Interactive data anomaly detection method and system adapted to power load management terminal, and storage medium

Abstract

The invention belongs to the technical field of anomaly detection, and particularly relates to an interactive data anomaly detection method and system adaptive to a power load management terminal and a storage medium. According to the method, normal historical user interactive data and abnormal historical user interactive data are processed by using a sequence pattern mining algorithm; and obtaining a corresponding user periodic interaction mode, training the machine learning model based on the periodic interaction mode to obtain an anomaly detection model, and during actual use, inputting user interaction data to be detected into the anomaly detection model to obtain an anomaly probability prediction value. Whether the to-be-detected user interaction data and the to-be-detected user interaction behavior are abnormal or not is judged according to the abnormal probability predicted value, the inherent periodic interaction rule of the user can be recognized, a high-quality standard is provided for anomaly detection, the anomaly detection precision is improved, hidden anomaly can be effectively recognized, and the user experience is improved. Data-level anomaly and behavior-level anomaly can be effectively detected at the same time.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of anomaly detection, and particularly relates to an interactive data anomaly detection method and system suitable for a power load management terminal and a storage medium. BACKGROUND

[0002] With the access of large-scale new energy equipment to the power grid, the uncertainty of both the source and the load is significantly enhanced, and the load regulation on the demand side has become an important means to ensure the stability of the new power system. In order to better manage the demand side, in recent years, the construction of new load management systems has been gradually carried out in an orderly manner, and a series of user-side terminals suitable for power load management have been deployed, including special transformer terminals (modular) and smart energy units. With the advancement of intelligent terminals and Internet of Things technology, user-side terminals (such as smart meters, e-commerce platform user clients, and smart home devices) have generated a large amount of time-series interactive data. Through in-depth analysis of these time-series interactive data, the user's power consumption habits can be understood, thereby improving service quality. However, the abnormal interactive data (such as interactive data caused by equipment failure, network attacks, or atypical operations) existing in these time-series interactive data seriously interferes with the in-depth analysis and service provision based on data, and therefore, anomaly detection needs to be performed on the interactive data of the terminal to improve data quality.

[0003] In the prior art, graph neural network models or K-nearest neighbor algorithms are usually used for anomaly detection. For example, the Chinese patent application file with publication number CN114077811A discloses a power Internet of Things device anomaly detection method based on a graph neural network. The traffic data and business data of different power Internet of Things devices are collected, the collected data is subjected to Koopman analysis, a graph structure of the power Internet of Things is constructed and input to a graph neural network model, the output of the graph neural network model is classified using K-Means clustering, and an anomaly detection result is obtained. However, when processing user interactive data of power load management, graph neural network models or K-nearest neighbor algorithms have poor adaptability, and generally have low recognition of hidden anomalies and low anomaly detection accuracy. SUMMARY

[0004] The purpose of the present application is to provide an interactive data anomaly detection method and system suitable for a power load management terminal and a storage medium, to solve the problem of low anomaly detection accuracy of existing anomaly detection methods for interactive data of a power load management terminal.

[0005] The application provides an interactive data anomaly detection method for an electric power load management terminal, which comprises the following steps: inputting user interactive data to be detected into an anomaly detection model to obtain an anomaly probability prediction value, and determining whether the user interactive data to be detected and user interactive behavior are abnormal according to the anomaly probability prediction value; the anomaly detection model is obtained by training a machine learning model using a user periodic interactive mode obtained from normal historical user interactive data and abnormal historical user interactive data; and the user periodic interactive mode used for training is obtained by processing the normal historical user interactive data and the abnormal historical user interactive data using a sequence pattern mining algorithm.

[0006] Further, the sequence pattern mining algorithm adopts a MAPD-PRO algorithm.

[0007] Further, the machine learning model is an XGBoost model.

[0008] Further, if the anomaly probability prediction value is greater than an anomaly determination threshold, it is determined that the user interactive data to be detected and the user interactive behavior are both abnormal.

[0009] Further, the user interactive data comprises a timestamp, an interactive type, an interactive frequency and an interactive intensity.

[0010] The technical scheme has the following beneficial effects: the application is an improved invention, and is used for user interactive data of a terminal for electric power load management; a sequence pattern mining algorithm is used to mine user periodic interactive modes with frequent occurrence and fixed time rules from the data; the normal and abnormal user interactive data are converted into the user periodic interactive modes according to the mode; a machine learning model is iteratively trained to obtain an anomaly detection model; in use, the user interactive data to be detected is directly input into the anomaly detection model for anomaly detection; the inherent periodic interactive rules of the user can be recognized; high-quality standards are provided for anomaly detection; the anomaly detection precision is improved; hidden anomalies can be effectively recognized; data-level anomalies and behavior-level anomalies can be simultaneously and effectively detected.

