Power load abnormality early warning method and device combined with power consumption behavior big data mining

Abstract

The application discloses a power load abnormality early warning method and device combined with power consumption behavior big data mining, and relates to the technical field of smart power grids.The method comprises the following steps: collecting multi-source heterogeneous data of a composite power unit, generating power consumption behavior data sequences and data correlation relationships of each power consumption entity; performing multi-dimensional feature extraction on the power consumption behavior data sequences, and combining the data correlation relationships to establish a benchmark behavior portrait set and a response behavior portrait set of each type of power consumption entity; training a first model for detecting internal abnormalities of the entity based on the benchmark behavior portrait set, and training a second model for evaluating abnormalities between power consumption entities based on the response behavior portrait set; acquiring a monitoring data set in real time, extracting real-time feature vectors, inputting the real-time feature vectors into the first model and the second model for abnormality identification, and triggering a hierarchical early warning according to an identification result.The application realizes accurate positioning, early discovery and associated risk cooperative prevention and control of abnormality early warning.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and more specifically to a method and device for early warning of abnormal power loads by combining big data mining of electricity consumption behavior. Background Technology

[0002] With the development of smart grids, power load anomaly early warning technology has evolved from alarms based on fixed thresholds to a model that uses historical data to construct load baselines and determine anomalies. Current methods collect total load time-series data of users or lines, use machine learning algorithms to train normal behavior models, and identify anomalies by detecting deviations between real-time data and the baseline. This approach primarily serves the overall anomaly monitoring of distribution networks or individual users.

[0003] However, existing technologies have significant limitations. On the one hand, current methods typically treat a single metering point as an indivisible monitoring object, resulting in overly coarse granularity and an inability to effectively distinguish multiple independent electrical entities that may be included under the same metering point. On the other hand, most existing methods only focus on macroscopic changes in total load, lacking in-depth correlation analysis of the underlying multiple electrical quantities constituting the load, such as voltage and current waveforms and phase relationships. Consequently, it is difficult to define and characterize the unique behavioral patterns of different electrical entities, leading to relatively singular early warning targets. This makes it difficult to effectively warn of early performance degradation of specific electrical entities, and even more difficult to identify the cascading safety risks that may be caused by a single entity failure. Ultimately, when the system detects an anomaly in total load, the lack of refined behavioral correlation and location capabilities makes it difficult to quickly and accurately trace back to the specific equipment or electrical behavior that caused the anomaly, resulting in low fault diagnosis efficiency. Summary of the Invention

[0004] This invention addresses the technical problems in existing technologies, such as coarse granularity of monitoring objects, lack of behavioral connotation, single early warning target, and vague location of abnormal root causes, by providing a method and device for early warning of abnormal power loads that combines big data mining of electricity consumption behavior.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] In a first aspect, the present invention provides a method for early warning of abnormal power loads by combining big data mining of electricity consumption behavior, including:

[0007] Based on regional big data, multi-source heterogeneous data of composite power units are collected, and the electricity consumption behavior data sequence and data correlation relationship of each electricity-consuming entity are analyzed and obtained.

[0008] Multidimensional features are extracted from the electricity consumption behavior data sequence of each electricity-consuming entity, and the correlation between the multidimensional feature extraction results and the data is analyzed to establish a baseline behavior profile set and a response behavior profile set for each type of electricity-consuming entity.

[0009] A first model for detecting anomalies within an entity is trained based on the baseline behavior profile set, and a second model for evaluating anomalies between electricity-consuming entities is trained based on the response behavior profile set.

[0010] The monitoring dataset of the composite power unit is obtained in real time. The first model and the second model are combined to identify power load anomalies, and an early warning is triggered based on the power load anomaly identification results.

[0011] Secondly, the present invention provides an early warning device for abnormal power load based on big data mining of electricity consumption behavior, comprising:

[0012] The data acquisition and analysis module is used to collect multi-source heterogeneous data of composite power units based on regional big data, and analyze and obtain the data sequence of electricity consumption behavior and data correlation of each electricity-consuming entity.

[0013] The feature extraction and profile building module is used to perform multi-dimensional feature extraction on the electricity consumption behavior data sequence of each electricity-consuming entity, and combine the multi-dimensional feature extraction results with the data association to analyze and establish a baseline behavior profile set and a response behavior profile set for each type of electricity-consuming entity.

[0014] The model training module is used to train a first model for detecting anomalies within entities based on the baseline behavior profile set, and to train a second model for evaluating anomalies between electricity-consuming entities based on the response behavior profile set.

[0015] The real-time monitoring and early warning module is used to monitor and acquire the monitoring dataset of the composite power unit in real time, combine the first model and the second model to identify power load anomalies, and trigger early warnings based on the power load anomaly identification results.

[0016] The beneficial effects of this invention are:

[0017] Compared to existing technologies, this invention first achieves precise location of anomaly warnings, refining the monitoring and warning targets from coarse-grained metering points to every specific power-consuming entity inside and outside the composite power unit. This effectively distinguishes entity types, thereby improving the efficiency of troubleshooting root causes. Secondly, this invention can warn of early performance degradation of the power unit's own equipment, simultaneously monitor the abnormal states of external service objects, and assess the potential cascading risks, constructing a proactive internal and external collaborative security defense system. Thirdly, by mining and utilizing multi-dimensional behavioral characteristics reflecting the deep operating state of equipment, it can issue warnings in the early stages of performance degradation or during the latent fault phase, achieving early warning timing. Finally, by training and jointly applying an internal entity anomaly detection model and an inter-entity correlation anomaly assessment model, a multi-level, collaborative risk identification and warning mechanism is formed, improving the safety and reliability of power system operation. Attached Figure Description

[0018] Figure 1 A flowchart illustrating the power load anomaly early warning method combining big data mining of electricity consumption behavior provided by this invention;

[0019] Figure 2 A schematic diagram of the power load anomaly early warning device that combines big data mining of electricity consumption behavior provided by the present invention.

[0020] In the attached diagram, the components represented by each number are as follows:

[0021] Data acquisition and analysis module 11, feature extraction and profile construction module 12, model training module 13, real-time monitoring and early warning module 14. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0025] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for early warning of abnormal power load by combining big data mining of electricity consumption behavior, including:

[0026] S10: Based on regional big data, collect multi-source heterogeneous data of composite power units, and analyze and obtain the data sequence of electricity consumption behavior and data correlation of each electricity-consuming entity;

[0027] The multi-source heterogeneous data includes total incoming line electrical quantity data, electrical quantity data of each branch circuit, equipment status data, and physical topology data.

