An integrated energy consumption monitoring management system

Abstract

The application provides a comprehensive energy consumption monitoring management system, and relates to the technical field of data processing.The system is used for collecting current values and voltage values of different nodes in a power distribution network, arranging and aligning the current values and voltage values according to time sequences, analyzing parameter change trends of the different nodes, calculating change gradient sequences and fluctuation variances of the nodes, obtaining change trajectory data sets, analyzing change gradient distributions and fluctuation persistences of various change windows, performing clustering processing, dividing the change trajectory data sets into multiple change category clusters, obtaining change index sets, counting change amplitude upper bounds and duration boundaries of the change category clusters, constructing dynamic numerical value judgment intervals according to the change amplitude upper bounds and the duration boundaries, marking operation windows exceeding threshold values, determining node identities and operation window positions, generating operation instructions and performing power-off.The application can monitor current values and voltage values of different electrical appliances in a power distribution network, and can timely turn on and off currents.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a comprehensive energy consumption monitoring and management system. Background Technology

[0002] In existing building or industrial power distribution systems, distribution boxes and cabinets are typically used for zoned control and protection of circuits. These components are widely used to distribute electrical energy and provide short-circuit or overload protection. Their operating status is mostly monitored and managed through manual inspections, current transformer measurements, and power metering. However, this structured distributed monitoring method may not be able to grasp the working status of all nodes in real time, making it difficult to locate abnormal power consumption in specific equipment, thus hindering timely power outages or restoration operations.

[0003] For example, in smart building scenarios, existing technologies cannot accurately monitor the status of individual electrical components such as circuit breakers and distribution boxes, such as their opening and closing status or abnormal arcing. They also cannot automatically make power-off decisions based on real-time load data, which may result in the inability to cut off local power in a timely manner when the load fluctuates abnormally, thus affecting the energy management capabilities of the entire building. Summary of the Invention

[0004] The purpose of this invention is to provide a comprehensive energy consumption monitoring and management system, which aims to solve the problems mentioned in the background art.

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

[0006] A comprehensive energy consumption monitoring and management system, the system comprising:

[0007] The data acquisition module is used to collect the current and voltage values ​​of different nodes in the power distribution network, and arrange and align them in chronological order to obtain a node dataset.

[0008] The feature module is used to analyze the parameter change trends of different nodes based on the node dataset, calculate the change gradient sequence and fluctuation variance of each node, and divide them into continuous change windows to obtain the change trajectory dataset.

[0009] The clustering module is used to analyze the gradient distribution and fluctuation persistence of each change window based on the change trajectory dataset and perform clustering processing to divide the change trajectory dataset into multiple change category clusters and obtain a change index set;

[0010] The threshold module is used to calculate the upper bound of the change magnitude and the duration boundary of each change category cluster based on the change index set, obtain boundary combination data, and construct a dynamic numerical judgment interval based on it to obtain the node judgment threshold set.

[0011] The identification module is used to collect the real-time dataset of each node, compare it with the node judgment threshold set, mark the running windows that exceed the threshold, and obtain the anomaly identification result set.

[0012] The instruction module is used to determine the node identifier and running window position based on the anomaly identification result set, generate operation instructions, and execute a power-off.

[0013] Furthermore, the feature modules include:

[0014] The gradient calculation unit is used to calculate the parameter difference value based on the current value and voltage value at adjacent time points in the node dataset, and obtain the initial change sequence;

[0015] The normalization transformation unit is used to perform linear normalization on each parameter based on the maximum and minimum values ​​in the initial change sequence to obtain the change gradient sequence;

[0016] The trend recognition unit is used to identify the time points of positive and negative changes in parameter difference values ​​based on the changing gradient sequence, and to mark the corresponding gradient direction inflection points to obtain a set of gradient inflection points.

[0017] The fluctuation assessment unit is used to divide the changing gradient sequence into different sub-regions based on the gradient inflection point set, and calculate the variance value within each sub-region to obtain the fluctuation variance set.

[0018] Furthermore, the feature module also includes:

[0019] The window partitioning unit is used to divide the data of each node into multiple continuously changing windows according to the inflection point position in the gradient change sequence, so as to obtain the initial window structure.

