A refined analysis system for user electricity consumption behavior based on non-intrusive load monitoring.

Abstract

This invention relates to the field of power load monitoring technology, specifically a refined analysis system for user electricity consumption behavior based on non-intrusive load monitoring. The system includes: collecting time-series data of the total load at the incoming line of the target power user; decomposing the load data using an improved Hidden Markov Model algorithm; dynamically adjusting the state transition probability based on the multi-scale abrupt changes caused by load switching; extracting the operating feature vectors of each independent load; fusing environmental correlation information to construct a comprehensive electricity consumption behavior feature matrix; identifying various electricity consumption behavior patterns using a deep learning behavior clustering network; building a fine-grained electricity consumption behavior profile using sequence generation technology; generating a structured analysis report; and pushing the report to the platform. This solution optimizes the accuracy of load decomposition, improves multi-dimensional data parsing capabilities, and achieves refined identification and analysis of user electricity consumption behavior.

Description

Technical Field

[0001] This invention relates to the field of power load monitoring technology, and in particular to a system for fine analysis of user electricity consumption behavior based on non-intrusive load monitoring. Background Technology

[0002] Non-intrusive load monitoring is the mainstream method for power consumption data collection and analysis in the power operation and maintenance field. It can complete equipment power consumption monitoring and behavior statistics based on the total load data at the user's incoming line. Conventional load analysis schemes mostly use traditional hidden Markov models to process the total load time series data, rely on fixed state transition probabilities to complete the load splitting, and only extract the load operation characteristics of individual equipment for subsequent behavior analysis calculations. The parameters of traditional models are kept at fixed settings for a long time and cannot keep up with the real-time changes in on-site power conditions. The start-up, shutdown and switching operations of power equipment will produce power fluctuations of varying magnitudes, and the diverse forms of power mutations are difficult to accurately capture by fixed parameter models.

[0003] The simplistic feature extraction model only retains the load's own operational data, ignoring the correlation between environmental factors and electricity consumption. The feature data has a single dimension and narrow coverage. The shallow data analysis model can only complete basic electricity consumption data statistics, making it difficult to deeply explore the differentiated electricity consumption patterns hidden in time-series data. The behavior classification standards are relatively coarse, and the generated electricity consumption descriptions are coarse-grained, making it difficult to form a complete and detailed description of electricity consumption behavior. It cannot output standardized and refined electricity consumption analysis content, making it difficult to adapt to the refined management and control requirements of modern electricity management platforms. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a refined analysis system for user electricity consumption behavior based on non-intrusive load monitoring.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a user electricity consumption behavior fine analysis system based on non-intrusive load monitoring, comprising:

[0006] The data acquisition module acquires the total load time-series data at the incoming line of the target power user;

[0007] The load decomposition module processes the total load time series data using an improved hidden Markov model algorithm to decompose it into a sequence of operating states of multiple independent power loads. The improved hidden Markov model algorithm dynamically adjusts the state transition probability based on the multi-scale mutation characteristics of the load switching events in the total power.

[0008] The feature extraction and fusion module performs feature extraction processing on the operating state sequence of each independent power load, generates the operating mode feature vector of each load, and fuses the operating mode feature vectors of all independent power loads with the environmental correlation information in the total load time series data to generate a comprehensive power consumption behavior feature matrix.

[0009] The behavior analysis and summarization module calls a deep learning behavior clustering network to analyze the comprehensive electricity consumption behavior feature matrix, identify and summarize multiple typical electricity consumption behavior patterns of users;

[0010] The behavior profile building module constructs a fine-grained user electricity behavior profile based on the aforementioned multiple typical electricity consumption behavior patterns through behavior pattern matching and sequence generation technology.

[0011] The report generation and push module transforms the user's fine-grained electricity consumption behavior profile into a structured electricity consumption behavior analysis report and pushes it to the electricity management platform.

[0012] As a further aspect of the present invention, the total load time-series data is processed using an improved hidden Markov model algorithm to decompose it into operating state sequences of multiple independent electrical loads, including:

[0013] The total load time-series data includes voltage waveform sequence, current waveform sequence, and active power sequence;

[0014] From the total load time series data, local segments of the active power sequence are extracted in a sliding window manner. Multi-scale edge detection is performed on each local segment to identify power abrupt change points at different time scales. The power abrupt change points correspond to potential load switching events.

[0015] Based on all identified power abrupt change points, the continuous total active power sequence is divided into multiple steady-state power intervals, each of which corresponds to a stable operating state of a set of load combinations.

[0016] The improved hidden Markov model algorithm is initialized by setting different steady-state power ranges as the implicit states of the algorithm, setting the observed total active power value as the observation state, and initializing the state transition probability matrix and observation probability matrix according to the historical load feature library.

[0017] By utilizing the improved Hidden Markov Model algorithm and combining the multi-scale mutation features provided by the power mutation point, the state transition probability matrix is ​​dynamically adjusted during the algorithm iteration process, so that state transitions tend to occur near the power mutation point and unreasonable state jumps in the steady state interval are suppressed.

[0018] Using the Viterbi decoding algorithm, based on the optimized state transition probability matrix and observation probability matrix, the hidden state sequence is calculated, and the hidden state sequence is mapped back to the specific load combination state, thereby decomposing the operating state sequence of each independent power load in reverse.

[0019] As a further aspect of the present invention, the method of calculating the hidden state sequence based on the optimized state transition probability matrix and observation probability matrix using the Viterbi decoding algorithm, mapping the hidden state sequence back to the specific load combination state, and then inversely decomposing the operating state sequence of each independent electrical load includes:

[0020] Using the optimized state transition probability matrix and observation probability matrix, combined with the observed total active power sequence, the dynamic programming path probability and state backtracking pointer of the Viterbi decoding algorithm are initialized.

[0021] At each time step of the total active power sequence, all hidden states are traversed, the maximum probability among all paths to reach each current hidden state is calculated, and the maximum probability value and the state index of the previous moment that generated the maximum probability are stored.

[0022] After processing the observation data of all time steps, starting from the hidden state with the highest probability in the last time step, the complete sequence of hidden states is obtained by backtracking backward according to the stored state backtracking pointer.

[0023] Each state in the obtained implicit state sequence is matched with the load combination state represented by the steady-state power range recorded in the historical load feature library to determine the specific load on and off combination at each time point.

[0024] Based on the specific combination of load on / off states and the power values ​​of each on-state load inferred from the steady-state power range, the operating state of each independent electrical load in the corresponding time period is separated from the total active power signal, generating an independent operating state sequence for each independent electrical load.

[0025] As a further aspect of the present invention, the improved Hidden Markov Model algorithm dynamically adjusts the state transition probability based on the multi-scale abrupt changes in total power caused by load switching events, including:

[0026] During the iterative process of the improved Hidden Markov Model algorithm, the values ​​of the corresponding elements in the state transition probability matrix are dynamically adjusted according to whether there is a power mutation point near the current time point and the time scale of the power mutation point.

[0027] If the current time point is within the time window of a power mutation point that is identified as a load switching event, the transition probability between load combination states related to the load switching event is increased, while the transition probability between unrelated states is suppressed.

[0028] The size of the time window for the power mutation point is determined by the edge detection scale used when the power mutation point is detected. A mutation point with a larger time scale corresponds to a longer time window.

[0029] In each iteration, after dynamically adjusting the state transition probability matrix, it needs to be normalized again to ensure that the sum of the transition probabilities in each row is one.