[0028] Specifically, based on regional big data, multi-source heterogeneous data of composite power units are collected, and the electricity consumption behavior data sequence of each electricity-consuming entity is analyzed and obtained, including:

[0029] By combining big data methods and deploying data sensing layer devices, multi-source sensing data from multiple monitoring points in several composite power units within a preset area is collected.

[0030] Acquire prior entity topology data of the composite power unit, and merge it with the multi-source sensing data to output the multi-source heterogeneous data;

[0031] Based on the prior entity topology data, an association map of composite power units is constructed, and the multi-source sensing data is decomposed and associated according to the association map and the energy conservation relationship to generate an independent electricity consumption behavior data sequence and data association relationship for each electricity-consuming entity.

[0032] First, data sensing layer devices are deployed within a pre-defined area, utilizing big data methods. This pre-defined area refers to the geographical scope or partition of the power network requiring refined load anomaly early warning management, such as a microgrid in an industrial park containing energy storage power stations, distributed photovoltaic systems, and electric vehicle charging facilities. The data sensing layer devices are physical sensors and intelligent acquisition terminals installed at key nodes of the power network to directly measure electrical and non-electrical parameters, such as high-precision power quality monitoring devices, intelligent measurement and control units, and gateways supporting specific communication protocols. The deployed data sensing layer devices target multiple monitoring points within several composite power units within this pre-defined area, continuously collecting multi-source sensing data.

[0033] Specifically, the multi-source sensing data includes total incoming line electrical quantity data, electrical quantity data of each branch circuit, equipment status data, and physical topology data. Preferably, the total incoming line electrical quantity data includes voltage, current, and power timing information measured by the main switch meter of the power unit. The electrical quantity data of each branch circuit includes voltage and current waveform data of key nodes such as the DC and AC sides of the energy storage converter and the output terminal of the charging pile, as well as switch status signals. The equipment status data includes voltage, temperature, state of charge, and health parameters of the battery clusters or modules, as well as requested voltage, requested current, and battery parameter information in the communication messages between the charging pile and the electric vehicle battery management system. The physical topology data refers to the predefined power station equipment connection diagram and the physical layout diagram of the charging pile.

[0034] Secondly, prior entity topology data of the composite power unit is acquired. This prior entity topology data describes the inherent connections and relationships between various power-consuming entities within the power unit. The multi-source sensing data collected above is merged with this prior entity topology data to form a complete multi-source heterogeneous dataset containing time-stamped synchronization information.

[0035] Finally, based on prior entity topology data, an association graph of the composite power unit is constructed. This association graph uses nodes to represent specific power-consuming entities, such as specific energy storage units or DC charging piles, and edges to represent electrical connections or physical adjacencies between entities. Based on the constructed association graph and energy conservation relationships, signal decomposition techniques are applied to process the aforementioned multi-source sensing data. Specifically, this process combines the waveform characteristics of each branch current with information such as switching events to decompose the overall load data at the main incoming line layer by layer, and accurately associates it with the corresponding specific entity nodes in the graph according to topological relationships. Through the decomposition and association process, a unique power consumption behavior data sequence reflecting the operating characteristics of each independent power-consuming entity is generated, while the data association relationships between different power-consuming entities are clearly defined and recorded.

[0036] Ultimately, each electricity-consuming entity obtains an independent electricity consumption behavior data sequence and data association. Specifically, the electricity consumption behavior data sequence is used to characterize the complete electrical measurement information of the electricity-consuming entity over time, reflecting its own operating characteristics, and representing the unique load characteristics, operating modes, and health status evolution process of the electricity-consuming entity; the data association is used to characterize the interactions between the electricity-consuming entity and other electricity-consuming entities, such as electrical connections, energy interactions, or physical proximity, and represents the potential mutual influence paths and risk transmission mechanisms between entities.

[0037] S20: Perform multi-dimensional feature extraction on the electricity consumption behavior data sequence of each electricity-consuming entity, and combine the multi-dimensional feature extraction results with the data association relationship to analyze and establish a baseline behavior profile set and a response behavior profile set for each type of electricity-consuming entity.

[0038] Specifically, multi-dimensional feature extraction is performed on the electricity consumption behavior data sequence for each electricity-consuming entity, including:

[0039] Based on the energy source of the power-consuming entity and its relationship within the composite power unit, the power-consuming entities are divided into external power-consuming entities and internal power-consuming entities.

[0040] Based on the early warning mission objectives of the composite power unit, the management depth of the power consumption entities of the composite power unit is defined, and the classification basis of the power consumption entities is determined accordingly.

[0041] Based on the classification criteria for electricity-consuming entities, a multi-level cascaded index relationship for electricity-consuming entity classification is constructed.

[0042] Specifically, before performing multi-dimensional feature extraction on the electricity consumption behavior data sequence of each electricity-consuming entity, a preprocessing step of entity classification and index construction is required. This is to establish a clear analytical framework and object scope for subsequent feature extraction, cluster analysis, and behavior profiling, ensuring that subsequent technical steps can be carried out in a targeted and hierarchical manner based on the entity's physical attributes and functional roles. This will improve the representativeness of behavioral features, the accuracy of behavioral profiling, and the precision of the final anomaly detection model.

[0043] First, based on the energy source of the electricity-consuming entities and their relationship within the composite power unit, all electricity-consuming entities are divided into two main categories: external electricity-consuming entities and internal electricity-consuming entities. Internal electricity-consuming entities refer to the equipment that constitutes the infrastructure of the composite power unit itself, such as energy storage converters and battery clusters; external electricity-consuming entities refer to external equipment or loads connected to the unit and receiving its power, such as electric vehicles and their connected charging stations.

[0044] Secondly, based on the early warning objectives to be achieved by the composite power unit, such as focusing on equipment health management, user service security, or overall risk prevention and control, the depth of power entity management for this composite power unit needs to be clearly defined. Specifically, the depth of power entity management is a predefined analysis level parameter that determines the scope of entity levels to be covered by behavior monitoring and feature analysis. It specifically indicates whether the analysis work starts from the overall incoming line level, delves into the main subsystem level, or is further refined to each independent equipment unit or even load port.