[0020] The window filtering unit is used to remove windows with low sample counts based on the minimum data length of each window in the initial window structure, thereby obtaining a set of effective windows.

[0021] The window feature aggregation unit is used to obtain a group of window feature vectors based on the average gradient value and fluctuation variance value of each window in the effective label index set window set;

[0022] The trajectory stitching unit is used to stitch together window feature vector groups in chronological order, determine the time sequence number of each window feature vector, and obtain the change trajectory dataset.

[0023] Furthermore, the clustering module includes:

[0024] The gradient feature extraction unit is used to extract the average gradient value and maximum gradient fluctuation amplitude of each change window based on the time series of each change window in the change trajectory dataset, and obtain the gradient feature group.

[0025] The persistent feature extraction unit is used to calculate the duration of each change window based on the start and end time points, and to count the number of gradient direction transformations within the window to obtain persistent feature groups.

[0026] The feature fusion unit is used to construct a multi-dimensional vector structure for each change window based on the gradient feature set and the persistent feature set, thereby obtaining the original feature vector set;

[0027] The normalization unit is used to normalize each vector based on the range of each dimension vector in the original feature vector set to obtain a standard feature vector set.

[0028] Furthermore, the clustering module also includes:

[0029] The density assessment unit is used to calculate the neighborhood density value of each vector based on the standard feature vector set, and obtain the density score sequence.

[0030] The category center identification unit is used to extract the top three positions of the neighborhood density value from the density score sequence and define them as the initial category centers.

[0031] The clustering partitioning unit is used to divide the standard feature vector set into multiple variable category clusters by calculating the distance between each vector and each initial category center and performing clustering processing, thereby obtaining the clustering assignment result;

[0032] The index generation unit is used to label each change category cluster with a category number based on the clustering results and align it with the corresponding time series to obtain a change index set.

[0033] Furthermore, the density assessment unit includes:

[0034] The neighborhood density value calculation unit is used to extract the fluctuation variance, average gradient value, duration length and time sequence number of each vector in the change window before normalization processing, based on the standard feature vector set.

[0035] Based on the standard eigenvector set vectors and their fluctuation variance values, the distance between different vectors is calculated to measure the spatial proximity between different vectors, resulting in a distance exponential decay term; based on the average gradient value of the vectors, the effect of gradient direction differences on density coupling between different vectors is calculated, resulting in a gradient similarity modulation term; based on the duration length of the vectors, a harmonic average is performed on vectors with similar duration lengths, resulting in a time-consistent weighting term; based on the time sequence number of the vectors, the influence of vectors in distant time periods on each other is controlled, resulting in a time proximity suppression term.

[0036] The neighborhood density value is obtained by fusing the distance exponential decay term, gradient similarity modulation term, time consistency weighting term, and time proximity suppression term.

[0037] Furthermore, the threshold module includes:

[0038] The amplitude statistics unit is used to extract the maximum gradient value within the corresponding change window for each change category cluster in the change index set, and obtain the upper bound set of change amplitude.

[0039] The time extraction unit is used to calculate the time span and calculate the range based on the start and end times of each change window within each change category cluster, thereby obtaining the duration boundary set.

[0040] Boundary pairing unit is used to combine the upper bound set of change amplitude and the boundary set of duration according to the time order of the change window to obtain boundary pairing data;

[0041] The boundary integration unit is used to establish a mapping relationship between the pairing data of each boundary and the category number of the change category cluster to obtain the boundary combination data.

[0042] Furthermore, the threshold module also includes:

[0043] The amplitude interval unit is used to extract the upper bound of the change amplitude of each group based on the boundary combination data, and calculate the upper and lower limit intervals of the gradient of the change category cluster by combining the average gradient value of each window within the corresponding change category cluster, thus obtaining the amplitude determination interval group.

[0044] The time interval unit is used to extract the duration boundary of each group based on the boundary combination data, and expand its minimum duration period boundary by combining the standard deviation of the time length within the corresponding change category cluster to obtain the time determination interval group.

