A method and system for tracing abnormal electricity marketing data based on multi-scale convolution.

Abstract

A power marketing abnormal data tracing method and system based on multi-scale convolution. The method comprises the following steps: according to a curve subset labeled by a preliminary high-response area, processing a periodic fluctuation feature under a week period scale by using a 7*7 convolution kernel to determine an expanded abnormal point set; based on the comprehensive key abnormal point set, grouping mutation points and periodic fluctuation points in the high-response area by using a clustering algorithm to obtain a hierarchically labeled abnormal point subset for tracing; for the hierarchically labeled abnormal point subset, calculating the overlap degree of the high-response area between different scales through cross-granularity correlation calculation to determine a key point path with the highest correlation strength; and based on the key point path, locating a seasonal inflection point source in an original curve by using a feature map backtracking mechanism of a convolutional neural network to obtain a starting abnormal point of a tracing chain. The scheme improves the abnormal positioning accuracy and the tracing integrity.

Description

Technical Field

[0001] This invention belongs to the field of power marketing data processing, and specifically relates to a method and system for tracing abnormal power marketing data based on multi-scale convolution. Background Technology

[0002] In power systems, a core and complex technical challenge lies in the anomaly detection and source tracing analysis of daily electricity consumption curves: how to accurately identify and correlate anomalies in electricity consumption curves at multiple time granularities, and trace their root causes to reveal deep-seated patterns and potential risks in user electricity consumption behavior. This problem stems from the complex fluctuation characteristics of electricity consumption data across multiple time scales, such as daily, weekly, and monthly. Analysis at a single time granularity often fails to fully capture the inherent relationships between abrupt changes, periodic fluctuations, and seasonal inflection points, resulting in insufficient accuracy in anomaly identification and incomplete source tracing chains.

[0003] At a daily scale, a sudden change in a user's electricity consumption might be detected, but it's impossible to determine whether it's related to regular weekly fluctuations or whether it's affected by monthly seasonal trends. This could lead to misjudging the nature and importance of the anomaly. Cross-time granularity feature correlation analysis presents technical challenges. Calculating the overlap of high-response regions at different scales is complex, and improper threshold settings can easily overlook key correlation paths, affecting the final aggregation effect of anomalies.

[0004] Regarding the aforementioned core issues, existing technologies also present other technical challenges: the balance between sensitivity and noise interference in the extraction of mutation point features at the daily scale, and the distinction between periodic fluctuations and seasonal trends at the weekly and monthly scales. Therefore, it is urgent to construct a comprehensive analysis framework with multiple time granularities to ensure the logical consistency and business interpretability of the entire chain from anomaly detection to source tracing. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for tracing abnormal data in power marketing based on multi-scale convolution, which solves the technical problem of tracing abnormal data across the entire chain from daily mutations to weekly cycles and monthly seasons, thereby improving the accuracy of abnormal location and the completeness of tracing.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0007] This invention first discloses a method for tracing the source of abnormal electricity marketing data based on multi-scale convolution, which includes the following steps:

[0008] Obtain daily electricity consumption curve data of users, apply 3×3 convolution kernel to the daily electricity consumption curve data of users through convolutional neural network, extract the features of abrupt change points at the single-day scale, and obtain a subset of curves with preliminary high response area annotation;

[0009] Based on the subset of curves labeled in the initial high-response region, a 7×7 convolution kernel is used to process the periodic fluctuation characteristics at the periodic scale to determine the expanded set of outliers.

[0010] By using the expanded set of outliers, a 15×15 convolutional kernel is applied to analyze the seasonal inflection point trend at the monthly seasonal scale. It is determined whether the deviation between the seasonal inflection point trend and the weekly pattern exceeds a preset threshold. If it does, multi-time granular information is fused to obtain a comprehensive set of key outliers.

[0011] Based on the comprehensive set of key anomalies, a clustering algorithm is used to group the mutation points and periodic fluctuation points in the high-response region to obtain a hierarchically labeled subset of anomalies.

[0012] For a subset of outlier points with hierarchical annotation, the overlap of high-response regions at different scales is calculated through cross-granularity association to determine the key point path with the highest association strength.

[0013] Based on the key point path, the feature map backtracking mechanism of the convolutional neural network is used to locate the source of the seasonal inflection point in the original curve, and the starting abnormal point of the traceability chain is obtained.

[0014] Based on the starting anomaly point of the traceability chain, aggregate the relevant anomaly point set at multiple time granularities, determine whether the aggregation result shows that periodic fluctuation points dominate, and if so, output the final set of key anomaly points for cross-granularity related traceability.

[0015] The present invention further includes the following preferred embodiments:

[0016] The step of applying a 3×3 convolution kernel to the user's daily electricity consumption curve data using a convolutional neural network to extract abrupt change features at a single-day scale further includes:

[0017] The subset of curves marked in the initial high-response area is standardized to obtain the daily power consumption sequence;

[0018] A convolutional neural network is used to traverse the daily electricity consumption sequence using a 3×3 convolutional kernel, and the feature response value at each position is calculated to obtain the abrupt feature response map.

[0019] By setting a response intensity threshold, the mutation feature response map is scanned, and the locations where the response value exceeds the threshold are marked to obtain a set of daily mutation points.

[0020] For each set of daily mutation points, the corresponding time points are marked on the original daily electricity consumption sequence to generate a labeled daily curve with mutation labels;

[0021] Extract curves containing at least one mutation label from all labeled daily curves to form a subset of curves labeled with high response regions;

[0022] The subset of curves labeled with high response regions is aggregated according to user identifiers to obtain the initial subset of curves labeled with high response regions for each user.

[0023] The method of using a 7×7 convolution kernel to process periodic fluctuation features on a periodic scale further includes:

[0024] A subset of curves with high response regions is obtained from the initially labeled data. Feature extraction is performed on the data at the periodic scale to obtain a basic feature description of the periodic fluctuations.

[0025] Based on the basic characteristics of periodic fluctuations, identify points that deviate from the normal range, form an initial set of outliers, determine whether there is a significant periodic deviation, and if a deviation is detected, record the relevant point information.

[0026] For the initial set of outliers, an expansion rule is used to analyze the surrounding data points, and combined with the fluctuation characteristics on a periodic scale, the expanded set of outliers is determined.

[0027] By analyzing the feature distribution related to the weekly pattern in the expanded set of outliers, the matching degree of the periodic pattern is obtained, and it is determined whether it meets the preset correlation threshold. If it does, it is marked as a related point.

[0028] Based on the marked relevant points, fluctuation features related to the weekly pattern are extracted and in-depth analysis is performed on the features to obtain an expanded set of outliers.