[0030] As a further aspect of the present invention, the step of performing feature extraction processing on the operating state sequence of each independent electrical load to generate an operating mode feature vector for each load includes:

[0031] The operating mode feature vector includes start-stop period characteristics, operating power curve shape, and cycle working mode;

[0032] For a sequence of operating states of an independent electrical load, identify the times of all start-up and stop events;

[0033] Based on the continuous start and stop event times, the start and stop time period characteristics are extracted, which include the start frequency, average continuous running time and the most frequently running time period interval for each hour of the day.

[0034] During the duration of each operation, the power segment corresponding to the load is extracted from the total active power sequence, the power segment is standardized and resampled to generate a fixed-length power time series, and the power time series is morphologically described to obtain the shape of the operating power curve.

[0035] Analyze the repetition pattern of the operating state sequence over a long period of time, identify the operating patterns of daily cycle, weekly cycle or specific date type, and quantify them as the periodic working mode;

[0036] The start-stop period characteristics, operating power curve shape, and cyclic operating mode are concatenated and encoded into a fixed-length numerical vector, which serves as the operating mode feature vector of the load.

[0037] As a further aspect of the present invention, the step of fusing the operation mode feature vectors of all independent electricity loads with the environmental correlation information in the total load time-series data to generate a comprehensive electricity consumption behavior feature matrix includes:

[0038] The environmental information includes timestamps, temperature, and date type;

[0039] Extract time information, weekday information, and whether it is a holiday information from the timestamps of the total load time series data to form a time context vector;

[0040] Temperature information is extracted from environmental data recorded synchronously with the total load time series data and then normalized.

[0041] For each independent electrical load, its operating mode feature vector is concatenated with the time context vector and temperature information of the corresponding time period to form an enhanced feature vector of the load under a specific environment;

[0042] The enhanced characteristic vectors of all independent electrical loads within the same or consecutive time periods are arranged in chronological order to form a two-dimensional matrix. The rows of the two-dimensional matrix represent time points or time periods, and the columns represent the characteristics of different loads under specific environments. This two-dimensional matrix is ​​the comprehensive electricity consumption behavior characteristic matrix.

[0043] As a further aspect of the present invention, the step of calling a deep learning behavior clustering network to analyze the comprehensive electricity consumption behavior feature matrix, identifying and summarizing multiple typical electricity consumption behavior patterns of users, including:

[0044] The deep learning behavior clustering network consists of a feature encoder and a cluster allocator, and the feature encoder is a deep autoencoder structure.

[0045] The comprehensive electricity consumption behavior feature matrix is ​​input into the feature encoder, and the encoder part compresses it into a low-dimensional feature representation vector.

[0046] The clustering allocator receives the feature representation vector and calculates its distance or association metric with several cluster center vectors maintained within the network.

[0047] Based on the measurement results, the electricity consumption behavior characteristics of the current time period are softly assigned to each cluster center to obtain the cluster assignment probability distribution;

[0048] During training, the reconstruction error of the feature encoder and the clustering loss of the cluster allocator are jointly optimized, so that the deep learning behavior clustering network can learn the ability to map consistent electricity consumption behavior features to the same cluster center.

[0049] After training, the comprehensive electricity consumption behavior feature matrix of all historical time periods is input into the network. Based on the clustering and allocation results, multiple typical electricity consumption behavior patterns of the user are summarized. Each pattern is represented by a cluster center and the feature of all time period samples belonging to the cluster.

[0050] As a further aspect of the present invention, the step of jointly optimizing the reconstruction error of the feature encoder and the clustering loss of the clustering allocator during training includes:

[0051] The comprehensive electricity consumption behavior feature matrix is ​​forward propagated, the feature representation vector is obtained by the encoder, and the input matrix is ​​reconstructed by the decoder. The mean square error between the original input matrix and the reconstructed matrix is ​​calculated as the reconstruction error.

[0052] Calculate the similarity between the feature representation vector and the student t-distribution of each cluster center vector to obtain the cluster assignment probability distribution. Calculate the KL divergence between the cluster assignment probability distribution and the target distribution as the clustering loss. The target distribution is obtained by increasing the confidence of the assignment probability distribution.

[0053] The total loss function of the joint optimization is the weighted sum of the reconstruction error and the clustering loss. The parameters of the feature encoder, decoder and cluster center vector are updated through the backpropagation algorithm to minimize the total loss function.

[0054] As a further aspect of the present invention, based on the aforementioned multiple typical electricity consumption behavior patterns, a fine-grained user electricity consumption behavior profile is constructed using behavior pattern matching and sequence generation technology, including:

[0055] The user's fine-grained electricity consumption behavior profile includes habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments;

[0056] From the various typical electricity consumption behavior patterns, representative features of each pattern are extracted. These representative features include the typical operating state combinations of each load under the various typical electricity consumption behavior patterns and the time and environmental characteristics of the occurrence.

[0057] For new or historical electricity consumption periods of users, the real-time or historical comprehensive electricity consumption behavior feature matrix is ​​matched with the representative features of various patterns to identify the typical electricity consumption behavior pattern category of each period.

[0058] Connect the typical electricity consumption behavior patterns identified in all time periods in chronological order to form a sequence of user behavior patterns;

[0059] Analyze the behavioral pattern sequence, and statistically analyze the frequency and duration of various behavioral patterns under different date types and seasons to extract the habitual electricity consumption periods;

[0060] In the sequence of behavioral patterns, frequently co-occurring or sequentially occurring load state combinations are identified to form the load linkage rules;

[0061] Identify and mark abnormal time periods in the behavioral pattern sequence that have a low match with any typical pattern as the abnormal power consumption segments.

[0062] By integrating the habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments, a complete fine-grained profile of the user's electricity consumption behavior is formed.

[0063] As a further aspect of the present invention, the user's fine-grained electricity consumption behavior profile is transformed into a structured electricity consumption behavior analysis report, including:

[0064] Define the template structure of the electricity consumption behavior analysis report. The template structure includes a summary of basic user information, a description of typical behavior patterns, a statistical table of habitual electricity consumption periods, a list of load linkage rules, and a list of abnormal electricity consumption segments.

[0065] From the user's fine-grained electricity consumption behavior profile, extract specific information on habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments;

[0066] Descriptive features of each type of typical electricity consumption behavior pattern are extracted from the aforementioned patterns to form the typical behavior pattern library description.

[0067] Fill all the extracted information into the corresponding parts according to the specifications of the template structure;

[0068] The completed report is formatted to generate an easy-to-read and parse document, which serves as the structured electricity consumption behavior analysis report.

[0069] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0070] Based on the multi-scale abrupt changes in total power caused by load switching events, the state transition probabilities of the Hidden Markov Model are dynamically adjusted, abandoning the traditional fixed-parameter operation mode. It conforms to the real-time fluctuation patterns of power load, adapts to the power fluctuation differences caused by the switching of different equipment, and is suitable for complex and ever-changing on-site power consumption scenarios. It reduces data deviations caused by load status determination, accurately captures the start-up, shutdown, and operational changes of various independent power loads, and fully reconstructs the continuous temporal changes of multiple sets of power loads. It improves the fit between load decomposition results and actual power consumption conditions, enhances the adaptability of time-series load analysis, maintains stable operation of the load splitting calculation process, and fully presents the temporal changes of multiple types of power loads operating synchronously or alternately.