[0045] Based on the determined depth of power consumption entity management, the classification criteria for power consumption entities can be determined accordingly. This classification criteria is a set of rules and standards used to divide and group different power consumption entities, which can comprehensively consider multiple dimensions such as entity type, electrical level, functional role, or safety level. For example, in scenarios where the depth requirement is refined to the level of independent equipment units, the classification criteria for power consumption entities can stipulate that the power station's own equipment and external service equipment should be distinguished first, and then classified into subcategories such as energy storage converters, battery clusters, and DC charging piles according to equipment type, and can be further grouped according to electrical circuits or physical locations.

[0046] Finally, based on the established classification criteria for electricity-consuming entities, a multi-level cascaded classification index relationship for electricity-consuming entities is constructed. This classification index relationship adopts a tree or network structure, hierarchically organizing and grouping all electricity-consuming entities according to their subordinate, parallel, or functional relationships.

[0047] For example, "energy storage power station" can be set as the root node of the index relationship, and its next level can be divided into main branches such as "energy storage system" and "charging system". On this basis, the "energy storage system" branch can be further refined into specific equipment nodes such as "energy storage converter #1" and "battery cluster #01", while the "charging system" branch can be refined into terminal load nodes such as "DC charging pile #A03", thus forming a clear entity hierarchy structure from root to leaf.

[0048] The core purpose of constructing this electricity-consuming entity classification index is to provide a unified and clear structured framework for subsequent multi-dimensional feature clustering analysis and behavioral profiling based on entity categories. This ensures that the feature extraction and model training processes can implement targeted and differentiated processing based on the physical attributes, electrical characteristics, and functional roles of different categories of electricity-consuming entities. For example, for device nodes classified as internal electricity-consuming entities in the electricity-consuming entity classification index, subsequent feature extraction can focus on indicators that directly reflect the device's own operational health status, such as charging and discharging efficiency and the trend of total harmonic distortion of current. For nodes classified as external electricity-consuming entities, the focus can be on indicators that characterize their external electricity consumption behavior patterns and safety status, such as the morphological characteristics of the charging curve and the rate of change of battery management system parameter requests.

[0049] Specifically, multi-dimensional features are extracted from the electricity consumption behavior data sequence of each electricity-consuming entity, and the correlation between the multi-dimensional feature extraction results and the data is analyzed to establish a baseline behavior profile set and a response behavior profile set for each type of electricity-consuming entity, including:

[0050] Based on the prior feature list, the electricity consumption behavior data sequence of each electricity-consuming entity is traversed, multi-dimensional feature sequences are extracted, and the multi-dimensional feature extraction result is output.

[0051] Based on the preset electricity consumption entity classification index relationship, the multidimensional feature extraction results are classified and divided to form several multidimensional category feature clusters;

[0052] Cluster analysis is performed on each of the multidimensional category feature clusters, and fitting processing is performed on the multiple multidimensional category feature sequences corresponding to each cluster in the cluster analysis results to obtain the baseline behavior profile set of each type of electricity-consuming entity.

[0053] First, based on a prior feature list, the electricity consumption behavior data sequence of each electricity-consuming entity is traversed and processed. Specifically, this feature list includes various feature calculation rules from the time domain, frequency domain, and energy domain, aiming to quantify different aspects of electricity consumption behavior. For each electricity-consuming entity's independent voltage and current data stream, the corresponding multi-dimensional feature sequence is extracted according to this feature list.

[0054] For example, for power-consuming entities that are part of the power station's own equipment, extracted features might include the charging and discharging efficiency of the energy storage converter, the trend of the total harmonic distortion rate of the output current over time, the standard deviation of voltage between different battery clusters, and the rate of battery temperature rise under the same charge and discharge rate. For power-consuming entities that are external service recipients, extracted features might include the overall shape of the charging curve, specific quantitative indicators such as the duration of the constant current charging phase and the starting point of the constant voltage charging phase, the rate of change of parameters requested by the battery management system, the maximum difference in individual battery cell voltage at the end of charging, and the total energy consumption of a single charging session. The feature sequences extracted from all entities together constitute the multidimensional feature extraction result.

[0055] Secondly, based on the pre-defined electricity entity classification index, the aforementioned multidimensional feature extraction results are divided and aggregated. This electricity entity classification index clearly defines the category and level to which different entities belong. According to this electricity entity classification index, the multidimensional feature sequences of multiple entities belonging to the same category or having similar attributes can be aggregated and summarized to form several multidimensional category feature clusters. Each multidimensional category feature cluster represents the complete set of feature data presented by a specific category of electricity entity during its historical operation.

[0056] Finally, cluster analysis is performed on each of the resulting multidimensional category feature clusters. Specifically, this cluster analysis employs unsupervised learning methods, such as hierarchical clustering algorithms based on dynamic time warping distance, aiming to automatically discover inherent typical patterns from the behavioral characteristics of numerous electricity-consuming entities within the same category.

[0057] First, cluster analysis further divides a multidimensional category feature cluster into several finer-grained clusters. The feature sequences within each cluster exhibit high similarity in morphology and numerical range. Second, for each cluster in the cluster analysis results, its corresponding multidimensional feature sequences are fitted. This fitting process aims to extract and summarize the core statistical characteristics and variation range of the feature sequences within the cluster, such as calculating the mean vector, variance range, or probability distribution of each feature dimension. Through this process, each cluster is abstracted and represented as a baseline behavioral profile, which defines the normal numerical range and variation law of each feature dimension under the typical behavioral pattern of that type. The set of all baseline behavioral profiles derived from the same type of electricity-consuming entity constitutes the baseline behavioral profile set of that type of electricity-consuming entity.

[0058] For example, by performing cluster analysis on all electric vehicle charging characteristic clusters, several baseline behavioral profiles such as "typical fast charging" and "slow charging replenishment" can be obtained; by analyzing the operating characteristic clusters of energy storage units, baseline behavioral profiles such as "automatic generation control frequency regulation mode" and "peak shaving and valley filling mode" can be obtained. Each baseline behavioral profile establishes a quantified normal behavioral baseline for the mode it represents.