[0011] To solve the above technical problems, the application further provides an interactive data anomaly detection system for an electric power load management terminal, which comprises a processor configured to process user interactive data to be detected using an anomaly detection model to obtain an anomaly probability prediction value, and determine whether the user interactive data to be detected and user interactive behavior are abnormal according to the anomaly probability prediction value; the anomaly detection model is obtained by training a machine learning model using a user periodic interactive mode obtained from normal historical user interactive data and abnormal historical user interactive data; and the user periodic interactive mode used for training is obtained by processing the normal historical user interactive data and the abnormal historical user interactive data using a sequence pattern mining algorithm.

[0012] Further, the sequence pattern mining algorithm adopts a MAPD-PRO algorithm.

[0013] Further, the machine learning model is an XGBoost model.

[0014] Further, if the abnormal probability prediction value is greater than the abnormality determination threshold, it is determined that the user interaction data to be detected and the user interaction behavior are both abnormal.

[0015] To solve the above technical problems, the application further provides a storage medium, wherein the storage medium stores computer executable instructions, and the computer executable instructions are used to implement steps in an interactive data anomaly detection method for an electric power load management terminal when executed, and the method comprises the following steps: inputting user interaction data to be detected into an anomaly detection model to obtain an abnormal probability prediction value, and determining whether the user interaction data to be detected and a user interaction behavior are abnormal according to the abnormal probability prediction value; the anomaly detection model is obtained by training a machine learning model using a user periodic interaction pattern, and the user periodic interaction pattern is obtained by processing normal historical user interaction data and abnormal historical user interaction data using a sequence pattern mining algorithm.

[0016] Further, the sequence pattern mining algorithm adopts a MAPD-PRO algorithm.

[0017] Further, the machine learning model is an XGBoost model.

[0018] Further, if the abnormal probability prediction value is greater than the abnormality determination threshold, it is determined that the user interaction data to be detected and the user interaction behavior are both abnormal.

[0019] Further, the user interaction data comprises a timestamp, an interaction type, an interaction frequency and an interaction intensity.

[0020] The above technical solution has the following beneficial effects: the application is an improved invention, and is used for user interaction data of a terminal for electric power load management; a sequence pattern mining algorithm is used to mine frequent user periodic interaction patterns with fixed time rules from the data; normal and abnormal user interaction data are converted into user periodic interaction patterns according to the mode; a machine learning model is iteratively trained to obtain an anomaly detection model; when used, user interaction data to be detected is directly input into the anomaly detection model for anomaly detection; the inherent periodic interaction rules of the user can be identified; high-quality standards are provided for anomaly detection; the anomaly detection precision is improved; hidden anomalies can be effectively identified; data-level anomalies and behavior-level anomalies can be simultaneously effectively detected. Attached Figure Description

[0021] Figure 1 This is a flowchart of the interactive data anomaly detection process adapted to the power load management terminal in an embodiment of the method of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0023] This invention utilizes a sequence pattern mining algorithm to discover periodic user interaction patterns. Based on these patterns, a machine learning model is trained to obtain an anomaly detection model. The trained anomaly detection model is then used to process the user interaction data to be detected, thereby obtaining accurate anomaly detection results.

[0024] Method Implementation

[0025] This invention provides a method for detecting interactive data anomalies in power load management terminals, such as... Figure 1 As shown, it includes the following steps:

[0026] 1. Construct an anomaly detection model.

[0027] The anomaly detection model is trained on a machine learning model using periodic user interaction patterns obtained from both normal and abnormal historical user interaction data. These periodic interaction patterns are derived from sequence pattern mining algorithms that process both normal and abnormal historical user interaction data. This will be explained in detail below.

[0028] By using sequence pattern mining algorithms to process user interaction data, periodic user interaction patterns can be obtained.

[0029] Raw user interaction data is collected from terminals adapted for power load management. This data includes timestamps, interaction types, interaction frequency, and interaction intensity.