[0045] The dual-interval fusion unit is used to pair amplitude determination interval groups and time determination interval groups according to category number to construct a two-dimensional determination interval combination set containing gradient upper and lower limits and time upper and lower limits;

[0046] The node mapping unit is used to allocate the two-dimensional judgment interval combination set to the corresponding node identifier according to the node identifier in the change index set, so as to obtain the node judgment threshold set.

[0047] Furthermore, the identification module includes:

[0048] The normalization processing unit is used to standardize the current and voltage values ​​according to the real-time dataset of each node, so as to obtain a normalized real-time dataset.

[0049] The difference calculation unit is used to calculate the deviation between the current value and the upper bound at each time point based on the normalized real-time dataset and the node judgment threshold set, and obtain the threshold difference vector.

[0050] The over-limit monitoring unit is used to mark the time periods in the threshold difference vector that continuously exceed the upper bound of the gradient and meet the minimum duration as abnormal windows, thus obtaining a preliminary set of abnormal windows.

[0051] The anomaly integration unit is used to merge anomaly windows in the initial anomaly window set whose adjacent intervals are less than a preset recovery period, and integrate them according to node identifiers to obtain an anomaly identification result set.

[0052] Furthermore, the instruction module includes:

[0053] The node localization unit is used to find the associated node identifier in the change index set according to the index position corresponding to each abnormal window in the abnormal identification result set, and obtain the abnormal node list.

[0054] The action determination unit is used to calculate the cumulative abnormal cycle value of each node based on the number of abnormal windows and time distribution of each node in the abnormal node list, and compare it with the preset power outage threshold to obtain the determination result data.

[0055] The parameter assembly unit is used to determine the node address, power-off method and execution time of each trigger node based on the trigger node number in the judgment result data, and obtain the execution parameter data.

[0056] The instruction generation unit is used to encapsulate the execution parameter data into a recognizable control instruction structure and output it to the execution interface to complete the power-off operation of the trigger node.

[0057] The above-described solution of the present invention has at least the following beneficial effects:

[0058] This invention overcomes the problems of information fragmentation and low timeliness caused by relying on manual inspection and single-point measurement in existing technologies by acquiring the current and voltage values ​​of multiple nodes in a power distribution network at different times and arranging them in chronological order to construct a node dataset. Through synchronous data acquisition from all nodes, it not only improves the visualization of the power system's operating status but also enhances data continuity and integrity, ensuring that subsequent processing modules can obtain unified structured data across multiple nodes and time periods. Compared with existing power distribution metering instruments or transformer solutions, this system can achieve more refined energy consumption structure analysis, providing solid data support for anomaly identification and dynamic modeling, thereby better meeting the management needs of large-scale power supply and distribution systems.

[0059] This invention models the evolution trend of electrical parameters at different nodes in the time dimension by introducing a changing gradient sequence and fluctuation variance. It divides the continuously changing window and generates a change trajectory dataset. The system can accurately reflect the micro-change characteristics of the equipment's operating status. Especially in scenarios with frequent nonlinear load fluctuations, it can effectively identify hidden phenomena such as slow transitions before sudden changes and periodic anomalies. It avoids the misjudgment and missed judgment problems caused by static thresholds or excessively large sampling intervals in existing systems. The system has time sensitivity and multi-dimensional expression capabilities and can also run stably in edge computing frameworks, which helps to realize lightweight energy consumption change monitoring at the edge.

[0060] This invention utilizes gradient distribution and fluctuation persistence in changing trajectories, combined with density and gradient features, to perform multidimensional clustering. It can automatically extract features based on the evolution of node trajectories. This mechanism is superior to traditional classification methods that rely on empirical rules or manually set templates. It can automatically identify potential high-energy-consuming categories, anomalous fluctuation behaviors, and other abnormal clusters without requiring a large number of labeled samples. In the event of load disturbances or node impacts in the power supply and distribution network, the system can identify abnormal signs in advance through the evolution trend of cluster categories, thereby improving the energy consumption management system's ability to predict emergencies.

