, typically caused by sensor aging, electromagnetic interference, or other environmental factors. For example, high-frequency fluctuations caused by sensor aging account for 8%-12% of the original data. Noise-type anomalies are identified by calculating the variance or standard deviation of the data. If the variance or standard deviation exceeds a preset threshold (e.g., standard deviation exceeds 10%), the data point is considered a noise-type anomaly. Abrupt change anomaly cleaning rules: For abrupt change anomalies, a sliding window smoothing method is used. For example, a sliding window of 3 data points is used, and the average value of the data points within the window is taken as the new data point value. Trend anomaly cleaning rules: For trend anomalies, trend analysis and adjustment methods are used. For example, a trend line is fitted to the data using a linear regression model, and data points deviating from the trend line are adjusted to the trend line. Noise-type anomaly cleaning rules: For noisy anomalies, filtering methods are used. For example, median filtering or wavelet transform is used to suppress high-frequency noise.

[0133] Based on the repair quality and the cleaning rules, a preliminary cleaning strategy is generated.

[0134] Specifically, data reconstruction using generative adversarial networks (GANs) can generate high-quality normal operating condition data to replace abnormal data, ensuring data integrity and accuracy. Adaptive interpolation repair, combining equipment operating status and historical data, can effectively handle missing data and low-confidence anomalies, reducing false alarms and missed alarms. Selecting appropriate interpolation methods based on the contextual information of abnormal samples improves the flexibility and accuracy of data repair.

[0135] Step S5: Perform a quality assessment on the cleaning results to obtain an assessment result, and adjust the parameters of the preliminary cleaning strategy according to the assessment result to obtain the target cleaning strategy.

[0136] Specifically, step S5 includes the following process:

[0137] Calculate the quality indicators of the cleaned data based on the cleaning results;

[0138] Specifically, calculate the missing rate of the cleaned data to ensure data integrity. For example, the missing rate should be controlled below 0.5%. Calculate the similarity (e.g., JS divergence) and repair error (e.g., MAE) between the cleaned data and the original data. For example, the JS divergence should reach above 0.92, and the MAE should be controlled within 15% of the standard deviation of the original data. Check whether the cleaned data is consistent with historical data and equipment operating status. For example, ensure the consistency of the data over time by calculating statistical indicators such as the mean and variance of the data.

[0139] Based on the quality indicators, determine whether the cleaning results meet the preset quality threshold to obtain the judgment result;

[0140] Specifically, data quality thresholds are set, such as a missing rate threshold of 0.5%, a JS divergence threshold of 0.92, and a MAE threshold of 15% of the original data standard deviation. Based on the calculated quality indicators, it is determined whether the cleaning results meet the preset thresholds. If all indicators meet the thresholds, the cleaning results are considered acceptable; otherwise, proceed to the next step of analysis.

[0141] For cleaning results that do not meet the quality threshold, analyze the deviation of the quality indicators, determine the cleaning strategy parameters that need to be adjusted, and obtain a parameter adjustment plan.

[0142] Specifically, a deviation analysis is performed on quality indicators that do not meet the thresholds to determine the specific nature of the deviation. For example, if the JS divergence is below 0.92, the distribution difference between the generated data and the real data is analyzed; if the MAE exceeds 15%, the error sources in the repaired data are analyzed. Based on the deviation, the parameters of the cleaning strategy that need to be adjusted are determined. For example, if the JS divergence deviation is large, the training parameters of the generator and discriminator of the generative adversarial network need to be adjusted; if the MAE deviation is large, the interpolation parameters of the adaptive interpolation method may need to be adjusted.

[0143] The initial cleaning strategy is optimized and adjusted according to the parameter adjustment scheme to obtain the target cleaning strategy.

[0144] Specifically, the process of optimizing and adjusting the preliminary cleaning strategy according to the parameter adjustment scheme to obtain the target cleaning strategy includes:

[0145] The training parameters of the generator and discriminator of the generative adversarial network are adjusted according to the parameter adjustment scheme described above.