[0071] This system integrates the operational mode feature vectors of all independent electricity loads and combines them with environmental correlation information covered by total load time-series data to achieve multi-dimensional data fusion, constructing a comprehensive electricity consumption behavior feature matrix. This broadens the coverage of electricity consumption data analysis, compensates for the data shortcomings of single load features, and strengthens the correlation expression between various electricity consumption influencing factors. Relying on a deep learning behavioral clustering network to process the fused feature data, it clearly delineates the classification boundaries of different electricity consumption behaviors, distinguishes various differentiated electricity consumption operation modes, and expands the identification coverage of electricity consumption behaviors. It enriches the detailed content of electricity consumption behavior representation, expands the dimensions of user electricity consumption behavior profiles, and improves the overall content system of electricity consumption behavior analysis. It ensures that electricity-related data can be smoothly converted into standardized text content, adapting to the data transmission and daily business needs of the electricity management platform. Attached Figure Description

[0072] Figure 1 This is a timing diagram of the user electricity consumption behavior fine analysis system based on non-intrusive load monitoring described in this invention.

[0073] Figure 2 A flowchart for calculating the hidden state sequence and decomposing the running state sequence for the Viterbi decoding algorithm;

[0074] Figure 3 This is a flowchart for dynamically adjusting the state transition probability matrix. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0076] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0077] See Figure 1 This invention provides a refined analysis system for user electricity consumption behavior based on non-intrusive load monitoring. The specific system includes:

[0078] The system comprises a data acquisition module, a load decomposition module, a feature extraction and fusion module, a behavior analysis and summarization module, a behavior profile construction module, and a report generation and push module. It acquires the total load time-series data at the incoming line of the target power user. The load decomposition module processes the total load time-series data using an improved Hidden Markov Model algorithm. This algorithm dynamically adjusts the state transition probability based on the multi-scale abrupt changes in total power caused by load switching events, thereby decomposing the data into multiple independent power load operating state sequences. The feature extraction and fusion module performs feature extraction processing on the operating state sequence of each independent power load, generating operating mode feature vectors for each load. It then fuses these operating mode feature vectors with environmental correlation information from the total load time-series data to generate a comprehensive power consumption behavior feature matrix. The behavior analysis and summarization module uses a deep learning behavior clustering network to analyze the comprehensive power consumption behavior feature matrix, identifying and summarizing multiple typical power consumption behavior patterns of the user. The behavior profile construction module, based on these typical power consumption behavior patterns, constructs a fine-grained power consumption behavior profile of the user through behavior pattern matching and sequence generation techniques. The report generation and push module transforms users' fine-grained electricity consumption behavior profiles into structured electricity consumption behavior analysis reports and pushes them to the electricity management platform.

[0079] In one embodiment of the present invention, the total load time-series data includes voltage waveform sequences, current waveform sequences, and active power sequences. Local segments of the active power sequence are extracted from the total load time-series data using a sliding window approach. Multi-scale edge detection is performed on each local segment to identify power abrupt change points at different time scales. These power abrupt change points correspond to potential load switching events. Based on all identified power abrupt change points, the continuous total active power sequence is divided into multiple steady-state power intervals, each corresponding to a stable operating state of a load combination. An improved Hidden Markov Model (HMM) algorithm is initialized, setting different steady-state power intervals as the algorithm's implicit states and the observed total active power values ​​as the observed states. The state transition probability matrix and observation probability matrix are initialized based on a historical load feature database. Utilizing the improved HMM algorithm, combined with the multi-scale abrupt change features provided by the power abrupt change points, the state transition probability matrix is ​​dynamically adjusted during algorithm iteration, making state transitions more likely to occur near power abrupt change points and suppressing unreasonable state jumps within steady-state intervals. Using the Viterbi decoding algorithm, based on the optimized state transition probability matrix and observation probability matrix, the hidden state sequence is calculated, and the hidden state sequence is mapped back to the specific load combination state, thereby decomposing the operating state sequence of each independent electrical load in reverse.

[0080] See Figure 2Using the optimized state transition probability matrix and observation probability matrix, combined with the observed total active power sequence, the dynamic programming path probability and state backtracking pointer of the Viterbi decoding algorithm are initialized. At each time step of the total active power sequence, all hidden states are traversed, the maximum probability among all paths to reach each hidden state is calculated, and the maximum probability value and the state index of the time step preceding the generation of the maximum probability are stored. After processing the observation data of all time steps, starting from the hidden state with the highest probability in the last time step, the complete hidden state sequence is obtained by backtracking backward according to the stored state backtracking pointer. Each state in the obtained hidden state sequence is matched with the load combination state represented by the steady-state power interval recorded in the historical load feature database to determine the specific load on / off combination at each time point. Based on the specific load on / off combination and the power value of each on-load inferred from the steady-state power interval, the operating state of each independent electrical load in the corresponding time period is separated from the total active power signal in reverse, generating an independent operating state sequence for each independent electrical load.

[0081] In practice, the acquired total load time-series data includes voltage waveform sequences, current waveform sequences, and active power sequences. From the total load time-series data, local segments of the total active power sequence are extracted using a sliding window with a width of 5 seconds. Each local segment undergoes first- and second-order differential edge detection using a Gaussian kernel to identify power abrupt changes at different time scales. For example, a 2-second step change with an amplitude of 1500 watts is identified as a small-scale abrupt change, while a 10-second slow increase with an amplitude of 80 watts is identified as a large-scale abrupt change. These power abrupt changes correspond to potential load switching events. Based on all identified power abrupt changes, the continuous total active power sequence is divided into multiple steady-state power intervals. For example, in one steady-state interval, the power is stable around 180 watts, and in another, it is stable around 1250 watts. Each steady-state power interval corresponds to the stable operating state of a set of load combinations. An improved Hidden Markov Model (HMM) algorithm is initialized by setting the power value sets represented by different steady-state power intervals as the implicit states of the algorithm, and the observed average total active power per minute as the observed state. The state transition probability matrix and observation probability matrix are initialized based on the known load power values ​​recorded in the historical load feature database. Utilizing the improved HMM algorithm and combining the multi-scale mutation features provided by power mutation points, the state transition probability matrix is ​​dynamically adjusted during algorithm iteration. This makes state transitions more likely to occur near power mutation points and suppresses unreasonable state jumps within the steady-state interval.

[0082] In some embodiments, the improved Hidden Markov Model (HMM) algorithm dynamically adjusts the state transition probability based on the multi-scale abrupt changes in total power caused by load switching events. This adjustment is implemented through a function based on the time window of the abrupt change point. Specifically, in each iteration of the improved HMM algorithm, for each element in the state transition probability matrix—that is, the probability of transitioning from hidden state i to hidden state j—it is corrected based on whether there is a power abrupt change point near the current time point t and the time scale of the power abrupt change point. The correction rule can be expressed as follows: if time point t is within the time window of a power abrupt change point m, the state transition probability corresponding to the load switching event associated with the power abrupt change point m is increased. For example, the transition probability from the "refrigerator only running" state to the "refrigerator and air conditioner running simultaneously" state is enhanced, while the transition probability between unrelated states is suppressed. For example, the transition probability from the "refrigerator only running" state to the "television running" state is weakened. The size of the time window for the power abrupt change point is determined by the edge detection scale used when the power abrupt change point m is detected; a larger time scale corresponds to a longer time window. In each iteration, after dynamically adjusting the state transition probability matrix, each row of the matrix needs to be re-normalized to ensure that the sum of the transition probabilities in each row is one. A feasible mathematical expression is as follows:

[0083]

[0084] in: It is the probability of transitioning from state i to state j before adjustment. It is the probability of transitioning from state i to state k before adjustment. It is the transition probability adjusted at time point t. It is an adjustment function based on the set M of all power abrupt change points near time point t. A value greater than 1 indicates enhancement, and a value less than 1 indicates suppression. The normalization operation ensures... .