[0061] This invention extracts the upper bound of amplitude and the boundary of duration by clustering the change category clusters. By constructing a two-dimensional judgment interval that includes the amplitude and duration of electrical parameter changes, it effectively overcomes the false alarm or missed alarm problem caused by the static fixed threshold of traditional energy management systems. In power grids where node loads exhibit periodic or intermittent characteristics, the system can automatically adapt to threshold changes based on category characteristics, significantly enhancing the system's understanding of complex fluctuation behavior. The judgment interval supports online dynamic adjustment, improving the system's versatility and transferability in multiple scenarios, and ensuring that the system can adaptively control and make high-precision predictions.

[0062] This invention combines real-time data with node judgment threshold sets, employing normalization and difference monitoring mechanisms to continuously track abnormal intervals in node operating status over time. It also introduces a dual-constraint mechanism to filter abnormal windows, effectively reducing the risk of false alarms caused by short-term fluctuations or instantaneous jumps in existing technologies. Through abnormal window merging and node integration processing, the system can construct a complete chain of abnormal events, providing a basic basis for power outage response and energy consumption alarms. Compared to existing methods that only compare maximum or average values, this mechanism significantly enhances adaptability to complex fluctuation scenarios, enabling the identification and processing of periodic mutations and hidden anomalies, thereby improving the safety and stability of power distribution system operation. Attached Figure Description

[0063] Figure 1 This is a flowchart of a comprehensive energy consumption monitoring and management system provided by an embodiment of the present invention. Detailed Implementation

[0064] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0065] like Figure 1 As shown, an embodiment of the present invention proposes a comprehensive energy consumption monitoring and management system, the system comprising:

[0066] The data acquisition module is used to collect the current and voltage values ​​of different nodes in the power distribution network, and arrange and align them in chronological order to obtain a node dataset.

[0067] The feature module is used to analyze the parameter change trends of different nodes based on the node dataset, calculate the change gradient sequence and fluctuation variance of each node, and divide them into continuous change windows to obtain the change trajectory dataset.

[0068] The clustering module is used to analyze the gradient distribution and fluctuation persistence of each change window based on the change trajectory dataset and perform clustering processing to divide the change trajectory dataset into multiple change category clusters and obtain a change index set;

[0069] The threshold module is used to calculate the upper bound of the change magnitude and the duration boundary of each change category cluster based on the change index set, obtain boundary combination data, and construct a dynamic numerical judgment interval based on it to obtain the node judgment threshold set.

[0070] The identification module is used to collect the real-time dataset of each node, compare it with the node judgment threshold set, mark the running windows that exceed the threshold, and obtain the anomaly identification result set.

[0071] The instruction module is used to determine the node identifier and running window position based on the anomaly identification result set, generate operation instructions, and execute a power-off.

[0072] In this embodiment of the invention, the acquisition module is used to acquire current and voltage values ​​of different nodes in the power distribution network, and arrange and align them in chronological order to obtain a node dataset. By synchronously acquiring multiple nodes and multiple parameters, the data accuracy and real-time performance are improved, and the accuracy of parameter comparison is ensured by time-series alignment, providing a data foundation with continuous time sequence and unified dimensions. The feature module is used to analyze the parameter change trends of different nodes based on the node dataset, calculate the change gradient sequence and fluctuation variance of each node, and divide it into continuous change windows to obtain a change trajectory dataset. This can accurately capture the state transition points in equipment operation and overcome the limitations of traditional methods based on the absolute value of static electrical parameters. The clustering module is used to analyze the change gradient distribution and fluctuation persistence of each change window based on the change trajectory dataset and perform clustering processing. The change trajectory dataset is divided into multiple change category clusters to obtain a change index set, which effectively copes with the complexity of energy consumption caused by the variety of equipment or differences in operating habits. Through the indexing mechanism, the subsequent system can directly reference the cluster categories for dynamic threshold modeling.

[0073] The threshold module is used to statistically analyze the upper bound of the change amplitude and the duration boundary of each change category cluster based on the change index set, obtain boundary combination data, and construct a dynamic numerical judgment interval based on it to obtain the node judgment threshold set. This effectively breaks through the limitation of traditional static threshold for anomaly judgment. Through dynamic interval setting, the system can adaptively adjust the judgment criteria according to power fluctuations. The identification module is used to collect the real-time dataset of each node and compare it with the node judgment threshold set. It marks the running window that exceeds the threshold and obtains the anomaly identification result set. This can effectively improve the ability to identify equipment failures or power anomalies under complex conditions and provide accurate basis for automatic control. The instruction module is used to determine the node identifier and running window position based on the anomaly identification result set, generate operation instructions and execute power cut-off. Power cut-off control can be quickly completed after an abnormal node is triggered.