[0030] Sequence pattern mining algorithms are used to extract frequently occurring, periodic user interaction patterns with fixed time regularities from user interaction sequence databases, serving as a benchmark and basis for anomaly detection. A periodic user interaction pattern refers to a user's recurring behavior, such as frequently turning on the lights at 6:00 AM and the air conditioner at 6:30 AM. During the mining process, the sequence pattern mining algorithm initially uses turning on the lights at 6:00 AM as the initial interaction pattern based on frequency, and iteratively identifies related behavioral data such as turning on the lights and air conditioner, defining them as a periodic interaction pattern. Different interaction behavior parameters involve time, interaction type, interaction frequency, and interaction intensity.

[0031] As a preferred embodiment, the sequence pattern mining algorithm adopts a Multi-Agent Path Discovery-Priority-based Replanning Optimization (MAPD-PRO) algorithm, which is used to mine the periodic interaction patterns of users based on the MAPD-PRO algorithm, which is described below.

[0032] 1) Preprocess and serialize the original user interaction data to construct a user interaction sequence dataset .

[0033] where each user interaction sequence represents a certain interaction behavior of a group of users within a period of time, and D represents the total number of sequences. Each user interaction sequence includes a timestamp, an interaction type, an interaction frequency, and an interaction intensity, etc.

[0034] 2) Construct an initial frequent pattern.

[0035] ① Traverse each behavior data a in the user interaction sequence S i and record its time coordinate. Taking the above periodic interaction pattern containing the user's habit of turning on the light at 6 o'clock and turning on the air conditioner at 6:30 as an example, the behavior data a refers to the user interaction data corresponding to turning on the light. Recording the time coordinate is used to determine whether the behavior data is a periodic behavior, such as the user turning on the light at around 6 o'clock for three days a week, which is a periodic behavior.

[0036] ② For each interaction behavior a identified within a specific time period unit Cell[t], establish it as an initial frequent pattern to be mined P=a, construct the data structure Forest[i] (forest) of the pattern, and create a mining starting point (root node). The Number of Root Paths (NRP) is initialized to 1, indicating that one possible starting point has been found to construct more complex user behavior sequences.

[0037] ③ Count the frequency of each behavior a in the entire dataset (i.e., the user interaction sequence dataset constructed from the input original interaction data):

[0038]

[0039] where, represents the frequency of behavior a in the entire dataset, represents the sum of the number of root paths of all nodes in the data structure.

[0040] ④ Set a minimum frequency threshold ρ, if:

[0041]

[0042] If a is judged as an initial frequent pattern P, it is stored in the frequent pattern set A.

[0043] 3) Depth-first extending periodic interaction patterns and setting flexible gap constraints.

[0044] For a known frequent pattern P, try to extend other related or continuous behaviors to form a longer candidate pattern Again traverse the user interaction sequence dataset S D , get the net tree of the frequent pattern P , introduce periodic gap constraint mechanism , search for part of records in sequence S i that meet the gap constraints as candidate patterns , that is, allow a flexible and reasonable time interval between two continuous behaviors, so that the pattern sequence is more consistent with the continuity and diversity of real user behavior, and avoid mechanical matching. is the minimum value of the periodic gap, is the maximum value of the periodic gap, and these two values are constantly adjusted according to the results of mining to finally determine the appropriate periodic gap constraint, so that the results of mining are the most accurate.

[0045] Create a node q in the net tree under the candidate pattern . Calculate the NRP of q, that is, the total number of tree root paths of . That is, the frequency of the super pattern :

[0046]

[0047] If there is:

[0048]

[0049] Add to the frequent pattern set A of the user interaction sequence.

[0050] 4) Termination condition and result output.

[0051] Repeat the above extension process, and each newly discovered frequent pattern will be used as a new starting point to continue exploring longer behavior sequences until all super patterns of the current user interaction sequence pattern P are processed. The frequent pattern set A is a complete user periodic interaction pattern library containing various sequences from simple to complex.

[0052] ​According to the above process, the normal historical user interaction data and the abnormal historical user interaction data are processed by using the sequence pattern mining algorithm to obtain corresponding periodic interaction patterns. The periodic interaction patterns are used to train and evaluate the performance of the machine learning model, and the model is trained iteratively to accurately detect the interaction abnormal data.

[0053] The machine learning model converts the periodic interaction patterns of each user into a machine learning feature vector to identify the features of the user interaction data. The conversion method is as follows: for any piece of user interaction data, at least one of the occurrence number, intensity and frequency of each periodic interaction pattern contained in the user interaction data is calculated, and a feature vector is formed based on the calculated at least one of the occurrence number, intensity and frequency of each periodic interaction pattern, the feature vector being a multi-dimensional feature vector.