[0059] Specifically, by combining the multi-dimensional feature extraction results with the data correlation, the analysis establishes a baseline behavior profile set and a response behavior profile set for each type of electricity-consuming entity, which also includes:

[0060] Based on the data association, multiple directed entity relationship pairs are extracted, wherein the orientation of the directed entity relationship pair represents the controlled relationship between the corresponding entity pairs;

[0061] Based on the multidimensional feature extraction results, determine the key triggering events for each of the directed entity relationship pairs;

[0062] Obtain the subset of multidimensional feature sequences corresponding to the key triggering events from the multidimensional feature extraction results, perform common feature dimension filtering, and label the key triggering event summary according to the common feature dimension filtering results;

[0063] Based on the baseline behavior profile set and the key triggering event summary, response behavior feature pairs of multiple directed entity relationship pairs are determined;

[0064] The directed entity relationship pairs, the response behavior feature pairs, and the key triggering event summaries are associated and stored to form the response behavior profile set.

[0065] After establishing the baseline behavioral profile set, it is further necessary to establish a response behavioral profile set. This is because there are electrical connections, energy interactions, and functional couplings between the various power-consuming entities within a complex power unit. An abnormal behavior in a single entity may trigger a chain reaction in related entities. This response behavioral profile set is a knowledge set constructed based on data relationships and multi-dimensional feature extraction results. It describes the paired behavioral response patterns of specific entity relationships under the triggering of critical events and can be used to assess the associated abnormal states and potential chain risks among power-consuming entities.

[0066] First, based on the established data relationships, multiple directed entity relationship pairs are extracted. These pairs are represented as directional entity tuples, with the direction indicating the controlled relationship or influence flow between the corresponding entity pairs.

[0067] Secondly, based on the completed multidimensional feature extraction results, for each directed entity relationship pair, the corresponding key triggering event is identified and determined. Specifically, the key triggering event refers to the specific behavioral pattern or state transition exhibited by the source entity in the entity relationship pair that can trigger or significantly affect the behavioral changes of the target entity.

[0068] Furthermore, a subset of multidimensional feature sequences associated with each key triggering event is obtained from the multidimensional feature extraction results. This subset of multidimensional feature sequences is then filtered for common feature dimensions to identify those that exhibit recurring, regular changes within similar events. Based on the filtering results, a summary of key triggering events used to refine and describe the core features of the event is determined.

[0069] Furthermore, by combining the established baseline behavioral profile set with the key triggering event summary, a response behavior feature pair is determined for each directed entity relationship pair. The response behavior feature pair defines a set of interrelated characteristic states that the source entity and the target entity in the directed entity relationship pair should exhibit after a specific key triggering event occurs.

[0070] Finally, each directed entity relationship pair, its corresponding response behavior feature pair, and a summary of key triggering events are associated and stored. The resulting associated data set constitutes the response behavior profile set. This response behavior profile set systematically encapsulates the predictable inter-entity behavioral response knowledge triggered by specific events under different entity relationships.

[0071] S30: Train a first model for detecting anomalies within an entity based on the baseline behavior profile set, and train a second model for evaluating anomalies between electricity-consuming entities based on the response behavior profile set.

[0072] Specifically, a first model for detecting anomalies within entities is trained based on the baseline behavior profile set, and a second model for evaluating anomalies between electricity-consuming entities is trained based on the response behavior profile set, including:

[0073] The first model takes the real-time electricity consumption behavior feature vector of a single electricity-consuming entity as input and outputs a Boolean value representing whether the electricity-consuming entity has internal anomalies and the corresponding confidence level.

[0074] The second model takes a real-time electricity consumption behavior feature vector containing at least two electricity consumption entities with data association as input, and outputs a Boolean value representing whether there is an abnormal association between the entities and the corresponding confidence level.

[0075] The first model is used to detect internal anomalies in electricity-consuming entities. This first model takes the real-time electricity consumption behavior feature vector of a single electricity-consuming entity extracted in the preceding steps as its input. The training process of the first model relies on a baseline behavioral profile set, which defines the feature range of the corresponding category of entities under normal conditions. Through unsupervised learning algorithms, such as reconstruction-based autoencoders or prediction-based long short-term memory network models, the first model learns and memorizes the normal behavior patterns of a single entity. After training, the first model can calculate and compare the input real-time feature vectors, ultimately outputting a Boolean value and its corresponding confidence score. This Boolean value directly represents whether the electricity-consuming entity has an internal anomaly deviating from its normal behavior baseline at the current moment, while the confidence score quantifies the reliability of this judgment.

[0076] For example, since there is a complex time dependency and nonlinear relationship between the normal operation behavior pattern of an electricity-consuming entity and its multidimensional feature sequence, and deep neural network models have outstanding capabilities in sequence feature learning and pattern reconstruction, the first model can be constructed based on the reconstructed autoencoder neural network.

[0077] Specifically, the first model mainly consists of an encoder network and a decoder network. The input layer receives a normalized feature vector of the real-time electricity consumption behavior of a single electricity-consuming entity. This vector contains quantized features extracted from multiple dimensions in the time, frequency, and energy domains. The encoder network employs a multi-layer long short-term memory (LSTM) network structure. The number of hidden layer units is configured according to the time step and dimension of the input feature vector. Each LSM layer uses the Tanh activation function to capture long-term dependencies in the sequence, and layer normalization operations are embedded between network layers to stabilize the training process. The decoder network employs a symmetrical LSM network structure to the encoder, aiming to reconstruct the original input feature sequence from the latent feature representation output by the encoder. The final output of the model is the reconstructed feature sequence.

[0078] During training, key hyperparameters included a learning rate of 0.0005, 200 training epochs, and a batch size of 32. The learning rate was set to balance the stability and convergence speed of gradient updates; the number of training epochs ensured the model fully learned the potential distribution of normal behavior sequences; and the batch size balanced computational efficiency with noise control in gradient estimation. Specifically, an unsupervised learning approach was adopted. Multidimensional feature sequences representing normal behavior patterns were selected from historical normal operation data based on a baseline behavior profile set to form the training sample set.