[0074] The acquisition module is used to collect current and voltage values ​​from different nodes in the power distribution network, and arrange them in chronological order to obtain a node dataset, which specifically includes:

[0075] The data acquisition module is installed at the front end of the integrated energy consumption monitoring and management system. Its function is to acquire current and voltage values ​​from multiple physical nodes in the power distribution network. Therefore, sampling-capable electrical parameter acquisition units need to be deployed at key nodes in the power distribution network (including but not limited to switchgear inside distribution boxes, cable inlets and outlets, terminal sockets, and load device access points). Each acquisition unit integrates a current transformer and a voltage divider module, enabling real-time sampling of the instantaneous current and voltage values ​​at its location. The acquisition unit uses a local microcontroller to perform A / D conversion on the raw analog signals to obtain digital sampled data; simultaneously, the acquisition unit has built-in unified time synchronization logic, which can synchronize with the system master station's time base to ensure that the data from multiple nodes has a unified timestamp format.

[0076] Each set of current and voltage data items acquired by the acquisition unit is denoted as follows: and ,in Number the nodes. The timestamp is used for data sampling. Each acquisition unit samples data at a fixed period (e.g., once per second or higher) and uploads the data to the central aggregation unit of the acquisition module via a communication interface. In the aggregation unit, the system binds each node number to its uploaded data and sorts and aligns it according to its timestamp. Since there may be millisecond-level latency between multiple nodes during data upload, to ensure data consistency across all nodes within the same sampling period, the system uses a window sliding mechanism to merge data with different timestamps. For example, with a sampling period of 1 second and a tolerance range of ±10ms, all node data falling within a certain second window are uniformly merged into the corresponding time period dataset. A node data matrix is ​​then constructed based on that second's timestamp. Finally, the acquisition module merges all nodes... At each sampling time current value With voltage value Organize the data into a time-aligned node dataset structure.

[0077] In a preferred embodiment of the present invention, the feature module includes:

[0078] The gradient calculation unit is used to calculate the parameter difference value based on the current value and voltage value at adjacent time points in the node dataset, and obtain the initial change sequence;

[0079] The normalization transformation unit is used to perform linear normalization on each parameter based on the maximum and minimum values ​​in the initial change sequence to obtain the change gradient sequence;

[0080] The trend recognition unit is used to identify the time points of positive and negative changes in parameter difference values ​​based on the changing gradient sequence, and to mark the corresponding gradient direction inflection points to obtain a set of gradient inflection points.

[0081] The fluctuation assessment unit is used to divide the changing gradient sequence into different sub-regions based on the gradient inflection point set, and calculate the variance value within each sub-region to obtain the fluctuation variance set.

[0082] In this embodiment of the invention, the gradient calculation unit is used to calculate the parameter difference value based on the current and voltage values ​​at adjacent time points in the node dataset, obtain the initial change sequence, realize the discrete extraction of the time-series changes of electrical parameters, and effectively reflect the power fluctuation trend of each node between adjacent time points; the normalization conversion unit is used to perform linear normalization processing on each parameter based on the maximum and minimum values ​​in the initial change sequence, obtain the change gradient sequence, standardize the differences of each parameter, and eliminate the influence of differences in physical characteristics or load size in the original collected values; the trend identification unit is used to identify the time points of positive and negative changes in parameter difference values ​​based on the change gradient sequence, and mark the corresponding gradient direction inflection points, obtain the gradient inflection point set, which can efficiently extract key inflection points in the change curve and help identify the stage characteristics in energy consumption behavior; the fluctuation evaluation unit is used to divide the change gradient sequence into different sub-regions based on the gradient inflection point set, and calculate the variance value in each sub-region to obtain the fluctuation variance set, which provides a reliable numerical basis for subsequent change trajectory clustering and threshold generation.