[0054] The machine learning model can use support vector machines or random forests, etc. As a preferred embodiment, the machine learning model uses an eXtreme Gradient Boosting (XGBoost) model.

[0055] The XGBoost model is integrated by multiple decision trees, and the weighted sum of the prediction results of all decision trees is the final prediction output of the XGBoost model. That is, for the feature vector x i of the i-th user, the prediction value is represented as:

[0056]

[0057] Each decision tree is responsible for a part of the judgment of the user behavior features, and the conclusions of all trees are integrated to form the final comprehensive judgment. When constructing each decision tree, a greedy strategy is used for node splitting. For each node to be split, the algorithm traverses all user behavior features and all possible split points. The gain brought by each split method is calculated, which reflects the improvement of the distinction between normal and abnormal user behaviors after the feature is split. Finally, the feature and the split point that can bring the maximum gain are selected for node splitting, thereby ensuring the efficiency and effectiveness of the model learning.

[0058] When the XGBoost model is iteratively trained, in each iteration, the newly added decision tree does not directly learn the original label, but fits the residual between the current model prediction and the true label. Specifically, the newly built tree in the m-th round learns the error of the model in the previous m-1 rounds in judging the user behavior. In this way, each new tree is committed to correcting the mistakes made by the previous stage model in distinguishing normal and abnormal behaviors, so that the model prediction ability gradually approaches the true situation.

[0059] The XGBoost model introduces a regularization mechanism, and the loss function is expressed as:

[0060]

[0061] wherein, represents an error function of the real abnormal value and the predicted value; represents a regularization term, and N is the number of data, specifically:

[0062]

[0063] wherein, and both represent adjustment parameters; and respectively represent the number of leaf nodes and the leaf node weight of the decision tree. This mechanism can effectively control the complexity of the model, and ensure that the trained detection model still has strong discriminant ability when facing new and unseen user interaction data, that is, has good generalization.

[0064] 2. The user interaction data to be detected is input into the anomaly detection model to obtain an abnormal probability prediction value, and whether the user interaction data to be detected and the user interaction behavior are abnormal is judged according to the abnormal probability prediction value.

[0065] For a new user interaction data, the user interaction data is input into the trained XGBoost detection model, and the model outputs an abnormal probability prediction value , and whether it is abnormal is judged according to an abnormality judgment threshold When , it is judged that the user interaction data to be detected and the user interaction behavior are both abnormal.

[0066] System implementation

[0067] An interactive data anomaly detection system suitable for a power load management terminal of the application comprises a processor, and the processor is used to realize an interactive data anomaly detection method suitable for a power load management terminal.

[0068] The method comprises the following steps: inputting user interaction data to be detected into the processor, processing the user interaction data to be detected by using an anomaly detection model by the processor to obtain an abnormal probability prediction value, and judging whether the user interaction data to be detected and the user interaction behavior are abnormal according to the abnormal probability prediction value.

[0069] The abnormality detection model is trained by using normal historical user interaction data and abnormal historical user interaction data to obtain a user periodic interaction pattern. The user periodic interaction pattern used in the training is obtained by processing the normal historical user interaction data and the abnormal historical user interaction data by using a sequential pattern mining algorithm. Details are described below.

[0070] S1, processing the user interaction data by using a sequential pattern mining algorithm to obtain a user periodic interaction pattern.

[0071] Raw user interaction data is collected from terminals adapted to power load management. The user interaction data includes a timestamp, an interaction type, an interaction frequency, and an interaction intensity.

[0072] The sequential pattern mining algorithm is used to mine frequent user periodic interaction patterns with fixed time rules from the user interaction sequence database, which serves as a reference and basis for abnormality detection.

[0073] As a preferred embodiment, the sequential pattern mining algorithm adopts a Multi-Agent Path Discovery-Priority-based Replanning Optimization (MAPD-PRO) algorithm. The user periodic interaction pattern is mined based on the MAPD-PRO algorithm, which is described below.

[0074] 1) Preprocessing and sequencing the raw user interaction data to construct a user interaction sequence dataset .

[0075] Each user interaction sequence represents a certain interaction behavior of a group of users within a period of time, and D represents the total number of sequences.

[0076] 2) Constructing an initial frequent pattern.

[0077] ① Traverse each behavior data a in the user interaction sequence S i and record its time coordinate.