[0029] The application of 15×15 convolution kernels to analyze the seasonal inflection point trend at the monthly seasonal scale further includes:

[0030] The original set of anomalies is matched and expanded with the weekly pattern association rules to obtain an expanded set of anomalies.

[0031] By sliding a 15×15 convolutional kernel across the monthly seasonal sequence, seasonal-scale trend feature maps are obtained within each time window.

[0032] Perform first-order and second-order difference operations on the seasonal trend feature map to determine the location and intensity of seasonal inflection points.

[0033] For each seasonal inflection point, extract the weekly pattern association strength sequence within the corresponding time period;

[0034] The mean square error between the seasonal inflection point intensity sequence and the weekly pattern correlation intensity sequence is calculated to obtain the deviation value;

[0035] If the deviation value exceeds the preset threshold, the current seasonal inflection point is marked as a candidate key anomaly; otherwise, it is discarded.

[0036] Daily, weekly, and monthly granular features were extracted from all candidate key anomalies and then concatenated to obtain a comprehensive set of key anomalies.

[0037] The method of grouping abrupt change points and periodic fluctuation points within a high-response region using a clustering algorithm further includes:

[0038] Data is extracted from the comprehensive set of key anomalies. Preliminary screening is performed on the location information in the high response area. Preset thresholds are used to distinguish between mutation points and periodic fluctuations, resulting in a preliminary set of classified locations.

[0039] For the initial classification of the point set, clustering methods are used to group the mutation points and periodic fluctuations. By analyzing the spatial distribution and temporal correlation between the points, the hierarchical anomaly subset is determined.

[0040] Based on the hierarchically labeled anomaly subsets, the grouping characteristics of mutation points and periodic fluctuations are obtained. Time series comparisons are performed on the point data within each group to determine whether there are significant fluctuation patterns.

[0041] If significant fluctuation patterns exist, in-depth analysis of the point data within the relevant groups is conducted to obtain potential anomaly triggering patterns by calculating the correlation strength between points.

[0042] Based on potential anomaly triggering patterns, further feature extraction is performed on point groups within the high response zone. By comparing the fluctuation analysis results between different groups, the core anomaly points in the key set are identified.

[0043] For core anomalies, and in accordance with the needs of source tracing, the location information within the anomaly subset is organized in a structured manner, and a hierarchically labeled subset of anomalies is obtained by dividing the region.

[0044] The calculation of high-response region overlap across different scales through cross-granularity correlation further includes:

[0045] Obtain the set of anomalies labeled at each level based on the subset of anomalies labeled at each level;

[0046] High-response activation maps of the corresponding scale are extracted from the set of outliers at each level using a multi-scale convolutional neural network.

[0047] Calculate the pixel-level intersection-union matrix for two adjacent high-response activation maps to obtain the overlap distribution map;

[0048] If the same location has a cross-union ratio (CUI) value higher than a preset threshold in three consecutive layers of the cross-union matrix, then that location is marked as a high-intensity association point.

[0049] A cross-scale interconnected point graph is formed by connecting high-strength correlation points at adjacent levels;

[0050] A graph search algorithm is used to find the strongest path that runs from the finest to the coarsest level in a connected graph of related points;

[0051] The path with the highest association strength is obtained by back-mapping the original subset of anomalies back to the path with the highest strength.

[0052] The method of using a convolutional neural network to backtrack the feature map to locate the source of the seasonal inflection point in the original curve further includes:

[0053] Deep feature maps are obtained by performing multi-layer feature extraction on the original time series using a convolutional neural network.

[0054] High-response regions are identified based on the strong activation locations of seasonal inflection points on the last layer of the feature map.

[0055] The activation contribution map of each layer is obtained by backpropagating from the high response region to the input layer using a gradient backtracking method.

[0056] Obtain the channel attention weights corresponding to high-response regions in each feature map layer;

[0057] The weighted activation map sequence is obtained by weighting the feature maps of each layer by channel attention weights;

[0058] Path strength accumulation calculation is performed on the weighted activation graph sequence to obtain multiple backtracking paths from the output layer to the input layer and their strengths;

[0059] The main path of anomaly propagation is obtained by sorting and selecting the single backtracking path with the highest strength based on path strength.

[0060] The starting point of the anomaly was determined by tracing the main path of the anomaly propagation to the corresponding position in the original time series.

[0061] The step of aggregating a set of related anomalies at multiple time granularities based on the starting anomaly point of the tracing chain, and determining whether the aggregation result shows that periodic fluctuation points dominate, further includes:

[0062] The initial anomaly point is determined based on the source chain, and an initial anomaly point sequence is obtained. The KMeans clustering algorithm is used to aggregate the initial anomaly point sequence at multiple time granularities to obtain multiple anomaly point sets.

[0063] By calculating the periodic score of timestamps within each set of outliers, a subset of periodic fluctuation points is determined, and the set with the highest proportion of periodic fluctuation points is extracted from multiple sets of outliers as the dominant period set.

[0064] If the dominant period set has the highest proportion, then all points in the dominant period set are marked as cross-granularity association candidate points, and a candidate key point sequence is obtained.

[0065] Based on the common predecessor nodes of the candidate key point sequence at different time granularities, the associated predecessor node sequence is obtained by tracing back the source chain. The nodes that appear at at least two granularities at the same time are retained by the intersection operation, and the final set of abnormal key points for cross-granularity association tracing is output.

[0066] This invention also discloses a multi-scale convolution-based power marketing anomaly data tracing system that utilizes the aforementioned multi-scale convolution-based power marketing anomaly data tracing method, further comprising:

[0067] The curve subset acquisition module is used to acquire daily electricity consumption curve data of users. It applies a 3×3 convolution kernel to the daily electricity consumption curve data of users through a convolutional neural network, extracts the features of abrupt change points at the single-day scale, and obtains a preliminary high-response region labeled curve subset.

[0068] The periodic fluctuation feature processing module is used to process the periodic fluctuation features at the periodic scale based on the curve subset labeled in the initial high response region using a 7×7 convolution kernel, and to determine the expanded set of outliers.

[0069] The seasonal inflection point trend judgment module is used to analyze the seasonal inflection point trend at the monthly seasonal scale by applying a 15×15 convolution kernel to the expanded set of outliers. It determines whether the deviation between the seasonal inflection point trend and the weekly pattern exceeds a preset threshold. If it does, it integrates multi-time granular information to obtain a comprehensive set of key outliers.

[0070] The hierarchical annotation module is used to group mutation points and periodic fluctuation points in high-response regions based on the comprehensive set of key anomalies, using a clustering algorithm to obtain a subset of anomalies with hierarchical annotation.