[0085] Specifically, using the Viterbi decoding algorithm, based on the optimized state transition probability matrix and observation probability matrix, a hidden state sequence is calculated. This hidden state sequence is then mapped back to specific load combination states, and the operating state sequence of each independent electrical load is decomposed in reverse. Using the optimized state transition probability matrix and observation probability matrix, combined with the observed total active power sequence, the dynamic programming path probability and state backtracking pointer of the Viterbi decoding algorithm are initialized. At each time step of the total active power sequence, all hidden states are traversed, the maximum probability among all paths to reach each hidden state is calculated, and the maximum probability value and the state index of the time step preceding the maximum probability are stored. After processing the observation data for all time steps, starting from the hidden state with the highest probability in the last time step, the complete hidden state sequence is obtained by backtracking according to the stored state backtracking pointer. Each state in the obtained hidden state sequence is matched with the load combination state represented by the steady-state power range recorded in the historical load feature database. For example, the hidden state "S2" matches the load combination state of "air conditioning (cooling) and refrigerator running simultaneously" recorded in the historical load feature database. Based on the specific combination of load on / off states and the power values ​​of each on-state load inferred from the steady-state power range, the operating state of each independent electrical load in the corresponding time period is separated from the total active power signal, generating an independent operating state sequence for each independent electrical load, such as the on / off state sequence of the refrigerator and the on / off state sequence of the air conditioner.

[0086] In one embodiment of the present invention, see [reference] Figure 3 In the iterative process of the improved Hidden Markov Model algorithm, the values ​​of corresponding elements in the state transition probability matrix are dynamically adjusted based on whether there are power abrupt changes near the current time point and the time scale of these power abrupt changes. If the current time point is within the time window of a power abrupt change identified as a load switching event, the transition probability between load combination states related to the load switching event is increased, while the transition probability between unrelated states is suppressed. The size of the time window for the power abrupt change is determined by the edge detection scale used when the power abrupt change is detected; abrupt changes with a larger time scale correspond to a longer time window. In each iteration, after dynamically adjusting the state transition probability matrix, it needs to be re-normalized to ensure that the sum of the transition probabilities in each row is one.

[0087] In its implementation, the improved Hidden Markov Model (HMM) algorithm dynamically adjusts the state transition probability based on the multi-scale abrupt changes in total power caused by load switching events. During each iteration of the improved HMM algorithm, the system scans the power abrupt change detection results near the current time point. Based on the existence of power abrupt changes near the current time point and their time scale, the system dynamically adjusts the values ​​of the corresponding elements in the state transition probability matrix. Power abrupt changes are identified by multi-scale edge detection of the total active power sequence. Each power abrupt change is accompanied by its timestamp and detection scale information. If the current time point falls within the time window of a power abrupt change identified as a load switching event, the transition probability between load combination states related to the load switching event is increased, while the transition probability between unrelated states is suppressed.

[0088] In some embodiments, the size of the time window for power abrupt changes is determined by the edge detection scale used when the power abrupt change is detected; abrupt changes with a larger time scale correspond to a longer time window. For example, a smooth power rise edge detected using a large-scale Gaussian kernel, reflecting the startup of slow HVAC equipment, might have a time window set to 15 seconds; while a rapid power step detected using a small-scale Gaussian kernel, reflecting the switching on and off of lighting fixtures, might have a time window set to 2 seconds. When the improved Hidden Markov Model algorithm reaches a time point t within a power abrupt change time window, the algorithm consults the historical load feature database. Power abrupt changes are usually associated with the state changes of one or more specific loads. For example, if a power surge of approximately 2000 watts is detected at 7 PM, the historical load feature database indicates that this may be associated with the activation of the air conditioning heating mode. Therefore, the algorithm will increase the transition probability from the state of "no air conditioning operation" to the state of "air conditioning heating operation" in the state transition probability matrix, while decreasing the transition probability from the state of "no air conditioning operation" to unrelated states such as "electric kettle operation".

[0089] Optionally, the dynamic adjustment operation can be implemented using a gain function related to the strength and proximity of the power mutation point, which acts on the original values ​​of the state transition probability matrix. For each element in the state transition probability matrix... , indicating from the implicit state Transition to latent state The original probability, This represents the original state transition probability from hidden state i to hidden state k, and its adjusted probability. The calculation at time point t can take into account all active power abrupt change points, and one mathematical expression is as follows:

[0090]

[0091] in: This represents the set of power abrupt change points that are in an active state at time point t (i.e., t is within their time window). It is a gain factor for a specific power mutation point m and state transition pair (i,j). A value greater than 1 indicates enhancement of the transition, while a value less than 1 but greater than 0 indicates suppression. The specific value depends on whether the state transition pair (i,j) is associated with the load event implied by the power mutation point m, and how close time point t is to the center moment of the power mutation point m. The summation and normalization operations of the denominator ensure that the adjusted state transition... The sum of the probabilities of transitioning to all possible states is strictly equal to 1. In each iteration, after dynamically adjusting the state transition probability matrix, the above normalization must be performed again to ensure that the sum of the transition probabilities in each row is equal to 1.

[0092] It is understandable that the gain factor is... The specific definition of the gain factor can be designed in several ways. One direct approach is based on rule mapping: if the state transition pair (i,j) precisely matches the most probable load change event recorded in the historical load feature database for the power abrupt change point m, a fixed, higher gain value is assigned; otherwise, a fixed, lower gain value is assigned. Another more continuous approach is based on similarity: calculating the similarity between the total power change corresponding to the state transition and the magnitude and direction of the power abrupt change point m, and converting the similarity into a gain value. Regardless of the method used to define the gain factor, the core objective is to ensure that, near the power abrupt change point, the probability of the event-related state transition path is significantly higher than other unreasonable state transition paths, thereby guiding the improved Hidden Markov Model algorithm to decode a load state sequence that better reflects the real physical process.

[0093] In one embodiment of the present invention, feature extraction processing is performed on the operating state sequence of each independent electrical load. The operating mode feature vector includes start-stop period features, operating power curve shape, and periodic operating mode. For the operating state sequence of an independent electrical load, all start-up and stop-stop event times are identified. Based on the continuous start-stop event times, start-stop period features are extracted, including the start-up frequency, average continuous running duration, and most frequently operating time interval for each hour of the day. Within the continuous running period, the power segment corresponding to the load is extracted from the total active power sequence. The power segment is standardized and resampled to generate a fixed-length power time series. The power time series is morphologically described to obtain the operating power curve shape. The repetition pattern of the operating state sequence over a long time span is analyzed to identify daily, weekly, or date-specific operating modes, which are quantified as periodic operating modes. The start-stop period features, operating power curve shape, and periodic operating mode are concatenated and encoded into a fixed-length numerical vector, which serves as the operating mode feature vector of the load.

[0094] Environmental association information includes timestamps, temperature, and date type. The operating mode feature vectors of all independent electrical loads are fused with the environmental association information in the total load time-series data. From the timestamps of the total load time-series data, time information, day of the week information, and whether it is a holiday are extracted to form a time context vector. Temperature information is extracted from environmental data recorded synchronously with the total load time-series data and normalized. For each independent electrical load, its operating mode feature vector is concatenated with the time context vector and temperature information of the corresponding time period to form an enhanced feature vector of the load under a specific environment. The enhanced feature vectors of all independent electrical loads within the same or consecutive time periods are arranged chronologically to form a two-dimensional matrix. The rows of the two-dimensional matrix represent time points or time periods, and the columns represent the characteristics of different loads under a specific environment. This two-dimensional matrix is ​​the comprehensive electricity consumption behavior feature matrix.