[0146] Specifically, the generator's training parameters, such as the learning rate, the number of hidden layer neurons, and the number of training epochs, are adjusted based on the deviation of the JS divergence. For example, if the JS divergence is below 0.92, the number of training epochs can be increased or the learning rate adjusted. The discriminator's training parameters, such as the weights of the multi-scale discriminator and the number of discriminator layers, are adjusted based on the quality of the generated data. For example, if the local temporal fluctuations in the generated data are inaccurate, the weights of the multi-scale discriminator can be adjusted.

[0147] For the adaptive interpolation method, the interpolation parameters are dynamically adjusted according to the deviation of the quality index;

[0148] Specifically, interpolation parameters, such as interpolation window size and interpolation weights, are dynamically adjusted based on the deviation of MAE. For example, if MAE exceeds 15%, the interpolation window size can be adjusted to better adapt to the temporal characteristics of the data.

[0149] Based on the threshold parameters used in the data cleaning process, adjust the threshold range and sensitivity in range checks and correlation detection according to the deviation of quality indicators.

[0150] Specifically, the threshold range and sensitivity in range checks are adjusted based on the data distribution characteristics. For example, if the data fluctuates significantly, the threshold range can be appropriately widened. The sensitivity parameters in correlation detection are adjusted based on equipment operating status and historical data. For example, if equipment operating status changes frequently, the sensitivity of correlation detection can be increased.

[0151] The adjusted parameters are applied to the initial cleaning strategy, and the optimized cleaning strategy is regenerated to obtain the target cleaning strategy.

[0152] Specifically, the adjusted parameters are applied to the initial cleaning strategy to regenerate the optimized cleaning strategy. The optimized cleaning strategy is then validated to ensure it meets preset quality thresholds. For example, the quality metrics of the cleaned data are recalculated to ensure the missing rate is below 0.5%, the JS divergence is above 0.92, and the MAE is within 15% of the original data standard deviation.

[0153] Specifically, through quality assessment and parameter adjustment, the cleaned data is ensured to meet preset quality thresholds, significantly improving data integrity and accuracy. Based on deviations in quality indicators, the parameters of the cleaning strategy are dynamically adjusted to ensure its adaptability and optimization. Adjusting the parameters of range checks and correlation detection enhances the flexibility of the cleaning strategy, enabling it to better adapt to data from different equipment and operating conditions. Optimizing the cleaning strategy reduces false alarms and missed alarms caused by data anomalies, improving system reliability and stability.

[0154] Specifically, such as Figure 4As shown, the present invention also provides an energy anomaly data cleaning system based on artificial intelligence, comprising:

[0155] Data processing module 10 is used to acquire initial monitoring data of the energy system and preprocess the initial monitoring data to obtain processing results;

[0156] The threshold adjustment module 20 is connected to the data processing module 10 and is used to dynamically adjust the threshold range of the range check based on the processing results, data distribution characteristics, and equipment physical parameter constraints, and generate preliminary processed data with anomaly markers.

[0157] The pattern analysis module 30 is connected to the threshold adjustment module 20 and is used to perform correlation detection on the preliminary processed data, construct the correlation pattern under normal working conditions through the equipment topology map, and score abnormal samples that deviate from the pattern to obtain the scoring result.

[0158] The strategy generation module 40 is connected to the pattern analysis module 30 and is used to perform data cleaning on the scoring results to obtain a preliminary cleaning strategy.

[0159] The strategy optimization module 50 is connected to the strategy generation module 40 and is used to perform quality evaluation on the cleaning results to obtain evaluation results, and adjust the parameters of the preliminary cleaning strategy according to the evaluation results to obtain the target cleaning strategy.

[0160] Specifically, the AI-based energy anomaly data cleaning system provided by the present invention can execute the AI-based energy anomaly data cleaning method in the embodiments of the present invention and achieve the same technical effect, which will not be elaborated here.