[0078] ② For each interaction behavior a identified within a specific time period unit Cell[t], establish it as an initial frequent pattern P=a to be mined, construct the data structure Forest[i] (net tree) of the pattern, and create a mining starting point (root node). The Number of Root Paths (NRP) is initialized to 1, indicating that one possible starting point has been found to construct more complex user behavior sequences.

[0079] ③ Count the frequency of each behavior a in the whole dataset:

[0080]

[0081] where, represents the frequency of behavior a in the whole dataset, represents the sum of the tree root path numbers of all nodes in the data structure.

[0082] ④ Set the minimum frequency threshold ρ, if:

[0083]

[0084] then judge a as the initial frequent pattern P and store it in the frequent pattern set A.

[0085] 3) Depth-first expansion of periodic interaction patterns and setting of flexible gap constraints.

[0086] For a known frequent pattern P, try to extend other related or continuous behaviors to form a longer candidate pattern . Again traverse the user interaction sequence dataset S D , get the net tree of the frequent pattern P , introduce the periodic gap constraint mechanism , search for the part of the record in the sequence S i that meets the interval constraint as the candidate pattern , that is, allow a flexible and reasonable time interval between two consecutive behaviors, so that the pattern sequence is more consistent with the continuity and diversity of real user behavior, and avoid mechanical matching. is the minimum value of the periodic gap, is the maximum value of the periodic gap, and these two values are constantly adjusted according to the results of mining to finally determine the appropriate periodic gap constraint, so that the results of mining are the most accurate.

[0087] Create a node q in the net tree under the candidate pattern . Calculate the NRP of q, which is the total number of tree root paths of . That is, the frequency of the super-pattern :

[0088] If there is:

[0089]

[0090]

[0091] then add to the frequent pattern set A of the user interaction sequence. ​

[0092] 4) termination condition and result output.

[0093] The above expansion process is repeated, and each newly discovered frequent pattern is used as a new starting point to continue exploring longer behavior sequences until all super patterns of the current user interaction sequence pattern P are processed The frequent pattern set A is a complete library of user periodic interaction patterns containing various sequences from simple to complex.

[0094] S2: According to the above process, the sequence pattern mining algorithm is used to process normal historical user interaction data and abnormal historical user interaction data to obtain corresponding periodic interaction patterns. The periodic interaction patterns are used to train and evaluate the performance of the machine learning model, and the model is trained iteratively to accurately detect interaction abnormal data.

[0095] The machine learning model converts each user periodic interaction pattern into a machine learning feature vector to identify the characteristics of the user interaction data. The conversion method is as follows: for any piece of user interaction data, at least one of the occurrence number, intensity, and frequency of each user periodic interaction pattern contained in the user interaction data is calculated, and a feature vector is formed based on the calculated at least one of the occurrence number, intensity, and frequency of each user periodic interaction pattern. The feature vector is a multi-dimensional feature vector.

[0096] The machine learning model can use support vector machines or random forests, etc. As a preferred embodiment, the machine learning model uses an eXtreme Gradient Boosting (XGBoost) model.

[0097] The XGBoost model is integrated by multiple decision trees, and the weighted sum of the prediction results of all decision trees is the final prediction output of the XGBoost model. That is, for the feature vector x i of the i-th user, the prediction value is represented as:

[0098]

[0099] Each decision tree is responsible for a part of the judgment of user behavior characteristics, and the conclusions of all trees are integrated to form the final comprehensive judgment. When constructing each decision tree, a greedy strategy is used for node splitting. For each node to be split, the algorithm traverses all user behavior characteristics and all possible split points. The gain brought by each split method is calculated, which reflects the improvement of the distinction between normal and abnormal user behavior after splitting by this feature. Finally, the feature and split point that can bring the maximum gain are selected for node splitting, thereby ensuring the efficiency and effectiveness of model learning.

[0100] In the iterative training of the XGBoost model, in each iteration, the newly added decision tree does not directly learn the original label, but fits the residual between the current model prediction and the real label. Specifically, the newly built tree in the mth round learns the error of the model in the previous m-1 rounds in judging the user behavior. In this way, each new tree is committed to correcting the mistakes made by the previous stage model in distinguishing normal and abnormal behavior, so that the model prediction ability gradually approaches the real situation.

[0101] The XGBoost model introduces a regularization mechanism, and the loss function is expressed as:

[0102]

[0103] wherein, represents the error function of the real abnormal value and the predicted value; represents the regularization term, and N is the number of data, specifically:

[0104]

[0105] wherein, and both represent adjustment parameters; and respectively represent the number of leaf nodes and the weight of leaf nodes of the decision tree. This mechanism can effectively control the complexity of the model and ensure that the trained detection model still has strong discriminant ability when facing new and unseen user interaction data, that is, good generalization.