[0079] Furthermore, the normal behavioral feature sequences from the training set are used as input, and the goal is to reconstruct these input sequences. The network weight parameters of the encoder and decoder are iteratively optimized using the backpropagation algorithm and the Adam optimizer. The mean squared error loss function is used to measure the deviation between the original input feature sequence and the model's reconstructed sequence. During training, the model's learning progress is evaluated by continuously monitoring the reduction curves of the reconstruction errors on the training and independent validation sets. Specifically, when the average reconstruction errors on the training and validation sets remain consistently within 0.5% for twenty consecutive training epochs without showing a further significant decreasing trend (e.g., the training set reconstruction error fluctuates slightly between 0.015 and 0.0152, and the validation set reconstruction error fluctuates slightly between 0.016 and 0.0163), the model is considered to have fully converged. At this point, the training process is terminated, and the first converged model is obtained. The first model can learn and memorize the behavioral characteristic patterns of electricity-consuming entities under normal conditions. When a real-time feature vector is input, the internal anomalies that deviate from the normal pattern can be effectively identified by calculating the reconstruction error, and the reconstruction error can be mapped to the anomaly confidence level.

[0080] Furthermore, a second model is used to assess anomalies in the relationships between electricity-consuming entities. The input to this second model is a set of real-time electricity consumption behavior feature vectors, derived from at least two electricity-consuming entities with predefined data relationships, collectively forming a pair of related entities to be evaluated. Training of the second model relies on a response behavior profile set, which defines the pairwise collaborative behavior feature patterns that should be exhibited by various specific entity relationships under normal operating conditions. The training process employs a model architecture suitable for processing structured relational data, such as a graph neural network. Through training, the second model can learn and master the normal mutual influence mechanisms and collaborative behavior paradigms between different electricity-consuming entities.

[0081] After training, the second model receives a set of real-time feature vectors of associated entities as input. The second model performs deep analysis on the relationships between these feature vectors and compares them with the normal association patterns learned during training to determine whether the current interaction state has deviated. The output of the second model is a Boolean value and its corresponding confidence level. This Boolean value directly indicates whether there is an abnormal association between the evaluated group of electricity-consuming entities at the current moment, and whether their behavioral interactions have deviated from the normal range of cooperation; while the output confidence level quantitatively reflects the reliability of this anomaly judgment result.

[0082] S40: Real-time monitoring acquires the monitoring dataset of the composite power unit, combines the first model and the second model to identify power load anomalies, and triggers an early warning based on the power load anomaly identification results.

[0083] Specifically, the monitoring dataset of the composite power unit is acquired in real time, and power load anomalies are identified by combining the first model and the second model. Based on the power load anomaly identification results, an early warning is triggered, including:

[0084] Based on the monitoring dataset obtained from real-time monitoring, extract the real-time electricity consumption behavior feature vector for each type of electricity-consuming entity;

[0085] Each of the real-time electricity consumption behavior feature vectors is input into the corresponding first model to perform internal anomaly identification.

[0086] If the internal anomaly identification result indicates the presence of an anomaly, then the affected objects are identified based on the data correlation, and a load anomaly warning is executed based on the result of the affected object identification.

[0087] First, the monitoring dataset of the composite power unit is acquired in real time. This monitoring dataset is a multi-source time-series data set that is continuously collected and synchronized by data sensing layer devices deployed at each key node of the power unit and covers multiple monitoring points. It includes total incoming line electrical quantity data, electrical quantity data of each branch circuit, equipment status data and related topology information, etc., and is used to characterize the real-time operating status and behavioral interactions of the composite power unit as a whole and its internal components at the time of monitoring.

[0088] Based on the monitoring dataset, real-time electricity consumption behavior feature vectors for each type of electricity-consuming entity can be extracted in parallel at the current moment. The extraction rules for these real-time electricity consumption behavior feature vectors are consistent with the multi-dimensional feature extraction methods used in the training phase to ensure that the feature dimensions match the model input requirements.

[0089] Secondly, the extracted real-time electricity consumption behavior feature vectors are input into the pre-trained corresponding first model for calculation. The first model independently analyzes and evaluates the real-time electricity consumption behavior feature vector of a single electricity consumption entity, and outputs a Boolean judgment result, which directly indicates whether the current electricity consumption entity has internal anomalies; at the same time, it outputs a quantified anomaly confidence score, which numerically represents the reliability and confidence level of this anomaly judgment result.

[0090] If any identification result indicates an internal anomaly, the early warning process is immediately initiated. Specifically, the early warning process first identifies affected entities based on pre-stored data relationships. Specifically, based on these data relationships, other electrical entities directly or indirectly connected to the currently identified anomaly are identified. These other electrical entities may be potentially affected due to electrical connections, functional coupling, or physical proximity. The set of all identified affected entities constitutes the affected objects. Finally, based on the affected object identification results, a load anomaly early warning is issued. The generation and issuance of load anomaly early warnings will comprehensively consider the severity of the internal anomaly, the confidence level of the anomaly, and the scope and criticality of the affected objects.

[0091] For example, an early warning notification can be triggered, specifying the particular abnormal entity, the type of abnormality, the potential scope of impact, and recommended handling measures. This notification can be categorized into levels of concern, handling, or emergency, and sent to relevant operation and maintenance management platforms or personnel to initiate corresponding inspection, intervention, or isolation procedures. Ultimately, this achieves a closed loop from single-entity anomaly detection to related impact assessment, and then to proactive early warning.

[0092] Simultaneously, real-time monitoring data of the composite power unit is acquired, and power load anomalies are identified by combining the first model and the second model. Based on the power load anomaly identification results, early warnings are triggered, including:

[0093] Based on the data association relationship and combined with the profile paradigm of the response behavior profile set, the data of multiple real-time electricity consumption behavior feature vectors are reconstructed to obtain comparative response behavior features.

[0094] Trigger event matching is performed on the comparative response behavior features, and the comparative response behavior features are input into the corresponding second model to perform associated anomaly identification based on the trigger event matching results;

[0095] If the correlation anomaly identification result indicates the existence of anomalies, then obtain the power consumption entity pair corresponding to the comparative response behavior characteristics, and identify the union of affected objects in conjunction with the data correlation relationship;

[0096] Execute load anomaly warnings based on the union of the affected objects.

[0097] First, based on the established data relationships and referring to the profiling paradigm defined in the response behavior profiling set, specific data reconstruction processing is performed on multiple real-time electricity consumption behavior feature vectors extracted from the monitoring dataset. Specifically, the purpose of this reconstruction processing is to generate comparative response behavior features that can be used for comparative analysis. Specifically, based on the relationships between entities, the real-time feature vectors of related entity pairs are organized into a structured feature pair or feature group, i.e., comparative response behavior features, according to the feature combination and calculation methods specified in the response behavior profiling set, to simulate the behavioral response patterns that should be observed under specific related scenarios.