[0071] The cross-granularity association calculation module is used to determine the key point path with the highest association strength by calculating the overlap of high-response regions at different scales for a subset of hierarchically labeled outlier points.

[0072] The feature map backtracking module is used to locate the source of the seasonal inflection point in the original curve based on the key point path and using the feature map backtracking mechanism of the convolutional neural network, so as to obtain the starting abnormal point of the traceability chain.

[0073] The output module is used to aggregate a set of related anomalies at multiple time granularities based on the starting anomaly point of the traceability chain, determine whether the aggregation result shows that periodic fluctuation points dominate, and if so, output the final set of key anomaly points for cross-granularity related traceability.

[0074] Accordingly, this application also discloses a terminal, including a processor and a storage medium;

[0075] The storage medium is used to store instructions;

[0076] The processor is configured to operate according to the instructions to perform the steps of the aforementioned method for tracing abnormal electricity marketing data based on multi-scale convolution.

[0077] Accordingly, this application also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned method for tracing abnormal electricity marketing data based on multi-scale convolution.

[0078] The beneficial effects of this invention are that, compared with the prior art, it provides a method and system for tracing the source of abnormal electricity marketing data based on multi-scale convolution. Addressing the core business challenge of broken tracing chains and missing correlations between weekly, monthly, and seasonal patterns caused by single-scale analysis of anomalies in daily electricity consumption data, this invention achieves multi-time-granularity fusion detection and tracing of anomalies through a progressively layered convolutional kernel design. A 3×3 convolutional kernel is used to extract single-day mutation features to form high-response regions for initial labeling. A 7×7 convolutional kernel is used to capture weekly fluctuations to expand the anomaly set. A 15×15 convolutional kernel is used to analyze monthly and seasonal inflection point trends and compare them with weekly pattern deviations, fusing them to form a comprehensive set of key anomalies. Clustering and cross-granularity overlap calculations are used to construct correlation strength paths. A feature map backtracking mechanism is used to accurately locate the source of seasonal inflection points, aggregating to form a cross-granularity set of key anomalies dominated by periodic fluctuations. This achieves full-chain anomaly tracing from daily mutations to weekly cycles and then to monthly and seasonal patterns, significantly improving anomaly location accuracy and tracing completeness. Attached Figure Description

[0079] Figure 1 This is a flowchart of the method for tracing abnormal electricity marketing data based on multi-scale convolution in this invention. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0081] The embodiments described in this application are merely some, not all, embodiments of the present invention. Based on the spirit of the present invention, other embodiments obtained by those skilled in the art without inventive effort are all within the protection scope of the present invention.

[0082] To address the shortcomings of existing technologies, this invention proposes a method and system for tracing abnormal electricity marketing data based on multi-scale convolution. (See [link to relevant documentation]). Figure 1 As shown, the method of the present invention includes the following steps:

[0083] Step S101: Obtain user daily electricity consumption curve data, apply 3×3 convolution kernel to user daily electricity consumption curve data through convolutional neural network, extract abrupt change point features for single-day scale, and obtain a subset of curves with preliminary high response area annotation;

[0084] Step S102: Based on the subset of curves labeled with the initial high-response region, use a 7×7 convolution kernel to process the periodic fluctuation features at the periodic scale to determine the expanded set of outliers;

[0085] Step S103: Using the expanded set of outliers, apply a 15×15 convolution kernel to analyze the seasonal inflection point trend at the monthly seasonal scale, and determine whether the deviation between the seasonal inflection point trend and the weekly pattern exceeds a preset threshold. If it does, fuse multi-time granular information to obtain a comprehensive set of key outliers.

[0086] Step S104: Based on the comprehensive set of key anomalies, a clustering algorithm is used to group the mutation points and periodic fluctuation points in the high-response region to obtain a hierarchically labeled subset of anomalies;

[0087] Step S105: For the subset of outlier points with hierarchical annotation, calculate the overlap of high-response regions at different scales through cross-granularity association to determine the key point path with the highest association strength;

[0088] Step S106: Based on the key point path, the feature map backtracking mechanism of the convolutional neural network is used to locate the source of the seasonal inflection point in the original curve, and the starting abnormal point of the traceability chain is obtained.

[0089] Step S107: Aggregate the set of related anomalies at multiple time granularities based on the starting anomaly point of the traceability chain, and determine whether the aggregation result shows that periodic fluctuation points dominate. If so, output the final set of key anomaly points for cross-granularity related traceability.

[0090] In a preferred embodiment, step S101, which involves applying a 3×3 convolution kernel to the user's daily electricity consumption curve data using a convolutional neural network to extract abrupt change features at the daily scale, further includes:

[0091] The subset of curves marked in the initial high-response area is standardized to obtain the daily power consumption sequence;

[0092] A convolutional neural network is used to traverse the daily electricity consumption sequence using a 3×3 convolutional kernel, and the feature response value at each position is calculated to obtain the abrupt feature response map.

[0093] ;

[0094] in, Indicates the time position of the daily electricity consumption sequence Place, No. Input values ​​in adjacent dimensions, express The weight parameters at corresponding positions in the convolution kernel. For bias terms, For time location The mutation characteristic response value at the location, all This forms a mutation characteristic response map.

[0095] By setting a response intensity threshold, the mutation feature response map is scanned, and the locations where the response value exceeds the threshold are marked to obtain a set of daily mutation points.

[0096] For each set of daily mutation points, the corresponding time points are marked on the original daily electricity consumption sequence to generate a labeled daily curve with mutation labels;

[0097] Extract curves containing at least one mutation label from all labeled daily curves to form a subset of curves labeled with high response regions;

[0098] The subset of curves labeled with high response regions is aggregated according to user identifiers to obtain the initial subset of curves labeled with high response regions for each user.

[0099] In a preferred embodiment, step S102, which uses a 7×7 convolution kernel to process the periodic fluctuation features at the periodic scale, further includes:

[0100] A subset of curves with high response regions is obtained from the initially labeled data. Feature extraction is performed on the data at the periodic scale to obtain a basic feature description of the periodic fluctuations.

[0101] ;

[0102] in, This represents the subset of curves after initial high-response region labeling, the first... Within each cycle, the time position is Electricity consumption data values, Indicates the use of periodic-scale feature extraction Convolution kernel, This indicates the number of weekly cycles included in the statistics. , representing the radius of the convolution kernel in the time and period dimensions. Indicates time position The basic characteristic description value of the periodic fluctuation at the location;

[0103] Based on the basic characteristics of periodic fluctuations, identify points that deviate from the normal range, form an initial set of outliers, determine whether there is a significant periodic deviation, and if a deviation is detected, record the relevant point information.