[0161] Specifically, this invention achieves intelligent processing of the entire process from data acquisition to cleaning strategy optimization through a series of closely linked steps. Initial monitoring data, after preprocessing, dynamically adjusts threshold ranges based on data distribution characteristics and equipment physical parameters to accurately identify anomalies. Correlation detection utilizes equipment topology maps to construct correlation patterns for normal operating conditions, scoring abnormal samples and providing a basis for subsequent cleaning. Data cleaning generates a preliminary strategy, while quality assessment judges the cleaning effect by calculating the completeness, accuracy, and consistency indicators of the cleaned data. Parameters are adjusted based on the assessment results to optimize the cleaning strategy, forming a closed-loop feedback mechanism. This entire method not only significantly improves the efficiency and quality of data cleaning but also effectively reduces operation and maintenance costs and reliance on manual labor, enhancing the system's adaptability and scalability. The dynamic adjustment and optimization of parameters between steps, based on real-time feedback from data characteristics and equipment operating status, ensures that the cleaning strategy always aligns with actual needs, thereby achieving efficient and accurate data cleaning and providing solid data support for the stable operation and optimized management of energy systems.

[0162] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0163] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for cleaning energy anomaly data based on artificial intelligence, characterized in that, include: Step S1: Obtain initial monitoring data of the energy system, and preprocess the initial monitoring data to obtain the processing result; Step S2: Based on the processing results, dynamically adjust the threshold range of the range check according to the data distribution characteristics and equipment physical parameter constraints, and generate preliminary processing data with anomaly markers. Step S3: Perform correlation detection on the preliminary processed data, construct the correlation pattern under normal operating conditions through the equipment topology map, and score the abnormal samples that deviate from the pattern to obtain the scoring results. Step S4: Perform data cleaning on the scoring results to obtain a preliminary cleaning strategy; Step S5: Perform a quality assessment on the cleaning results to obtain an assessment result, and adjust the parameters of the preliminary cleaning strategy according to the assessment result to obtain the target cleaning strategy.

2. The energy anomaly data cleaning method based on artificial intelligence according to claim 1, characterized in that, The process of step S2 includes: Calculate the historical data distribution characteristics of the initial monitoring data based on the processing results; Based on the constraints of the equipment's physical parameters, the dynamic threshold ranges for each data dimension are determined. The constraints of the equipment's physical parameters include the allowable voltage fluctuation range, the rated current range, and the upper and lower limits of the equipment's operating status. Based on the historical data distribution characteristics and equipment physical parameter constraints, the threshold range of the range check is dynamically adjusted. Initial monitoring data that exceeds the dynamic threshold range is marked as abnormal data, and preliminary processing data containing abnormal tags is generated.

3. The energy anomaly data cleaning method based on artificial intelligence according to claim 2, characterized in that, The process of dynamically adjusting the threshold range for range checking based on the historical data distribution characteristics and equipment physical parameter constraints includes: The mean, standard deviation, and quantiles of each data dimension are calculated based on the historical data distribution characteristics, and an initial threshold range is determined based on these statistical indicators. The initial threshold range is calibrated in conjunction with the constraints of the equipment's physical parameters to ensure that the threshold range conforms to the physical operating characteristics of the equipment; The width and position of the threshold range are dynamically adjusted based on the real-time trend of data changes and the real-time operating status of the equipment. The dynamically adjusted threshold range is applied to the range check, and the initial monitoring data that exceeds the threshold range is marked to generate preliminary processing data with anomaly markers.

4. The energy anomaly data cleaning method based on artificial intelligence according to claim 3, characterized in that, The process of step S3 includes: Based on the preliminary processed data, the correlation features between each data dimension are extracted, and the variable correlation matrix is ​​constructed by calculating the Pearson correlation coefficient. Based on the variable correlation matrix and combined with the equipment topology of the energy system, an equipment topology diagram is constructed, where each node represents a device or data acquisition point, and the edges represent the physical connection or data association relationship between devices. The graph neural network is used to analyze the topology of the equipment and learn the association patterns between the equipment under normal operating conditions to generate an association pattern model under normal operating conditions. The abnormal samples in the preliminary processed data are compared with the association pattern model. By calculating the deviation between the abnormal samples and the normal pattern, the abnormal samples that deviate from the pattern are scored to obtain the scoring result.