[0098] Secondly, trigger event matching is performed on the reconstructed comparative response behavior features. The core purpose of this matching process is to calculate and compare the similarity between the comparative response behavior features presented by the current real-time data and the summaries of various key trigger events predefined and stored in the response behavior profile set. By calculating feature similarity or probability likelihood, the description of which key trigger event summary best matches the current feature pattern can be identified. Based on the matching result, the specific type of trigger event that may be occurring or has already occurred during the current monitoring period can be determined. Next, the current comparative response behavior features are directed into a second model pre-trained to handle this type of trigger event or corresponding to this specific set of entity relationships.

[0099] Specifically, the second model performs association anomaly identification on the input comparative response behavior features. The second model analyzes the inter-entity interaction pattern reflected by the comparative response behavior features, determines whether it deviates from the normal collaborative behavior defined by the response behavior profile set, and outputs a Boolean judgment on whether there is an association anomaly and the corresponding confidence level.

[0100] If the identification results of the second model indicate the existence of an anomaly, it is necessary to obtain the specific pair of electricity-consuming entities that generated this comparative response behavior characteristic, i.e., a source entity and a target entity involved in this anomaly interaction. Then, by combining the data correlations, an impact diffusion analysis is performed to identify all electricity-consuming entities that may be directly or indirectly affected by this anomaly association, which is the union of affected objects. This union of affected objects includes the direct participants in the anomaly association and its potential affected entities.

[0101] Finally, a load anomaly warning is issued based on the union of the identified affected objects. Specifically, the level and content of the load anomaly warning will comprehensively consider the severity of the associated anomalies, the confidence level of the second model output, and the scope and criticality of the union of affected objects.

[0102] For example, a tiered early warning can be triggered that specifies the type of abnormal association, the entity pairs involved, the scope of impact, and recommended collaborative handling measures, thereby achieving closed-loop management from abnormal interaction detection between entities to cascading impact assessment and collaborative early warning.

[0103] In summary, the embodiments of this application have at least the following technical effects:

[0104] Compared to existing technologies, this invention, by constructing a refined correlation model between electricity consumption behavior and physical entities, achieves precise definition and deconstructive monitoring of the behavior of each independent electricity-consuming entity both inside and outside the composite power unit. It refines the anomaly warning targets from coarse-grained overall metering points to specific equipment units, improving the accuracy and efficiency of fault root cause location and troubleshooting. Simultaneously, this invention establishes a two-way collaborative security protection mechanism, which not only provides early warning of early performance degradation of the power unit's own equipment but also effectively monitors the abnormal states of external service objects and assesses their cascading risks, forming a proactive internal and external security defense system.

[0105] Furthermore, by mining and utilizing multi-dimensional behavioral features that reflect the deep health status of equipment, this invention can issue early warnings in the early stages of performance degradation, achieving early and proactive warning timing. Simultaneously, by jointly applying a two-tiered model of internal entity anomaly detection and inter-entity correlation anomaly assessment, this invention constructs a multi-tiered, collaborative risk identification and early warning mechanism, enhancing the safety resilience of power systems in the face of complex anomalies and cascading failures.

[0106] Example 2, as Figure 2 As shown, based on the same inventive concept as the power load anomaly early warning method combining electricity consumption behavior big data mining provided in Embodiment 1, this embodiment of the invention also provides a power load anomaly early warning device combining electricity consumption behavior big data mining, including:

[0107] The data acquisition and analysis module 11 is used to collect multi-source heterogeneous data of composite power units based on regional big data, and analyze and obtain the data sequence of electricity consumption behavior and data correlation of each electricity-consuming entity.

[0108] The feature extraction and profile building module 12 is used to perform multi-dimensional feature extraction on the electricity consumption behavior data sequence of each electricity consumption entity, and combine the multi-dimensional feature extraction results with the data association relationship to analyze and establish a baseline behavior profile set and a response behavior profile set for each type of electricity consumption entity.

[0109] The model training module 13 is used to train a first model for detecting anomalies within an entity based on the baseline behavior profile set, and to train a second model for evaluating anomalies between electricity-consuming entities based on the response behavior profile set.

[0110] The real-time monitoring and early warning module 14 is used to monitor and acquire the monitoring dataset of the composite power unit in real time, combine the first model and the second model to identify power load anomalies, and trigger an early warning based on the power load anomaly identification results.

[0111] Specifically, the data acquisition and analysis module 11 is used for:

[0112] Based on regional big data, multi-source heterogeneous data of composite power units are collected, and the data sequence of electricity consumption behavior and data correlation of each electricity-consuming entity are analyzed and obtained.

[0113] The multi-source heterogeneous data includes total incoming line electrical quantity data, electrical quantity data of each branch circuit, equipment status data, and physical topology data.

[0114] Specifically, based on regional big data, multi-source heterogeneous data of composite power units are collected, and the electricity consumption behavior data sequence of each electricity-consuming entity is analyzed and obtained, including:

[0115] By combining big data methods and deploying data sensing layer devices, multi-source sensing data from multiple monitoring points in several composite power units within a preset area is collected.

[0116] Acquire prior entity topology data of the composite power unit, and merge it with the multi-source sensing data to output the multi-source heterogeneous data;

[0117] Based on the prior entity topology data, an association map of composite power units is constructed, and the multi-source sensing data is decomposed and associated according to the association map and the energy conservation relationship to generate an independent electricity consumption behavior data sequence and data association relationship for each electricity-consuming entity.

[0118] The feature extraction and portrait construction module 12 is specifically used for:

[0119] Specifically, multi-dimensional feature extraction is performed on the electricity consumption behavior data sequence for each electricity-consuming entity, including:

[0120] Based on the energy source of the power-consuming entity and its relationship within the composite power unit, the power-consuming entities are divided into external power-consuming entities and internal power-consuming entities.

[0121] Based on the early warning mission objectives of the composite power unit, the management depth of the power consumption entities of the composite power unit is defined, and the classification basis of the power consumption entities is determined accordingly.

[0122] Based on the classification criteria for electricity-consuming entities, a multi-level cascaded index relationship for electricity-consuming entity classification is constructed.