[0104] For the initial set of outliers, an expansion rule is used to analyze the surrounding data points, and combined with the fluctuation characteristics on a periodic scale, the expanded set of outliers is determined.

[0105] By analyzing the feature distribution related to the weekly pattern in the expanded set of outliers, the matching degree of the periodic pattern is obtained, and it is determined whether it meets the preset correlation threshold. If it does, it is marked as a related point.

[0106] Based on the marked relevant points, fluctuation features related to the weekly pattern are extracted and in-depth analysis is performed on the features to obtain an expanded set of outliers.

[0107] In a specific embodiment, a subset of curves is extracted from the initially labeled high-response regions. Time periods where the response intensity exceeds the mean plus 2.5 times the standard deviation (σ=0.018) are selected as high-response segments, resulting in a subset containing approximately 127 daily-dimensional curves with a total of approximately 8890 points. Next, each curve is aligned to a weekly cycle. Using a 7-day cycle, all curves are resampled to a uniform 168-hour grid (1-hour stride). An interpolation algorithm is used to fill in missing points, ensuring consistent positions within the week. A 7×7 convolutional kernel is constructed to capture local fluctuations at the weekly scale. Specifically, the kernel is designed with a Gaussian weighted differential kernel with σ=1.5 on the outer layer to highlight periodic jumps, a Sobel vertical operator filling the central 4×4 region to enhance intra-day and intra-weekly interaction edges, and a Laplacian operator filling the remaining edges to strengthen the second-order transition. The kernel values ​​are normalized to have a center weight of 0.28, a vertical weight of ±0.12, a horizontal weight of ±0.09, and a corner weight of -0.05. A 168×N feature matrix is ​​constructed for each curve, and then the 7×7 kernel is applied point-by-point for two-dimensional convolution. The absolute value of the resulting response map is taken, and 3×3 max pooling is performed. Coordinates greater than a threshold of 0.35 after pooling are extracted as candidate points for extended anomalies. These candidate points are then mapped back to the original timestamps and joined with the initial high-response points. Weekly pattern correlation is calculated: the hourly distribution entropy of each extended point within a week is calculated. If the entropy is less than 1.8 bits, it is marked as a strong weekly pattern correlation. The expanded set of anomalies increases from the initial 312 to 487, of which 268 have significant weekly periodic characteristics, with a mean correlation of 0.86, thus achieving week-pattern-driven anomaly expansion detection.

[0108] In a preferred embodiment, step S103, applying a 15×15 convolution kernel to analyze the seasonal inflection point trend at a monthly seasonal scale, further includes:

[0109] The original set of anomalies is matched and expanded with the weekly pattern association rules to obtain an expanded set of anomalies.

[0110] By sliding a 15×15 convolutional kernel across the monthly seasonal sequence, seasonal-scale trend feature maps are obtained within each time window.

[0111] ;

[0112] in, In the monthly seasonal sequence, the first... Within a seasonal cycle, the time position is Electricity consumption data values, This indicates the data used for monthly seasonal-scale trend extraction. Convolution kernel, , representing the radius of the convolution kernel in the time and seasonal dimensions. Indicates by time and location The response value of the seasonal-scale trend feature map centered on the center.

[0113] Perform first-order and second-order difference operations on the seasonal trend feature map to determine the location and intensity of seasonal inflection points.

[0114] ;

[0115] in, The seasonal trend feature plot shows the time position. The response value at that location, Indicates time position The seasonal inflection point intensity value at the location.

[0116] For each seasonal inflection point, extract the weekly pattern association strength sequence within the corresponding time period;

[0117] ;

[0118] in, This indicates a specific, established seasonal turning point. This represents the time offset centered on the seasonal inflection point, forming the corresponding time period. Indicates the first Within each cycle, the time position is Electricity consumption data This represents the convolution kernel used for weekly pattern correlation calculation. This indicates the number of weekly cycles included in the statistics. Indicates seasonal turning point The weekly pattern association strength sequence value within the corresponding time period.

[0119] The mean square error between the seasonal inflection point intensity sequence and the weekly pattern correlation intensity sequence is calculated to obtain the deviation value. If the deviation value exceeds the preset threshold, the current seasonal inflection point is marked as a candidate key anomaly; otherwise, it is discarded.

[0120] Daily, weekly, and monthly granular features were extracted from all candidate key anomalies and then concatenated to obtain a comprehensive set of key anomalies.

[0121] In a specific embodiment, a comprehensive set of key anomalies is extracted from multi-dimensional time-series indicators (such as CPU utilization, interface QPS, error rate, and latency). An anomaly detection threshold is set: Z-Score > 3.5 or Isolation Forest algorithm score < 0.45. This results in an initial set of anomalies accounting for approximately 0.87% of the total data points (12563 points). High-response regions, i.e., continuous segments with response times > 800ms during peak business hours (19:00-23:00), are extracted, resulting in 47 high-response time windows. Within these windows, feature engineering is performed on the anomalies to construct 7-dimensional feature vectors (original value, deviation from the mean, local density, Euclidean distance to the nearest normal point, amplitudes of the first three main frequencies before Fourier transform, Hurst exponent, and wavelet coefficient energy). The DBSCAN algorithm (eps=0.62, min_samples=8) is used for clustering, yielding 5 clusters. The main clusters are: Cluster 1 (mutant type, 4128 points, characterized by high bias and low periodicity), Cluster 2 (high-frequency periodic fluctuation, 3659 points, with a dominant frequency of 0.0167Hz corresponding to a 1-minute period), Cluster 3 (low-frequency daily periodic fluctuation, 2874 points, with a dominant frequency of 1.157×10^-5Hz), Cluster 4 (hybrid type, 1123 points), and Cluster 5 (noisy point-like noise, 779 points). Each cluster is hierarchically labeled: Cluster 1 is labeled as "mutant shock anomaly" with priority 1, Clusters 2 and 3 are labeled as "periodic fluctuation anomaly" with priority 2, Cluster 4 is labeled as "composite type" with priority 3, and Cluster 5 is directly filtered. The final output is a hierarchically labeled subset of 9781 anomaly points, with cluster labels and confidence scores (Silhouette coefficient 0.74), which are directly written into the feature database for use by the causal tracing model for root cause localization. The entire process is fully automated pipelined using Spark Streaming + PySpark MLlib, requiring no manual intervention.