[0106] For a new user interaction data, the user interaction data is input into the trained XGBoost detection model, and the model outputs an abnormal probability prediction value According to the abnormality judgment threshold whether the user interaction data and the user interaction behavior are abnormal is judged.

[0107] Storage medium implementation

[0108] The storage medium of the application, the storage medium stores computer executable instructions, the computer executable instructions are used when being executed to realize the steps in the interactive data anomaly detection method suitable for the power load management terminal. The method comprises:

[0109] The user interaction data to be detected is input into the anomaly detection model to obtain an abnormal probability prediction value, and whether the user interaction data to be detected and the user interaction behavior are abnormal is judged according to the abnormal probability prediction value.

[0110] ​The abnormality detection model is trained by using normal historical user interaction data and abnormal historical user interaction data to obtain a user periodic interaction mode.

[0111] Preferably, the sequence pattern mining algorithm adopts a MAPD-PRO algorithm.

[0112] The application uses a sequence pattern mining algorithm to mine a user periodic interaction mode, and uses the mined periodic interaction mode to train a machine learning model (such as an XGBoost model) to enable the XGBoost model to effectively detect data-level anomalies and behavior-level anomalies at the same time, with high detection accuracy. In the MAPD-PRO algorithm, the introduction of a periodic gap constraint enables more flexible and more accurate mining of a user's real, elastic time rule periodic behavior mode, providing high-quality standards for anomaly detection. The XGBoost model supports incremental updating, and when new user interaction data is added, the periodic interaction behavior mode library and the detection model can be dynamically updated without re-global calculation, with strong self-adaptation, adaptation to concept drift, and meeting the continuous needs of practical applications.

Claims

1. A method for detecting interactive data anomalies in a power load management terminal, characterized in that, include: The user interaction data to be detected is input into the anomaly detection model to obtain the anomaly probability prediction value. Based on the anomaly probability prediction value, it is determined whether the user interaction data and user interaction behavior to be detected are abnormal. The anomaly detection model is obtained by training a machine learning model with user periodic interaction patterns obtained from normal and abnormal historical user interaction data. The user periodic interaction patterns used during training are obtained by processing normal and abnormal historical user interaction data using a sequence pattern mining algorithm.

2. The interactive data anomaly detection method for power load management terminals according to claim 1, characterized in that, The sequence pattern mining algorithm uses the MAPD-PRO algorithm.

3. The interactive data anomaly detection method for power load management terminals according to claim 1 or 2, characterized in that, The machine learning model is the XGBoost model.

4. The interactive data anomaly detection method for power load management terminals according to claim 1 or 2, characterized in that, If the predicted probability of anomalies is greater than the anomaly determination threshold, then both the user interaction data and user interaction behavior to be detected are determined to be abnormal.

5. The interactive data anomaly detection method for power load management terminals according to claim 1, characterized in that, The user interaction data includes timestamps, interaction types, interaction frequency, and interaction intensity.

6. An interactive data anomaly detection system adapted to a power load management terminal, comprising a processor, characterized in that, The processor is used to process the user interaction data to be detected using an anomaly detection model to obtain an anomaly probability prediction value, and to determine whether the user interaction data and user interaction behavior to be detected are abnormal based on the anomaly probability prediction value. The anomaly detection model is obtained by training a machine learning model using the user periodic interaction pattern obtained from normal historical user interaction data and abnormal historical user interaction data. The user periodic interaction pattern used during training is obtained by processing normal historical user interaction data and abnormal historical user interaction data using a sequence pattern mining algorithm.

7. The interactive data anomaly detection system adapted to power load management terminals according to claim 6, characterized in that, The sequence pattern mining algorithm uses the MAPD-PRO algorithm.

8. The interactive data anomaly detection system adapted to a power load management terminal according to claim 6 or 7, characterized in that, The machine learning model is the XGBoost model.

9. The interactive data anomaly detection system adapted to a power load management terminal according to claim 6 or 7, characterized in that, If the predicted probability of anomalies is greater than the anomaly determination threshold, then both the user interaction data and user interaction behavior to be detected are determined to be abnormal.

10. A storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are used, when executed, to implement the steps in the interactive data anomaly detection method for an adaptive power load management terminal as described in any one of claims 1-5.

Images