[0123] Specifically, multi-dimensional features are extracted from the electricity consumption behavior data sequence of each electricity-consuming entity, and the correlation between the multi-dimensional feature extraction results and the data is analyzed to establish a baseline behavior profile set and a response behavior profile set for each type of electricity-consuming entity, including:

[0124] Based on the prior feature list, the electricity consumption behavior data sequence of each electricity-consuming entity is traversed, multi-dimensional feature sequences are extracted, and the multi-dimensional feature extraction result is output.

[0125] Based on the preset electricity consumption entity classification index relationship, the multidimensional feature extraction results are classified and divided to form several multidimensional category feature clusters;

[0126] Cluster analysis is performed on each of the multidimensional category feature clusters, and fitting processing is performed on the multiple multidimensional category feature sequences corresponding to each cluster in the cluster analysis results to obtain the baseline behavior profile set of each type of electricity-consuming entity.

[0127] Combining the multidimensional feature extraction results with the data correlation, the analysis establishes a baseline behavioral profile set and a response behavioral profile set for each type of electricity-consuming entity, and also includes:

[0128] Based on the data association, multiple directed entity relationship pairs are extracted, wherein the orientation of the directed entity relationship pair represents the controlled relationship between the corresponding entity pairs;

[0129] Based on the multidimensional feature extraction results, determine the key triggering events for each of the directed entity relationship pairs;

[0130] Obtain the subset of multidimensional feature sequences corresponding to the key triggering events from the multidimensional feature extraction results, perform common feature dimension filtering, and label the key triggering event summary according to the common feature dimension filtering results;

[0131] Based on the baseline behavior profile set and the key triggering event summary, response behavior feature pairs of multiple directed entity relationship pairs are determined;

[0132] The directed entity relationship pairs, the response behavior feature pairs, and the key triggering event summaries are associated and stored to form the response behavior profile set.

[0133] Specifically, the model training module 13 is used for:

[0134] A first model for detecting anomalies within entities is trained based on the baseline behavior profile set, and a second model for evaluating anomalies between electricity-consuming entities is trained based on the response behavior profile set, including:

[0135] The first model takes the real-time electricity consumption behavior feature vector of a single electricity-consuming entity as input and outputs a Boolean value representing whether the electricity-consuming entity has internal anomalies and the corresponding confidence level.

[0136] The second model takes a real-time electricity consumption behavior feature vector containing at least two electricity consumption entities with data association as input, and outputs a Boolean value representing whether there is an abnormal association between the entities and the corresponding confidence level.

[0137] The real-time monitoring and early warning module 14 is specifically used for:

[0138] Real-time monitoring acquires monitoring datasets of composite power units, combines the first model and the second model to identify power load anomalies, and triggers early warnings based on the power load anomaly identification results, including:

[0139] Based on the monitoring dataset obtained from real-time monitoring, extract the real-time electricity consumption behavior feature vector for each type of electricity-consuming entity;

[0140] Each of the real-time electricity consumption behavior feature vectors is input into the corresponding first model to perform internal anomaly identification.

[0141] If the internal anomaly identification result indicates the presence of an anomaly, then the affected objects are identified based on the data correlation, and a load anomaly warning is executed based on the result of the affected object identification.

[0142] Specifically, the monitoring dataset of the composite power unit is acquired in real time, and power load anomalies are identified by combining the first model and the second model. Based on the power load anomaly identification results, an early warning is triggered, including:

[0143] Based on the data association relationship and combined with the profile paradigm of the response behavior profile set, the data of multiple real-time electricity consumption behavior feature vectors are reconstructed to obtain comparative response behavior features.

[0144] Trigger event matching is performed on the comparative response behavior features, and the comparative response behavior features are input into the corresponding second model to perform associated anomaly identification based on the trigger event matching results;

[0145] If the correlation anomaly identification result indicates the existence of anomalies, then obtain the power consumption entity pair corresponding to the comparative response behavior characteristics, and identify the union of affected objects in conjunction with the data correlation relationship;

[0146] Execute load anomaly warnings based on the union of the affected objects.

[0147] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0148] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0149] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for early warning of abnormal power load based on big data mining of electricity consumption behavior, characterized in that, include: Based on regional big data, multi-source heterogeneous data of complex power units are collected, and the electricity consumption behavior data sequences and data correlations of each electricity-consuming entity are analyzed and obtained, including: By combining big data methods and deploying data sensing layer devices, multi-source sensing data from multiple monitoring points in several composite power units within a preset area is collected. Acquire prior entity topology data of the composite power unit, and merge it with the multi-source sensing data to output the multi-source heterogeneous data; Based on the prior entity topology data, an association map of the composite power unit is constructed, and the multi-source sensing data is decomposed and associated according to the association map and the energy conservation relationship to generate an independent power consumption behavior data sequence and data association relationship for each power consumption entity. Multidimensional features are extracted from the electricity consumption behavior data sequence of each electricity-consuming entity. The results of the multidimensional feature extraction are then combined with the data correlation to analyze and establish a baseline behavior profile set and a response behavior profile set for each type of electricity-consuming entity, including: Based on the prior feature list, the electricity consumption behavior data sequence of each electricity-consuming entity is traversed, multi-dimensional feature sequences are extracted, and the multi-dimensional feature extraction result is output. Based on the preset electricity consumption entity classification index relationship, the multidimensional feature extraction results are classified and divided to form several multidimensional category feature clusters; Cluster analysis is performed on each of the multidimensional category feature clusters, and fitting processing is performed on the multiple multidimensional category feature sequences corresponding to each cluster in the cluster analysis results to obtain the benchmark behavior profile set of each type of electricity-consuming entity. Based on the data association, multiple directed entity relationship pairs are extracted, wherein the orientation of the directed entity relationship pair represents the controlled relationship between the corresponding entity pairs; Based on the multidimensional feature extraction results, determine the key triggering events for each of the directed entity relationship pairs; Obtain the subset of multidimensional feature sequences corresponding to the key triggering events from the multidimensional feature extraction results, perform common feature dimension filtering, and label the key triggering event summary according to the common feature dimension filtering results; Based on the baseline behavior profile set and the key triggering event summary, response behavior feature pairs of multiple directed entity relationship pairs are determined; The directed entity relationship pairs, the response behavior feature pairs, and the key triggering event summaries are associated and stored to form the response behavior profile set; A first model for detecting anomalies within an entity is trained based on the baseline behavior profile set, and a second model for evaluating anomalies between electricity-consuming entities is trained based on the response behavior profile set. The monitoring dataset of the composite power unit is obtained in real time. The first model and the second model are combined to identify power load anomalies, and an early warning is triggered based on the power load anomaly identification results.