[0122] In a preferred embodiment, step S104, which uses a clustering algorithm to group mutation points and periodic fluctuation points within the high-response region, further includes:

[0123] Data is extracted from the comprehensive set of key anomalies. Preliminary screening is performed on the location information in the high response area. Preset thresholds are used to distinguish between mutation points and periodic fluctuations, resulting in a preliminary set of classified locations.

[0124] For the initial classification of the point set, clustering methods are used to group the mutation points and periodic fluctuations. By analyzing the spatial distribution and temporal correlation between the points, the hierarchical anomaly subset is determined.

[0125] Based on the hierarchically labeled anomaly subsets, the grouping characteristics of mutation points and periodic fluctuations are obtained. Time series comparisons are performed on the point data within each group to determine whether there are significant fluctuation patterns.

[0126] If significant fluctuation patterns exist, in-depth analysis of the point data within the relevant groups is conducted to obtain potential anomaly triggering patterns by calculating the correlation strength between points.

[0127] Based on potential anomaly triggering patterns, further feature extraction is performed on point groups within the high response zone. By comparing the fluctuation analysis results between different groups, the core anomaly points in the key set are identified.

[0128] For core anomalies, and in accordance with the needs of source tracing, the location information within the anomaly subset is organized in a structured manner, and a hierarchically labeled subset of anomalies is obtained by dividing the region.

[0129] In a preferred embodiment, step S105, which calculates the overlap of high-response regions at different scales through cross-granularity correlation, further includes:

[0130] Obtain the set of anomalies labeled at each level based on the subset of anomalies labeled at each level;

[0131] High-response activation maps of the corresponding scale are extracted from the set of outliers at each level using a multi-scale convolutional neural network.

[0132] ;

[0133] in, This represents the first step in a multi-scale convolutional neural network. Each scale level, Indicates the first The feature input of the set of outliers corresponding to the level, Indicates the first The convolution kernel corresponds to the level of the layer. Indicates the first Hierarchical convolution kernel radius, Represents a non-linear activation function. Indicates the first High-response activation maps extracted hierarchically.

[0134] Calculate the pixel-level intersection-union matrix for two adjacent high-response activation maps to obtain the overlap distribution map;

[0135] If the same location has a cross-union ratio (CUI) value higher than a preset threshold in three consecutive layers of the cross-union matrix, then that location is marked as a high-intensity association point.

[0136] A cross-scale interconnected point graph is formed by connecting high-strength correlation points at adjacent levels;

[0137] A graph search algorithm is used to find the strongest path that runs from the finest to the coarsest level in a connected graph of related points;

[0138] ;

[0139] in, This represents a connected graph of interconnected points consisting of highly interconnected points across scales. This represents a feasible path from the finest-grained node to the coarsest-grained node. () represents a pair of adjacent related points on the path. Represents a node and The correlation strength weight between them This represents the path with the highest cumulative correlation strength.

[0140] The path with the highest association strength is obtained by back-mapping the original subset of anomalies back to the path with the highest strength.

[0141] In a preferred embodiment, step S106, which uses a feature map backtracking mechanism of a convolutional neural network to locate the source of the seasonal inflection point in the original curve, further includes:

[0142] Deep feature maps are obtained by performing multi-layer feature extraction on the original time series using a convolutional neural network.

[0143] High-response regions are determined based on the strong activation positions of seasonal inflection points on the last layer of the feature map; the strong activation indicates that the feature response value of the position corresponding to the seasonal inflection point in the last layer of the feature map is higher than the background response level in the statistical distribution of the same layer of the feature map, indicating that the corresponding position has a dominant contribution to the seasonal trend change.

[0144] The activation contribution map of each layer is obtained by backpropagating from the high response region to the input layer using a gradient backtracking method.

[0145] Obtain the channel attention weights corresponding to high-response regions in each feature map layer;

[0146] The weighted activation map sequence is obtained by weighting the feature maps of each layer by channel attention weights;

[0147] Path strength accumulation calculation is performed on the weighted activation graph sequence to obtain multiple backtracking paths from the output layer to the input layer and their strengths;

[0148] The main path of anomaly propagation is obtained by sorting and selecting the single backtracking path with the highest strength based on path strength.

[0149] The starting point of the anomaly was determined by tracing the main path of the anomaly propagation to the corresponding position in the original time series.

[0150] In a preferred embodiment, step S107, which involves aggregating a set of related anomalies at multiple time granularities based on the starting anomaly point of the traceability chain, and determining whether the aggregation result shows that periodic fluctuation points dominate, further includes:

[0151] The initial anomaly point is determined based on the source chain, and an initial anomaly point sequence is obtained. The KMeans clustering algorithm is used to aggregate the initial anomaly point sequence at multiple time granularities to obtain multiple anomaly point sets.

[0152] By calculating the periodic score of timestamps within each set of outliers, a subset of periodic fluctuation points is determined, and the set with the highest proportion of periodic fluctuation points is extracted from multiple sets of outliers as the dominant period set.

[0153] If the dominant period set has the highest proportion, then all points in the dominant period set are marked as cross-granularity association candidate points, and a candidate key point sequence is obtained.

[0154] Based on the common predecessor nodes of the candidate key point sequence at different time granularities, the associated predecessor node sequence is obtained by tracing back the source chain. The nodes that appear at at least two granularities at the same time are retained by the intersection operation, and the final set of abnormal key points for cross-granularity association tracing is output.

[0155] In a real-world case of inspecting electricity consumption data quality and marketing anomalies for large-scale residential users, the power company's intelligent platform automatically collected nearly two months' worth of hourly electricity consumption data from thousands of users. With the arrival of high temperatures, the number of alarms on the platform's backend surged. The technical team decided to use the latest multi-scale convolutional electricity marketing anomaly data tracing method to deeply mine and accurately trace the source of these alarm data.

[0156] The raw electricity consumption curves of all users were standardized to remove abnormal and extreme outliers. Based on this, 24-hour daily electricity consumption data for each household was scanned using a 3×3 convolution kernel. Taking a typical residential user as an example, their electricity consumption curve remained relatively stable throughout the week, but on one particular day at 7 PM, 8 PM, and 9 PM, there were sudden increases in electricity consumption, reaching 2.2, 3.1, and 3.0 hours respectively, significantly higher than the daily average of 1.2. The abrupt change response values ​​calculated by the convolutional neural network were 2.14, 3.02, and 2.95, all exceeding the system's set abrupt change threshold of 1.8. These were directly identified as single-day abrupt change anomalies, and the user, date, time period, and outlier information were recorded in the anomaly database.