[0095] In practical implementation, feature extraction processing is performed on the operating state sequence of each independent electrical load. The operating mode feature vector includes start-stop period characteristics, operating power curve shape, and cyclical operating mode. For the operating state sequence of an independent electrical load, such as an air conditioner, the system scans the binary operating state sequence generated by the load decomposition module to identify all start-up and stop-stop event times. For example, on August 5, 2023, the air conditioner was recorded to start at 10:05:23 and stop at 12:30:45. Based on the continuous start-up and stop-stop event times, start-stop period characteristics are extracted. These characteristics include the start-up frequency, average continuous running time, and the most frequently operating time interval for each hour of the day. For example, by statistically analyzing the air conditioner's operating data over a week, the start-stop period characteristics can be calculated (see Table 1 for a summary). Within the duration of each operation, power segments corresponding to the air conditioning load are extracted from the total active power sequence. These power segments are standardized and resampled to generate a fixed-length power time series. For example, the power curve of each operation is resampled into an equally spaced sequence of length 100. The power time series is then morphologically described to obtain the shape of the operating power curve. The morphological description may include statistical features such as mean, variance, and skewness, as well as waveform descriptors extracted from the resampled sequence. The repetitive patterns of the operating state sequence over long time spans are analyzed to identify daily, weekly, or date-specific operating patterns, which are then quantified as periodic operating modes. For example, if the air conditioning runs significantly longer on weekend afternoons than on weekdays, this pattern can be quantified as an identifier code representing "long-term operation on weekend afternoons." The start-stop period characteristics, operating power curve shape, and periodic operating modes are concatenated and encoded into a fixed-length numerical vector, serving as the load's operating mode feature vector. For example, this could be a 100-dimensional vector containing the 24-hour start frequency, average running length, a time interval encoding, 10 waveform descriptor values, and 5 periodic mode codes.

[0096] Table 1: Summary of Statistical Characteristics of Air Conditioner Start-up and Shutdown Times within One Week

[0097] In some embodiments, the power segment is standardized using the Z-score normalization method, with the specific formula as follows:

[0098]

[0099] in: It is the nth point of the normalized power sequence. It is the nth point of the original power segment extracted from the total active power sequence. It is the average value of all points in the original power segment. This is the standard deviation of all points in the original power segment. After this standardization process, the power segment is dimensionless, its morphological characteristics are more prominent, and it is easier to conduct subsequent curve morphology analysis and compare different operating events.

[0100] Optionally, environmental association information includes timestamps, temperature, and date type. The operating mode feature vectors of all independent power loads are fused with the environmental association information in the total load time-series data to generate a comprehensive power consumption behavior feature matrix. From the timestamps of the total load time-series data, time information, weekday information, and whether it is a holiday are extracted to form a time context vector. For example, for a timestamp "2023-08-05 14:30:00", the extracted time information is 14.5 (hours and minutes converted to decimals), the weekday information is Saturday (coded as 6), and the holiday information is no (coded as 0), forming a three-dimensional time context vector [14.5, 6, 0]. From the environmental data recorded synchronously with the total load time-series data, temperature information is extracted and normalized. For example, the original temperature is 30 degrees Celsius, and after linear normalization to the [0, 1] interval, the value is 0.75. For each independent power load, its operating mode feature vector is concatenated with the corresponding time context vector and temperature information to form an enhanced feature vector of the load under specific environmental conditions. For example, the operating mode feature vector of an air conditioner is 100-dimensional. Concatenating it with a 3-dimensional time context vector and a 1-dimensional normalized temperature value forms a 104-dimensional enhanced feature vector. Arranging the enhanced feature vectors of all independent electrical loads within the same or consecutive time periods in chronological order creates a two-dimensional matrix. The rows of this matrix represent time points or time periods, and the columns represent the characteristics of different loads under specific environmental conditions. This two-dimensional matrix is ​​the comprehensive electricity consumption behavior feature matrix. For instance, if the system monitors three loads—air conditioner, refrigerator, and lighting—each with a 104-dimensional enhanced feature vector, arranging the features of the 10-minute time period from 9:00 AM to 9:10 AM in rows, with each row being a 312-dimensional vector concatenated from the three loads, ultimately forms a 10-row, 312-column comprehensive electricity consumption behavior feature matrix.

[0101] It's understandable that the construction method of the time context vector can be adjusted according to the analysis requirements. For example, month and season information can be added, or a day can be divided into time periods such as "early morning, morning, afternoon, and night" and uniquely coded. Temperature information normalization can also employ other methods such as min-max normalization. Regardless of the specific encoding and normalization method used, the core objective is to quantify environmental factors into numerical vectors that can be concatenated with load operation characteristics. This allows the comprehensive electricity consumption behavior feature matrix to simultaneously retain both the load's own operating characteristics and external environmental context information, providing multi-dimensional input data for subsequent electricity consumption behavior clustering analysis.

[0102] In one embodiment of the present invention, the deep learning behavior clustering network consists of a feature encoder and a cluster allocator, wherein the feature encoder is a deep autoencoder structure. The comprehensive electricity consumption behavior feature matrix is ​​input into the feature encoder, which compresses it into a low-dimensional feature representation vector. The cluster allocator receives the feature representation vector and calculates its distance or association metric with several cluster center vectors maintained internally by the network. Based on the metric results, the electricity consumption behavior features of the current time period are soft-assigned to various cluster centers, resulting in a cluster assignment probability distribution. During training, the reconstruction error of the feature encoder and the clustering loss of the cluster allocator are jointly optimized, enabling the deep learning behavior clustering network to learn the ability to map consistent electricity consumption behavior features to the same cluster centers. After training, the comprehensive electricity consumption behavior feature matrix of all historical time periods is input into the network. Based on the cluster assignment results, multiple typical electricity consumption behavior patterns of users are summarized, with each pattern represented by a cluster center and all time period sample features belonging to the cluster.

[0103] The forward propagation algorithm integrates the feature matrix of electricity consumption behavior. The encoder obtains the feature representation vector, which is then used to reconstruct the input matrix via the decoder. The mean squared error between the original input matrix and the reconstructed matrix is ​​calculated as the reconstruction error. The similarity between the feature representation vector and the Student t-distribution of each cluster center vector is calculated to obtain the cluster assignment probability distribution. The KL divergence between the cluster assignment probability distribution and the target distribution is calculated as the clustering loss. The target distribution is obtained by increasing the confidence of the assignment probability distribution. The joint optimization of the total loss function is a weighted sum of the reconstruction error and the clustering loss. The parameters of the feature encoder, decoder, and cluster center vectors are updated using the backpropagation algorithm to minimize the total loss function.