2. The power load anomaly early warning method combining big data mining of electricity consumption behavior as described in claim 1, characterized in that, The multi-source heterogeneous data includes total incoming line electrical quantity data, electrical quantity data of each branch circuit, equipment status data, and physical topology data.

3. The power load anomaly early warning method combining big data mining of electricity consumption behavior as described in claim 1, characterized in that, Before performing multi-dimensional feature extraction on the electricity consumption behavior data sequence for each electricity-consuming entity, the process includes: Based on the energy source of the power-consuming entity and its relationship within the composite power unit, the power-consuming entities are divided into external power-consuming entities and internal power-consuming entities. Based on the early warning mission objectives of the composite power unit, the management depth of the power consumption entities of the composite power unit is defined, and the classification basis of the power consumption entities is determined accordingly. Based on the classification criteria for electricity-consuming entities, a multi-level cascaded index relationship for electricity-consuming entity classification is constructed.

4. The power load anomaly early warning method combining big data mining of electricity consumption behavior as described in claim 1, characterized in that, A first model for detecting anomalies within entities is trained based on the baseline behavior profile set, and a second model for evaluating anomalies between electricity-consuming entities is trained based on the response behavior profile set, including: The first model takes the real-time electricity consumption behavior feature vector of a single electricity-consuming entity as input and outputs a Boolean value representing whether the electricity-consuming entity has internal anomalies and the corresponding confidence level. The second model takes a real-time electricity consumption behavior feature vector containing at least two electricity consumption entities with data association as input, and outputs a Boolean value representing whether there is an abnormal association between the entities and the corresponding confidence level.

5. The power load anomaly early warning method combining big data mining of electricity consumption behavior as described in claim 1, characterized in that, Real-time monitoring acquires monitoring datasets of composite power units, combines the first model and the second model to identify power load anomalies, and triggers early warnings based on the power load anomaly identification results, including: Based on the monitoring dataset obtained from real-time monitoring, extract the real-time electricity consumption behavior feature vector for each type of electricity-consuming entity; Each of the real-time electricity consumption behavior feature vectors is input into the corresponding first model to perform internal anomaly identification. If the internal anomaly identification result indicates the presence of an anomaly, then the affected objects are identified based on the data correlation, and a load anomaly warning is executed based on the result of the affected object identification.

6. The power load anomaly early warning method combining big data mining of electricity consumption behavior as described in claim 5, characterized in that, Real-time monitoring acquires monitoring datasets of composite power units, combines the first model and the second model to identify power load anomalies, and triggers early warnings based on the power load anomaly identification results, including: Based on the data association relationship and combined with the profile paradigm of the response behavior profile set, the data of multiple real-time electricity consumption behavior feature vectors are reconstructed to obtain comparative response behavior features. Trigger event matching is performed on the comparative response behavior features, and the comparative response behavior features are input into the corresponding second model to perform associated anomaly identification based on the trigger event matching results; If the correlation anomaly identification result indicates the existence of anomalies, then obtain the power consumption entity pair corresponding to the comparative response behavior characteristics, and identify the union of affected objects in conjunction with the data correlation relationship; Execute load anomaly warnings based on the union of the affected objects.

7. A power load anomaly early warning device combining big data mining of electricity consumption behavior, characterized in that, The method for performing the power load anomaly early warning method combining big data mining of electricity consumption behavior as described in any one of claims 1-6 includes: The data acquisition and analysis module is used to collect multi-source heterogeneous data from complex power units based on regional big data, and analyze and obtain the data sequence of electricity consumption behavior and data correlation of each electricity-consuming entity, including: By combining big data methods and deploying data sensing layer devices, multi-source sensing data from multiple monitoring points in several composite power units within a preset area is collected. Acquire prior entity topology data of the composite power unit, and merge it with the multi-source sensing data to output the multi-source heterogeneous data; Based on the prior entity topology data, an association map of the composite power unit is constructed, and the multi-source sensing data is decomposed and associated according to the association map and the energy conservation relationship to generate an independent power consumption behavior data sequence and data association relationship for each power consumption entity. The feature extraction and profile building module is used to perform multi-dimensional feature extraction on the electricity consumption behavior data sequence of each electricity-consuming entity, and combine the multi-dimensional feature extraction results with the data correlation to analyze and establish a baseline behavior profile set and a response behavior profile set for each type of electricity-consuming entity, including: Based on the prior feature list, the electricity consumption behavior data sequence of each electricity-consuming entity is traversed, multi-dimensional feature sequences are extracted, and the multi-dimensional feature extraction result is output. Based on the preset electricity consumption entity classification index relationship, the multidimensional feature extraction results are classified and divided to form several multidimensional category feature clusters; Cluster analysis is performed on each of the multidimensional category feature clusters, and fitting processing is performed on the multiple multidimensional category feature sequences corresponding to each cluster in the cluster analysis results to obtain the benchmark behavior profile set of each type of electricity-consuming entity. Based on the data association, multiple directed entity relationship pairs are extracted, wherein the orientation of the directed entity relationship pair represents the controlled relationship between the corresponding entity pairs; Based on the multidimensional feature extraction results, determine the key triggering events for each of the directed entity relationship pairs; Obtain the subset of multidimensional feature sequences corresponding to the key triggering events from the multidimensional feature extraction results, perform common feature dimension filtering, and label the key triggering event summary according to the common feature dimension filtering results; Based on the baseline behavior profile set and the key triggering event summary, response behavior feature pairs of multiple directed entity relationship pairs are determined; The directed entity relationship pairs, the response behavior feature pairs, and the key triggering event summaries are associated and stored to form the response behavior profile set; The model training module is used to train a first model for detecting anomalies within entities based on the baseline behavior profile set, and to train a second model for evaluating anomalies between electricity-consuming entities based on the response behavior profile set. The real-time monitoring and early warning module is used to monitor and acquire the monitoring dataset of the composite power unit in real time, combine the first model and the second model to identify power load anomalies, and trigger early warnings based on the power load anomaly identification results.

Images