[0157] The platform has detected approximately 3,000 daily power consumption anomalies, involving hundreds of users. The results fully demonstrate the high sensitivity and accuracy of this method in detecting daily power consumption anomalies, effectively overcoming the problem that the traditional mean-variance method is susceptible to extreme values.

[0158] For all identified single-day anomalous data points, further automated analysis was performed on a weekly cycle scale. The technical team merged and aligned historical electricity consumption data for all users at the same time period, using a 7-day cycle. Taking the aforementioned anomalous user as an example, their electricity consumption at 7 PM on Wednesdays over the past four weeks was 1.3, 1.2, 1.5, and 3.1 respectively. The system calculated its cycle entropy to be 1.05, lower than the cycle anomaly threshold of 1.5, indicating that the anomalous data point during this period was highly correlated with the cycle peak. A 7×7 convolution kernel was then applied to these high-response points, and by combining the convolution response value with the cycle pattern matching degree, anomalous points with a high correlation to the weekly pattern were automatically filtered out. The results showed that 70% of all anomalous points also exhibited strong periodicity (mean cycle correlation 0.89), thus labeling these points as cycle-fluctuation-dominated anomalies, enhancing the business interpretability of the anomalous points.

[0159] Continuing with the trend scanning of the previous periodic anomalies over a longer timeframe, using a 30-day window, the platform employs a 15×15 convolutional kernel sliding analysis to examine each user's monthly electricity consumption curve and identify seasonal trend inflection points. For this particular user, the average daily electricity consumption gradually increased from 3.4 to 5.9 and then to 6.1 within five days before and after the anomaly. The calculated trend inflection point strength was 2.07, representing the largest increase in the past two months. After first- and second-order difference operations, this point was also identified as a monthly inflection point anomaly. The platform automatically aggregated all points with a trend inflection point strength exceeding 2.0 into the monthly seasonal anomaly set. Statistical analysis revealed that nearly 30% of all anomalies were driven by monthly trends. This multi-timescale automatic feature extraction significantly improved the temporal depth of anomaly detection, compensating for the shortcomings of previous single-granularity analyses that missed periodic fluctuations and seasonal abrupt changes.

[0160] After obtaining three types of outliers—daily mutations, weekly cycles, and monthly inflection points—the system automatically generates a vector for each point containing multi-dimensional features such as mutation response value, cycle pattern entropy, trend inflection point strength, and deviation from the mean. All outliers are uniformly input into the KMeans clustering model (number of categories k=4). After clustering, most outliers are classified into mutation-dominated, cycle fluctuation-dominated, and seasonal inflection point-dominated types, with the remainder being considered noise points. The outliers of the aforementioned user are classified as "mutation + seasonal inflection point composite type" and given a high-priority label. The clustering results are written to the outlier feature database in real time by the system, serving as the basis for constructing the source tracing chain.

[0161] For each type of anomaly, high-response regions are generated in the daily, weekly, and monthly activation maps, and the cross-union ratio (CUNR) is automatically calculated to quantify the correlation strength between different granularities. The CUNRs of a certain anomaly in the three activation maps are 0.77, 0.81, and 0.84, all exceeding the system's set threshold of 0.7. The system automatically filters all points with high CUNRs as "cross-scale high-intensity anomalies," accumulating over 500 globally. Through this high-intensity path, anomalies at the daily, weekly, and monthly levels can be accurately connected to achieve a "point-line-surface" closed-loop tracing, completely eliminating chain breaks and isolated anomaly problems.

[0162] For critical anomalies with high overlap across multiple scales, a feature map backtracking algorithm is automatically executed to trace the anomaly back to its root cause from the finest granular level. In this embodiment, the backtracking path of a certain anomaly is 21:00 → 20:00 → 19:00 → the same time period of the previous cycle → the trend inflection point of the previous month. By comprehensively considering indicators such as channel attention weight and response intensity, the "main path of anomaly propagation" is output. Ultimately, it is confirmed that the fundamental driver of this point is the dual effect of seasonal high temperatures and concentrated weekly loads. This will be written into the anomaly report and provide authoritative reference for manual verification, significantly enhancing the transparency and interpretability of anomaly tracing.

[0163] After all anomalies are aggregated, the system automatically calculates the proportion of three types of anomalies for each user: periodic fluctuations, sudden changes, and seasonal inflection points. In Example 2, among the user's multiple periods of high response anomalies, periodic fluctuations accounted for as much as 73%, with a periodic pattern score of 3.2 (out of a maximum of 5). This was ultimately classified as a periodic fluctuation-dominated anomaly, and energy-saving and load balancing suggestions were automatically pushed to avoid meaningless security alerts, thus optimizing marketing strategies and user experience.

[0164] To fully validate the effectiveness of this method, the platform simultaneously tested the same user data using the traditional method of daily average plus three standard deviations. The results showed that the traditional method mainly detected a few significant abrupt changes in electricity consumption, identifying over 1,200 anomalies. However, it missed a large number of periodic and seasonal inflection point anomalies, with a near 50% miss rate for periodic anomalies and over 60% for seasonal inflection point anomalies, resulting in high false alarm rates and heavy manual verification workload. In contrast, the method of this invention improved the detection rate of periodic anomalies to over 90% and seasonal inflection point anomalies to nearly 90% for the same batch of data, reduced the false alarm rate to half, and decreased the average manual verification time per household from nine minutes to less than three minutes. The data demonstrates that the multi-scale convolution method significantly improves the detection completeness and business processing efficiency.

[0165] The above embodiments fully demonstrate the end-to-end, automated, and efficient nature of the power marketing anomaly data tracing method based on multi-scale convolution. It not only greatly improves the accuracy, completeness, and interpretability of anomaly detection and tracing, but also significantly reduces the manual burden and promotes the intelligent upgrade of power marketing business.

[0166] The beneficial effects of this invention are that, compared with the prior art, it provides a method and system for tracing the source of abnormal electricity marketing data based on multi-scale convolution. Addressing the core business challenge of broken tracing chains and missing correlations between weekly, monthly, and seasonal patterns caused by single-scale analysis of anomalies in daily electricity consumption data, this invention achieves multi-time-granularity fusion detection and tracing of anomalies through a progressively layered convolutional kernel design. A 3×3 convolutional kernel is used to extract single-day mutation features to form high-response regions for initial labeling. A 7×7 convolutional kernel is used to capture weekly fluctuations to expand the anomaly set. A 15×15 convolutional kernel is used to analyze monthly and seasonal inflection point trends and compare them with weekly pattern deviations, fusing them to form a comprehensive set of key anomalies. Clustering and cross-granularity overlap calculations are used to construct correlation strength paths. A feature map backtracking mechanism is used to accurately locate the source of seasonal inflection points, aggregating to form a cross-granularity set of key anomalies dominated by periodic fluctuations. This achieves full-chain anomaly tracing from daily mutations to weekly cycles and then to monthly and seasonal patterns, significantly improving anomaly location accuracy and tracing completeness.