[0104] In practical implementation, the deep learning behavior clustering network consists of a feature encoder and a cluster allocator. The feature encoder is a deep autoencoder structure, and its encoder part can contain multiple fully connected layers to compress the high-dimensional input into a low-dimensional representation. The decoder part then attempts to reconstruct the original input from the low-dimensional representation. The comprehensive electricity consumption behavior feature matrix is ​​input into the feature encoder, which compresses it into a low-dimensional feature representation vector. For example, a comprehensive electricity consumption behavior feature matrix containing 10 time points, each with 312-dimensional features, can be compressed into a 10-dimensional feature representation vector by the encoder. The cluster allocator receives the feature representation vector and calculates its distance or association metric with several cluster center vectors maintained internally by the network. The cluster center vectors are trainable parameters that can be randomly initialized initially. Based on the measurement results, the electricity consumption behavior features of the current time period are soft-assigned to each cluster center to obtain the cluster assignment probability distribution. For example, for a feature representation vector, after calculating its similarity with the three cluster centers, the assignment probability distribution is [0.1, 0.8, 0.1], which means that the electricity consumption behavior of this time period has an 80% probability of belonging to the second typical pattern.

[0105] In some embodiments, the training process uses a historically accumulated comprehensive electricity consumption behavior feature matrix as the dataset. A batch of comprehensive electricity consumption behavior feature matrices is forward-propagated, and the corresponding feature representation vectors are obtained through the encoder. The input matrix is ​​then reconstructed through the decoder, and the mean squared error between the original input matrix and the reconstructed matrix is ​​calculated as the reconstruction error. Simultaneously, the similarity between the feature representation vectors and the student t-distributions of each cluster center vector is calculated to obtain the cluster assignment probability distribution. The KL divergence between the cluster assignment probability distribution and the target distribution is calculated as the clustering loss. The target distribution is obtained by increasing the confidence of the assignment probability distribution. Specifically, the assignment probability distribution is squared and then normalized by clustering to make the distribution more "sharp," thereby encouraging the network to learn more deterministic cluster assignments. The total loss function of the joint optimization is the weighted sum of the reconstruction error and the clustering loss. The parameters of the feature encoder, decoder, and cluster center vectors are updated through the backpropagation algorithm to minimize the total loss function. See Table 2 for the forward propagation and loss calculation of a batch of training data.

[0106] Table 2: Training Batch Data Table for Deep Learning Behavioral Clustering Network

[0107] The total loss function of the joint optimization can be expressed by the following formula:

[0108]

[0109] in: This represents the total loss of the joint optimization. The reconstruction error of the feature encoder is usually calculated using the mean square error. This represents the clustering loss of the clustering allocator, i.e., the KL divergence between the clustering probability distribution and the target distribution. This is a hyperparameter used to balance the weights of the two loss terms. By minimizing... Deep learning behavioral clustering networks can simultaneously learn efficient compressed representations of data and feature space structures conducive to clustering. After training, the comprehensive electricity consumption behavior feature matrix of all historical time periods is input into the network. Based on the clustering assignment results, multiple typical electricity consumption behavior patterns of users are summarized. Each pattern is represented by a cluster center and the feature features of all time period samples belonging to the cluster. For example, the cluster center vector may represent "weekday daytime home mode", while another cluster center vector represents "nighttime sleep mode".

[0110] Optionally, the structure of the feature encoder can be adjusted according to the complexity of the input data. For example, for a comprehensive electricity consumption behavior feature matrix over a long time series, the encoder part can use a one-dimensional convolutional layer or a long short-term memory network layer to capture temporal dependencies. When calculating similarity in the cluster assigner, in addition to using the Student's t-distribution, a Gaussian distribution or a simple dot product similarity measure can also be used. After obtaining the cluster assignment probability distribution, it can be hard-assigned to a specific cluster by taking the maximum value, for subsequent pattern induction and analysis. After training, the cluster center vectors within the network are fixed; they represent the core prototypes of the learned multiple typical electricity consumption behavior patterns in the low-dimensional feature space.

[0111] It is understandable that the reconstruction error term prompts the network to retain the main information of the input data and prevents the feature representation from collapsing, while the clustering loss term guides the network to learn feature representations with good cohesion, making similar electricity consumption behaviors close to each other in the feature space. Hyperparameters The relative importance of the two losses needs to be controlled and adjusted according to the characteristics of the actual data. The entire training process is unsupervised. The network automatically discovers and summarizes recurring patterns from the comprehensive electricity consumption behavior feature matrix without the need for pre-labeling any behavior categories. This adapts to the real-world scenarios where patterns are unknown and diverse in electricity consumption behavior analysis.

[0112] In one embodiment of the present invention, the user's fine-grained electricity consumption behavior profile includes habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments. Representative features are extracted from multiple typical electricity consumption behavior patterns. These representative features include typical operating state combinations of each load under each typical electricity consumption behavior pattern and the time-environment characteristics of the occurrence. For new or historical electricity consumption periods, the real-time or historical comprehensive electricity consumption behavior feature matrix is ​​matched with the representative features of each pattern to identify the typical electricity consumption behavior pattern category to which each period belongs. The typical electricity consumption behavior pattern categories identified in all periods are connected in chronological order to form a user behavior pattern sequence. The behavior pattern sequence is analyzed to statistically analyze the frequency and duration of each behavior pattern under different date types and seasons, extracting habitual electricity consumption periods. Frequently co-occurring or sequentially occurring load state combinations are identified within the behavior pattern sequence to form load linkage rules. Abnormal periods in the behavior pattern sequence with low matching degree to any typical pattern are identified and marked as abnormal electricity consumption segments. The habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments are integrated to form a complete user fine-grained electricity consumption behavior profile.

[0113] Define a template structure for the electricity consumption behavior analysis report. This template includes a summary of basic user information, a description of typical behavior patterns, a statistical table of habitual electricity consumption periods, a list of load linkage rules, and a list of abnormal electricity consumption segments. Specific information on habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments is extracted from the user's fine-grained electricity consumption behavior profile. Descriptive features of various typical electricity consumption behavior patterns are extracted to form a description of the typical behavior pattern library. All extracted information is then filled into the corresponding sections according to the template structure. The completed report is then formatted to generate an easy-to-read and parse document, serving as a structured electricity consumption behavior analysis report.

[0114] In practical implementation, the fine-grained user electricity consumption behavior profile includes habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments. Representative features are extracted from multiple typical electricity consumption behavior patterns. These representative features include the typical operating state combinations of each load under each typical electricity consumption behavior pattern and the time and environmental characteristics of the occurrence. For example, a cluster center vector learned from a deep learning behavioral clustering network can be interpreted as a typical electricity consumption behavior pattern after decoding or mapping. This pattern is described as a "weekday evening home mode," and its representative features include: the main operating loads are air conditioners, televisions, and lighting; the air conditioner is set to 26 degrees Celsius; the television power is maintained at around 80 watts; all lighting is on; and the time and environmental characteristics of the occurrence are 18:00 to 22:00 from Monday to Friday, with indoor temperatures between 28 and 30 degrees Celsius. For new or historical electricity consumption periods, the real-time or historical comprehensive electricity consumption behavior feature matrix is ​​matched with representative features of various patterns to identify the typical electricity consumption behavior pattern category of each period. The matching process can be achieved by calculating the Euclidean distance or cosine similarity between the low-dimensional feature representation vector obtained after encoding the comprehensive electricity consumption behavior feature matrix of the current period and the cluster center vectors. The pattern with the smallest distance or the highest similarity is determined as the category of that period. The typical electricity consumption behavior pattern categories identified in all periods are connected in chronological order to form the user's behavior pattern sequence. For example, the behavior pattern sequence for a week can be represented as ["away from home", "away from home", "at home in the evening", "sleep", "at home in the evening", "away from home", "at home all day"].