5. The energy anomaly data cleaning method based on artificial intelligence according to claim 4, characterized in that, The process of analyzing and learning the association patterns between devices under normal operating conditions using a graph neural network to generate an association pattern model under normal operating conditions includes: Feature extraction is performed on each node in the device topology graph. The extracted features include the temporal features of device operating parameters, device status parameters, and interaction features between devices, in order to generate a node feature vector. We choose a graph neural network structure and use a graph attention network as the basic architecture to automatically learn the important correlation weights between devices through the attention mechanism. The node feature vectors are input into the graph attention network. Through the network's multi-layer propagation mechanism, the dynamic association patterns of each device in the device topology graph under normal operating conditions are learned to generate an association pattern model under normal operating conditions.

6. The energy anomaly data cleaning method based on artificial intelligence according to claim 5, characterized in that, The process of step S4 includes: Based on the scoring results, the abnormal samples are sorted according to their score values. Samples with score values ​​higher than a preset threshold are identified as high-confidence abnormal samples, and samples with score values ​​lower than a preset threshold are identified as low-confidence abnormal samples. For high-confidence anomalous samples, a generative adversarial network is used to reconstruct the data, generating data that conforms to normal operating conditions to replace the anomalous data and obtain the first repair result. For low-confidence abnormal samples, an adaptive interpolation method is used to repair the data by combining the equipment operating status and historical data to obtain a second repair result; Based on the first and second repair results, a preliminary cleaning strategy is generated by combining the types of abnormal samples.

7. The energy anomaly data cleaning method based on artificial intelligence according to claim 6, characterized in that, The process of generating a preliminary cleaning strategy based on the first repair result and the second repair result, combined with the type of abnormal sample, includes: Based on the first repair result and the second repair result, the repair quality of high-confidence outlier samples and low-confidence outlier samples is evaluated respectively. Based on the type of abnormal samples, abnormal samples are divided into mutation anomalies, trend anomalies, and noise anomalies, and different cleaning rules are formulated for different types of abnormal samples. Based on the repair quality and the cleaning rules, a preliminary cleaning strategy is generated.

8. The energy anomaly data cleaning method based on artificial intelligence according to claim 7, characterized in that, The process of step S5 includes: Calculate the quality indicators of the cleaned data based on the cleaning results; Based on the quality indicators, determine whether the cleaning results meet the preset quality threshold to obtain the judgment result; For cleaning results that do not meet the quality threshold, analyze the deviation of the quality indicators, determine the cleaning strategy parameters that need to be adjusted, and obtain a parameter adjustment plan. The initial cleaning strategy is optimized and adjusted according to the parameter adjustment scheme to obtain the target cleaning strategy.

9. The energy anomaly data cleaning method based on artificial intelligence according to claim 8, characterized in that, The process of optimizing and adjusting the preliminary cleaning strategy according to the parameter adjustment scheme to obtain the target cleaning strategy includes: The training parameters of the generator and discriminator of the generative adversarial network are adjusted according to the parameter adjustment scheme described above. For the adaptive interpolation method, the interpolation parameters are dynamically adjusted according to the deviation of the quality index; Based on the threshold parameters used in the data cleaning process, adjust the threshold range and sensitivity in range checks and correlation detection according to the deviation of quality indicators. The adjusted parameters are applied to the initial cleaning strategy, and the optimized cleaning strategy is regenerated to obtain the target cleaning strategy.

10. An AI-based energy anomaly data cleaning system based on the AI-based energy anomaly data cleaning method according to any one of claims 1-9, characterized in that, include: The data processing module is used to acquire initial monitoring data of the energy system and preprocess the initial monitoring data to obtain the processing result. A threshold adjustment module, connected to the data processing module, is used to dynamically adjust the threshold range of the range check based on the processing results, data distribution characteristics, and equipment physical parameter constraints, and generate preliminary processed data containing anomaly markers. The pattern analysis module, connected to the threshold adjustment module, is used to perform correlation detection on the preliminary processed data, construct the correlation pattern under normal operating conditions through the equipment topology map, and score abnormal samples that deviate from the pattern to obtain the scoring result. The strategy generation module is connected to the pattern analysis module and is used to perform data cleaning on the scoring results to obtain a preliminary cleaning strategy. The strategy optimization module, connected to the strategy generation module, is used to perform quality assessment on the cleaning results to obtain assessment results, and adjust the parameters of the initial cleaning strategy according to the assessment results to obtain the target cleaning strategy.

Images