[0167] This invention can be a system, method, and / or computer program product. This invention also discloses a multi-scale convolution-based power marketing anomaly data tracing system based on the aforementioned multi-scale convolution-based power marketing anomaly data tracing method, further comprising:

[0168] The curve subset acquisition module is used to acquire daily electricity consumption curve data of users. It applies a 3×3 convolution kernel to the daily electricity consumption curve data of users through a convolutional neural network, extracts the features of abrupt change points at the single-day scale, and obtains a preliminary high-response region labeled curve subset.

[0169] The periodic fluctuation feature processing module is used to process the periodic fluctuation features at the periodic scale based on the curve subset labeled in the initial high response region using a 7×7 convolution kernel, and to determine the expanded set of outliers.

[0170] The seasonal inflection point trend judgment module is used to analyze the seasonal inflection point trend at the monthly seasonal scale by applying a 15×15 convolution kernel to the expanded set of outliers. It determines whether the deviation between the seasonal inflection point trend and the weekly pattern exceeds a preset threshold. If it does, it integrates multi-time granular information to obtain a comprehensive set of key outliers.

[0171] The hierarchical annotation module is used to group mutation points and periodic fluctuation points in high-response regions based on the comprehensive set of key anomalies, using a clustering algorithm to obtain a subset of anomalies with hierarchical annotation.

[0172] The cross-granularity association calculation module is used to determine the key point path with the highest association strength by calculating the overlap of high-response regions at different scales for a subset of hierarchically labeled outlier points.

[0173] The feature map backtracking module is used to locate the source of the seasonal inflection point in the original curve based on the key point path and using the feature map backtracking mechanism of the convolutional neural network, so as to obtain the starting abnormal point of the traceability chain.

[0174] The output module is used to aggregate a set of related anomalies at multiple time granularities based on the starting anomaly point of the traceability chain, determine whether the aggregation result shows that periodic fluctuation points dominate, and if so, output the final set of key anomaly points for cross-granularity related traceability.

[0175] Based on the spirit of this invention, those skilled in the art will readily conceive of a computer program product derived from the aforementioned method for tracing abnormal electricity marketing data based on multi-scale convolution. The computer program product may include a computer-readable storage medium on which computer-readable program instructions are loaded to enable a processor to implement various aspects of this disclosure. That is, this application also includes a terminal comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the aforementioned method for tracing abnormal electricity marketing data based on multi-scale convolution.

[0176] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0177] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0178] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A multi-scale convolution-based power marketing abnormal data tracing method, characterized in that, The method comprises the following steps: obtaining user daily electricity consumption curve data, applying a 3*3 convolution kernel to the user daily electricity consumption curve data through a convolutional neural network, extracting mutation point features for a single day scale, and obtaining a preliminary high-response region labeled curve subset; According to the preliminary high-response region labeled curve subset, a 7*7 convolution kernel is used to process the periodic fluctuation features under the week cycle scale to determine an expanded abnormal point set; Through the expanded abnormal point set, a 15*15 convolution kernel is applied to analyze the seasonal inflection point trend under the month season scale, and it is judged whether the seasonal inflection point trend deviates from the week mode by more than a preset threshold value. If it exceeds, the multi-time granularity information is fused to obtain a comprehensive key abnormal point set; Based on the comprehensive key abnormal point set, a clustering algorithm is used to group the mutation points and periodic fluctuation points in the high-response region to obtain a hierarchical labeled abnormal point subset; For the hierarchical labeled abnormal point subset, the cross-granularity correlation calculation is used to calculate the overlap degree of the high-response region between different scales to determine the key point path with the highest correlation strength; Based on the key point path, the feature map backtracking mechanism of the convolutional neural network is used to locate the seasonal inflection point source in the original curve to obtain the starting abnormal point of the traceability chain; According to the starting abnormal point of the traceability chain, the related abnormal point set under the multi-time granularity is aggregated, and it is judged whether the aggregation result shows that the periodic fluctuation point is dominant. If it shows, the final cross-granularity correlation traceability abnormal key point set is output.

2. The method of claim 1, wherein, The method further comprises the following steps: standardizing the preliminary high-response region labeled curve subset to obtain a daily electricity consumption sequence; using a convolutional neural network, applying a 3*3 convolution kernel to traverse the daily electricity consumption sequence, calculating the feature response value of each position, and obtaining a mutation feature response map; By setting a response intensity threshold, the mutation feature response map is scanned, and the positions with response values exceeding the threshold are marked to obtain a single-day mutation point set; Labeling the corresponding time points on the original daily electricity consumption sequence for the single-day mutation point set to generate labeled daily curves with mutation labels; Extracting curves containing at least one mutation label from all labeled daily curves to form a high-response region labeled curve subset data; Aggregating the high-response region labeled curve subset data according to the user identifier to obtain a preliminary high-response region labeled curve subset for each user.

3. The method of claim 2, wherein, The method further comprises the following steps: Obtaining the curve subset of the high-response region from the preliminary labeled data, extracting features for the data under the week cycle scale, and obtaining basic feature descriptions of periodic fluctuations; According to the basic feature descriptions of the periodic fluctuations, the points deviating from the conventional range are identified to form an initial abnormal point set, and it is judged whether there is a significant periodic deviation. If deviation is detected, the related point information is recorded; For the initial abnormal point set, the surrounding data points are analyzed using an expansion rule, and the expanded abnormal point set is determined in combination with the fluctuation features of the week cycle scale; By the extended anomaly point set, the characteristic distribution related to the weekly mode is analyzed to obtain a matching degree of the periodic mode, and it is judged whether the matching degree meets a preset relevance threshold value. If the matching degree meets the relevance threshold value, the point is marked as a relevant point; According to the marked relevant point, a fluctuation feature related to the weekly mode is extracted, and the feature is deeply analyzed to obtain an extended anomaly point set.