[0115] In some embodiments, the matching process may employ a distance-based soft assignment calculation, whereby, for a feature representation vector over a given time period, the degree of matching with the k-th typical electricity consumption behavior pattern is calculated. It can be calculated using the following formula:

[0116]

[0117] in: This represents the matching score between the current time-period feature and the k-th type of pattern. It is an exponential function. It is a scaling parameter that controls the rate at which the fraction decays. Feature representation vector representing the current time period Cluster center vector of the k-th pattern Distance metric (e.g., Euclidean distance), matching score The value ranges from 0 to 1, with higher values ​​indicating a higher degree of matching. The system can calculate the matching score for all pattern categories for each time period and assign the pattern category with the highest score to that time period. By analyzing behavioral pattern sequences and statistically analyzing the frequency and duration of various behavioral patterns across different date types and seasons, habitual electricity usage periods can be extracted. For example, statistics show that the "staying at home in the evening" pattern occurs 5 times per week on winter weekdays, with an average duration of 4 hours, primarily between 18:00 and 22:00, thus forming a record of habitual electricity usage periods. Within the behavioral pattern sequences, frequently co-occurring or sequentially occurring load state combinations are identified to form load linkage rules. For instance, through association rule mining, it was found that when the "air conditioner starts" event occurs in the behavioral pattern sequence, the confidence level of the "living room lights turn off" event occurring within the following 5 minutes is as high as 85%, thus forming a load linkage rule that "air conditioner starting is often accompanied by living room lights turning off." Abnormal periods in the behavioral pattern sequence that show a low match with any typical pattern are marked as abnormal power consumption segments. For example, a period of continuous high-power operation that suddenly occurs at 3 a.m., whose match score with all known typical patterns is below a preset threshold of 0.2, is marked as an abnormal power consumption segment. Habitual power consumption periods, load linkage rules, and abnormal power consumption segments are integrated to form a complete, fine-grained profile of user power consumption behavior.

[0118] Define a template structure for the electricity consumption behavior analysis report. This template structure includes a summary of basic user information, a description of typical behavior patterns, a statistical table of habitual electricity consumption periods, a list of load linkage rules, and a list of abnormal electricity consumption segments. From the user's fine-grained electricity consumption behavior profile, extract specific information on habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments. For example, extract the habitual period for the "winter weekday nighttime mode" as "20:00-07:00," extract the linkage rule "water heater starts one hour after washing machine starts," and extract the abnormal segment "abnormally high energy consumption from 14:00-16:00 on December 24, 2023." From multiple typical electricity consumption behavior patterns, extract descriptive features for each pattern to construct a description of the typical behavior pattern library. For example, decode the cluster center vector into a human-readable description: "Pattern A: Summer weekday daytime, main loads are refrigerator and air conditioner, set temperature 26℃, average power 1.2kW." Fill all the extracted information into the corresponding sections according to the template structure specifications. The completed report is formatted to generate an easy-to-read and parse document, serving as a structured electricity consumption behavior analysis report. The report can be in JSON, XML, or a text document with a specific format.

[0119] Optionally, behavioral pattern sequence mining can utilize sequence pattern mining algorithms, such as the PrefixSpan algorithm, to discover longer and more complex cross-time period behavioral patterns. In addition to matching degree thresholds, the identification of abnormal electricity consumption segments can also incorporate statistical methods, such as building predictive models based on historical behavioral pattern sequences, and classifying periods that significantly deviate from the model's predictions as abnormal. The template for the electricity consumption behavior analysis report can be designed to be customizable, adjusting the included modules and level of detail according to the needs of different users or management platforms.

[0120] It is understandable that the distance metric used in the matching process... and scaling parameters Together, they determine the sensitivity and specificity of the matching; smaller... A larger or more lenient distance threshold will cause more time periods to be categorized into a particular typical pattern, while a larger one will... Higher values ​​or stricter distance thresholds will identify more abnormal segments. The mining of load linkage rules is not limited to co-occurrence within the same time period; it can also focus on sequential relationships across time periods, which helps in understanding the causal or habitual order in user electricity consumption behavior. The key to generating structured reports lies in transforming unstructured behavioral insights from user profiles into standardized, field-based data entries, facilitating automatic reading, storage, and further analysis by the electricity management platform.

[0121] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A system for fine analysis of user electricity consumption behavior based on non-intrusive load monitoring, characterized in that, The system includes: The data acquisition module acquires the total load time-series data at the incoming line of the target power user; The load decomposition module processes the total load time series data using an improved hidden Markov model algorithm to decompose it into a sequence of operating states of multiple independent power loads. The improved hidden Markov model algorithm dynamically adjusts the state transition probability based on the multi-scale mutation characteristics of the load switching events in the total power. The feature extraction and fusion module performs feature extraction processing on the operating state sequence of each independent power load, generates the operating mode feature vector of each load, and fuses the operating mode feature vectors of all independent power loads with the environmental correlation information in the total load time series data to generate a comprehensive power consumption behavior feature matrix. The behavior analysis and summarization module calls a deep learning behavior clustering network to analyze the comprehensive electricity consumption behavior feature matrix, identify and summarize multiple typical electricity consumption behavior patterns of users; The behavior profile building module constructs a fine-grained user electricity behavior profile based on the aforementioned multiple typical electricity consumption behavior patterns through behavior pattern matching and sequence generation technology. The report generation and push module transforms the user's fine-grained electricity consumption behavior profile into a structured electricity consumption behavior analysis report and pushes it to the electricity management platform.

2. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 1, characterized in that, An improved Hidden Markov Model algorithm is applied to the total load time-series data to decompose it into operating state sequences of multiple independent electrical loads, including: The total load time-series data includes voltage waveform sequence, current waveform sequence, and active power sequence; From the total load time series data, local segments of the active power sequence are extracted in a sliding window manner. Multi-scale edge detection is performed on each local segment to identify power abrupt change points at different time scales. The power abrupt change points correspond to potential load switching events. Based on all identified power abrupt change points, the continuous total active power sequence is divided into multiple steady-state power intervals, each of which corresponds to a stable operating state of a set of load combinations. The improved hidden Markov model algorithm is initialized by setting different steady-state power ranges as the implicit states of the algorithm, setting the observed total active power value as the observation state, and initializing the state transition probability matrix and observation probability matrix according to the historical load feature library. By utilizing the improved Hidden Markov Model algorithm and combining the multi-scale mutation features provided by the power mutation point, the state transition probability matrix is ​​dynamically adjusted during the algorithm iteration process, so that state transitions tend to occur near the power mutation point and unreasonable state jumps in the steady state interval are suppressed. Using the Viterbi decoding algorithm, based on the optimized state transition probability matrix and observation probability matrix, the hidden state sequence is calculated, and the hidden state sequence is mapped back to the specific load combination state, thereby decomposing the operating state sequence of each independent power load in reverse.

3. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 2, characterized in that, The method employs the Viterbi decoding algorithm to calculate the hidden state sequence based on the optimized state transition probability matrix and observation probability matrix. This hidden state sequence is then mapped back to specific load combination states, and subsequently, the operating state sequence of each independent electrical load is decomposed in reverse, including: Using the optimized state transition probability matrix and observation probability matrix, combined with the observed total active power sequence, the dynamic programming path probability and state backtracking pointer of the Viterbi decoding algorithm are initialized. At each time step of the total active power sequence, all hidden states are traversed, the maximum probability among all paths to reach each current hidden state is calculated, and the maximum probability value and the state index of the previous moment that generated the maximum probability are stored. After processing the observation data of all time steps, starting from the hidden state with the highest probability in the last time step, the complete sequence of hidden states is obtained by backtracking backward according to the stored state backtracking pointer. Each state in the obtained implicit state sequence is matched with the load combination state represented by the steady-state power range recorded in the historical load feature library to determine the specific load on and off combination at each time point. Based on the specific combination of load on / off states and the power values ​​of each on-state load inferred from the steady-state power range, the operating state of each independent electrical load in the corresponding time period is separated from the total active power signal, generating an independent operating state sequence for each independent electrical load.

4. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 2, characterized in that, The improved Hidden Markov Model algorithm dynamically adjusts the state transition probability based on the multi-scale abrupt changes in total power caused by load switching events, including: During the iterative process of the improved Hidden Markov Model algorithm, the values ​​of the corresponding elements in the state transition probability matrix are dynamically adjusted according to whether there is a power mutation point near the current time point and the time scale of the power mutation point. If the current time point is within the time window of a power mutation point that is identified as a load switching event, the transition probability between load combination states related to the load switching event is increased, while the transition probability between unrelated states is suppressed. The size of the time window for the power mutation point is determined by the edge detection scale used when the power mutation point is detected. A mutation point with a larger time scale corresponds to a longer time window. In each iteration, after dynamically adjusting the state transition probability matrix, it needs to be normalized again to ensure that the sum of the transition probabilities in each row is one.

5. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 1, characterized in that, The process of performing feature extraction on the operating state sequence of each independent electrical load to generate an operating mode feature vector for each load includes: The operating mode feature vector includes start-stop period characteristics, operating power curve shape, and cycle working mode; For a sequence of operating states of an independent electrical load, identify the times of all start-up and stop events; Based on the continuous start and stop event times, the start and stop time period characteristics are extracted, which include the start frequency, average continuous running time and the most frequently running time period interval for each hour of the day. During the duration of each operation, the power segment corresponding to the load is extracted from the total active power sequence, the power segment is standardized and resampled to generate a fixed-length power time series, and the power time series is morphologically described to obtain the shape of the operating power curve. Analyze the repetition pattern of the operating state sequence over a long period of time, identify the operating patterns of daily cycle, weekly cycle or specific date type, and quantify them as the periodic working mode; The start-stop period characteristics, operating power curve shape, and cyclic operating mode are concatenated and encoded into a fixed-length numerical vector, which serves as the operating mode feature vector of the load.

6. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 1, characterized in that, The process involves fusing the operation mode feature vectors of all independent electricity loads with environmental correlation information from the total load time-series data to generate a comprehensive electricity consumption behavior feature matrix, including: The environmental information includes timestamps, temperature, and date type; Extract time information, weekday information, and whether it is a holiday information from the timestamps of the total load time series data to form a time context vector; Temperature information is extracted from environmental data recorded synchronously with the total load time series data and then normalized. For each independent electrical load, its operating mode feature vector is concatenated with the time context vector and temperature information of the corresponding time period to form an enhanced feature vector of the load under a specific environment; The enhanced characteristic vectors of all independent electrical loads within the same or consecutive time periods are arranged in chronological order to form a two-dimensional matrix. The rows of the two-dimensional matrix represent time points or time periods, and the columns represent the characteristics of different loads under specific environments. This two-dimensional matrix is ​​the comprehensive electricity consumption behavior characteristic matrix.

7. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 1, characterized in that, The deep learning behavior clustering network is invoked to analyze the comprehensive electricity consumption behavior feature matrix, identify and summarize multiple typical electricity consumption behavior patterns of users, including: The deep learning behavior clustering network consists of a feature encoder and a cluster allocator, and the feature encoder is a deep autoencoder structure. The comprehensive electricity consumption behavior feature matrix is ​​input into the feature encoder, and the encoder part compresses it into a low-dimensional feature representation vector. The clustering allocator receives the feature representation vector and calculates its distance or association metric with several cluster center vectors maintained within the network. Based on the measurement results, the electricity consumption behavior characteristics of the current time period are softly assigned to each cluster center to obtain the cluster assignment probability distribution; During training, the reconstruction error of the feature encoder and the clustering loss of the cluster allocator are jointly optimized, so that the deep learning behavior clustering network can learn the ability to map consistent electricity consumption behavior features to the same cluster center. After training, the comprehensive electricity consumption behavior feature matrix of all historical time periods is input into the network. Based on the clustering and allocation results, multiple typical electricity consumption behavior patterns of the user are summarized. Each pattern is represented by a cluster center and the feature of all time period samples belonging to the cluster.

8. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 7, characterized in that, The step of jointly optimizing the reconstruction error of the feature encoder and the clustering loss of the clustering allocator during training includes: The comprehensive electricity consumption behavior feature matrix is ​​forward propagated, the feature representation vector is obtained by the encoder, and the input matrix is ​​reconstructed by the decoder. The mean square error between the original input matrix and the reconstructed matrix is ​​calculated as the reconstruction error. Calculate the similarity between the feature representation vector and the student t-distribution of each cluster center vector to obtain the cluster assignment probability distribution. Calculate the KL divergence between the cluster assignment probability distribution and the target distribution as the clustering loss. The target distribution is obtained by increasing the confidence of the assignment probability distribution. The total loss function of the joint optimization is the weighted sum of the reconstruction error and the clustering loss. The parameters of the feature encoder, decoder and cluster center vector are updated through the backpropagation algorithm to minimize the total loss function.

9. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 1, characterized in that, Based on the aforementioned typical electricity consumption behavior patterns, a fine-grained user electricity consumption behavior profile is constructed using behavior pattern matching and sequence generation techniques, including: The user's fine-grained electricity consumption behavior profile includes habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments; From the various typical electricity consumption behavior patterns, representative features of each pattern are extracted. These representative features include the typical operating state combinations of each load under the various typical electricity consumption behavior patterns and the time and environmental characteristics of the occurrence. For new or historical electricity consumption periods of users, the real-time or historical comprehensive electricity consumption behavior feature matrix is ​​matched with the representative features of various patterns to identify the typical electricity consumption behavior pattern category of each period. Connect the typical electricity consumption behavior patterns identified in all time periods in chronological order to form a sequence of user behavior patterns; Analyze the behavioral pattern sequence, and statistically analyze the frequency and duration of various behavioral patterns under different date types and seasons to extract the habitual electricity consumption periods; In the sequence of behavioral patterns, frequently co-occurring or sequentially occurring load state combinations are identified to form the load linkage rules; Identify and mark abnormal time periods in the behavioral pattern sequence that have a low match with any typical pattern as the abnormal power consumption segments. By integrating the habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments, a complete fine-grained profile of the user's electricity consumption behavior is formed.

10. The user electricity consumption behavior fine analysis system based on non-intrusive load monitoring according to claim 1, characterized in that, The user's fine-grained electricity consumption behavior profile is transformed into a structured electricity consumption behavior analysis report, including: Define the template structure of the electricity consumption behavior analysis report. The template structure includes a summary of basic user information, a description of typical behavior patterns, a statistical table of habitual electricity consumption periods, a list of load linkage rules, and a list of abnormal electricity consumption segments. From the user's fine-grained electricity consumption behavior profile, extract specific information on habitual electricity consumption periods, load linkage rules, and abnormal electricity consumption segments; Descriptive features of each type of typical electricity consumption behavior pattern are extracted from the aforementioned patterns to form the typical behavior pattern library description. Fill all the extracted information into the corresponding parts according to the specifications of the template structure; The completed report is formatted to generate an easy-to-read and parse document, which serves as the structured electricity consumption behavior analysis report.

Images