4. The method of claim 3, wherein, The application 15*15 convolution kernel analyzes the seasonal inflection point trend under the monthly seasonal scale, and further includes: According to the matching and expansion of the original anomaly point set and the weekly mode association rule, an extended anomaly point set is obtained; By sliding and scanning the 15*15 convolution kernel on the monthly seasonal sequence, a seasonal scale trend feature map in each time window is obtained; First-order difference and second-order difference operations are performed on the seasonal scale trend feature map to determine the seasonal inflection point position and its strength; For each seasonal inflection point position, a weekly mode association strength sequence in the corresponding time period is extracted; The mean square error between the seasonal inflection point strength sequence and the weekly mode association strength sequence is calculated to obtain a deviation degree value; If the deviation degree value exceeds a preset threshold value, the current seasonal inflection point is marked as a candidate key abnormal point, otherwise, it is discarded; The day granularity feature, the week granularity feature and the month granularity feature of all candidate key abnormal points are extracted and concatenated to obtain a comprehensive key abnormal point set.

5. The method of claim 4, wherein, The clustering algorithm is used to group the mutation points and periodic fluctuation points in the high response area, and further includes: From the comprehensive key abnormal point set, data is extracted, and point information in the high response area is preliminarily screened. A preset threshold value is used to distinguish the mutation points and the periodic fluctuations to obtain a preliminarily classified point set; For the preliminarily classified point set, a clustering method is used to group the mutation points and the periodic fluctuations. By analyzing the spatial distribution and time association between the points, an abnormal subset with hierarchical labeling is determined; According to the abnormal subset with hierarchical labeling, the grouping features of the mutation points and the periodic fluctuations are obtained. The time sequence of the point data in each group is compared to determine whether there is a significant fluctuation rule; If there is a significant fluctuation rule, the point data in the related group is deeply mined, and the correlation strength between the points is calculated to obtain a potential abnormal trigger mode; According to the potential abnormal trigger mode, the point groups in the high response area are further extracted. By comparing the fluctuation analysis results of different groups, the core abnormal points in the key set are determined; For the core abnormal points, the point information in the abnormal subset is structured and arranged according to the demand of traceability. The abnormal point subset with hierarchical labeling is obtained by region division.

6. The method of claim 5, wherein, The cross-granularity association is used to calculate the overlap degree between high response areas in different scales, and further includes: According to the abnormal point subset with hierarchical labeling, an abnormal point set at each level is obtained; A multi-scale convolutional neural network is used to extract a corresponding scale high response activation map in each level abnormal point set; The pixel-level intersection over union matrix is calculated for the high response activation maps of two adjacent levels to obtain an overlap distribution map; If the same position is higher than a preset threshold value in the intersection over union matrix for three consecutive layers, the position is marked as a high-intensity correlation point. A cross-scale correlation point connectivity graph is formed by connecting adjacent high-intensity correlation points of different levels; A graph search algorithm is used to find the highest-intensity path from the finest granularity to the coarsest granularity in the correlation point connectivity graph; The highest-intensity path is mapped back to the original abnormal point subset to obtain the key point path with the highest correlation intensity.

7. The method of claim 6, wherein, The feature map backtracking mechanism based on the convolutional neural network is further used to locate the seasonal inflection point source in the original curve, and further includes: A deep feature map is obtained by performing multi-layer feature extraction on the original time series through the convolutional neural network; A high response region is determined according to the strong activation position of the seasonal inflection point on the last layer feature map; A gradient backtracking method is used to perform reverse propagation from the high response region to the input layer to obtain an activation contribution map of each layer; Channel attention weights corresponding to the high response region in each layer feature map are obtained; A weighted activation map sequence is obtained by weighting each layer feature map through the channel attention weights; Path intensity accumulation calculation is performed on the weighted activation map sequence to obtain multiple backtracking paths from the output layer to the input layer and their intensities; The abnormal propagation main path with the highest intensity is obtained by sorting and screening the backtracking paths according to the path intensity; The starting abnormal point is determined by tracking the abnormal propagation main path to the corresponding position of the original time series.

8. The method of claim 7, wherein, The related abnormal point set under multiple time granularities is aggregated according to the starting abnormal point of the traceability chain, and whether the aggregation result shows that the periodic fluctuation point is dominant is determined, and further includes: The starting abnormal point is determined according to the traceability chain to obtain an initial abnormal point sequence, and the KMeans clustering algorithm is used to aggregate the initial abnormal point sequence under multiple time granularities to obtain multiple abnormal point sets; The periodic fluctuation point subset is determined by calculating the periodicity score of each abnormal point set, and the set with the highest proportion of periodic fluctuation point subset is extracted from the multiple abnormal point sets as the dominant periodic set; If the dominant periodic set has the highest proportion, all points in the dominant periodic set are marked as cross-granularity correlation candidate points to obtain a candidate key point sequence; The correlation predecessor node sequence is obtained by backtracking the candidate key point sequence under different time granularities according to the traceability chain, and the nodes that appear in at least two granularities are retained in the correlation predecessor node sequence through intersection operation, and the final abnormal key point set of cross-granularity correlation traceability is output.

9. A multi-scale convolution based power marketing abnormal data tracing system, characterized in that, Further includes: A curve subset acquisition module is configured to acquire user daily electricity consumption curve data, apply a 3*3 convolution kernel to the user daily electricity consumption curve data through a convolutional neural network, extract mutation point features for single-day scale, and obtain a curve subset with preliminary high response region labeling; A periodic fluctuation feature processing module is configured to determine an expanded abnormal point set by processing periodic fluctuation features under a week cycle scale using a 7*7 convolution kernel based on the curve subset with preliminary high response region labeling; A seasonal inflection point trend judgment module is configured to analyze seasonal inflection point trends under a month seasonal scale by applying a 15*15 convolution kernel to the expanded abnormal point set, determine whether the seasonal inflection point trend deviates from the week mode by more than a preset threshold, and if so, fuse multiple time granularity information to obtain a comprehensive key abnormal point set. The hierarchical labeling module is configured to group the mutation points and the periodic fluctuation points in the high-response area by using a clustering algorithm based on the comprehensive key abnormal point set, to obtain a hierarchically labeled abnormal point subset; The cross-granularity correlation calculation module is configured to determine a key point path with the highest correlation strength by calculating the overlap degree of the high-response areas at different scales through cross-granularity correlation calculation for the hierarchically labeled abnormal point subset; The feature map backtracking module is configured to locate a seasonal inflection point source in the original curve by using a feature map backtracking mechanism of a convolutional neural network based on the key point path, to obtain a starting abnormal point of a traceability chain; The output module is configured to aggregate relevant abnormal point sets at multiple time granularities according to the starting abnormal point of the traceability chain, to determine whether the aggregation result shows that the periodic fluctuation points are dominant, and if so, to output a final cross-granularity correlation traceability abnormal key point set.

10. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is configured to store instructions; The processor is configured to operate according to the instructions to perform the steps of the power marketing abnormal data traceability method based on multi-scale convolution according to any one of claims 1-8.

Images