[0083] The normalization transformation unit is used to perform linear normalization on each parameter based on the maximum and minimum values ​​in the initial change sequence to obtain the change gradient sequence, specifically including:

[0084] The system first iterates through the initial change sequence of each node at each time step, extracting the minimum and maximum values ​​for the current and voltage change sequences, respectively. The system then sets the node's current difference sequence as... The voltage differential sequence is set as Then, find their maximum values ​​respectively. , and minimum value , .

[0085] After obtaining the above extreme values, the system performs a standardization operation on each difference value according to the following linear normalization formula: for any difference value Its normalization result is in The normalized value can be any difference between current and voltage, and the normalized result is confined to the closed interval [0,1]. The normalized current change gradient sequence and voltage change gradient sequence will respectively constitute the feature sequence group used in subsequent processing. The entire normalization process will complete a full parameter normalization during system initialization, and will dynamically update the extreme value range and recalculate the normalization result after each new sampled data to ensure the scale consistency of data features.

[0086] The trend recognition unit is used to identify the time points of positive and negative changes in parameter difference values ​​based on the changing gradient sequence, and to mark the corresponding gradient direction inflection points to obtain a gradient inflection point set, specifically including:

[0087] After receiving the changing gradient sequence, the trend identification unit performs a direction determination operation on the trend change of the parameter difference value. Its main goal is to identify the key time points where the gradient value turns from positive to negative, thereby reflecting the potential boundaries or stage transitions of node operation status changes. During processing, the system first performs a linear traversal of the normalized gradient sequence. Let the current processing sequence be... and with Represents any time The gradient value. For any two consecutive gradient values... and The system determines its product. The sign of the result is determined by whether the result is less than 0. This indicates that the gradient direction has reversed, suggesting a trend change in the node's operating state at that point in time.

[0088] The system records all time points that satisfy the gradient direction inflection condition as a candidate inflection point set, and further filters out noise points by utilizing the trend continuity of neighboring points. Specifically, if a gradient change satisfies the condition that the product is negative but its magnitude is close to 0, and there is no continuous change trend before and after that moment, the system can identify it as a high-frequency noise point and remove it. For the retained inflection points, the system assigns a unified timestamp index number and generates a gradient inflection point set, which will be used for subsequent window division and regional volatility assessment.

[0089] In a preferred embodiment of the present invention, the feature module further includes:

[0090] The window partitioning unit is used to divide the data of each node into multiple continuously changing windows according to the inflection point position in the gradient change sequence, so as to obtain the initial window structure.

[0091] The window filtering unit is used to remove windows with low sample counts based on the minimum data length of each window in the initial window structure, thereby obtaining a set of effective windows.

[0092] The window feature aggregation unit is used to obtain a set of window feature vectors based on the average gradient value and fluctuation variance value of each window in the effective window set.

[0093] The trajectory stitching unit is used to stitch together window feature vector groups in chronological order, determine the time sequence number of each window feature vector, and obtain the change trajectory dataset.

[0094] In this embodiment of the invention, the window partitioning unit is used to divide the data of each node into multiple continuously changing windows according to the inflection point position in the gradient sequence, to obtain a preliminary window structure, effectively distinguishing the boundary between local drastic fluctuations and the overall stable trend, and improving the fine granularity of data expression; the window filtering unit is used to remove low sample number windows according to the minimum data length of each window in the preliminary window structure, to obtain an effective window set, and to exclude atypical, extremely small-scale, and easily misjudged interference windows without affecting the capture of the overall trend; the window feature summarization unit is used to obtain a window feature vector group according to the average gradient value and fluctuation variance value of each window in the effective window set, retaining both the numerical change amplitude (average gradient) and the fluctuation degree (variance); the trajectory stitching unit is used to stitch the window feature vector group in chronological order, to determine the time sequence number of each window feature vector, to obtain a change trajectory dataset, and to realize the temporal coherence construction of window-level features.

[0095] The window filtering unit is used to remove windows with low sample counts based on the minimum data length of each window in the initial window structure, thereby obtaining a set of effective windows. Specifically, it includes:

[0096] The system first counts the number of valid data points contained within each window in the initial window structure. These data points are the smallest units that form the basis for calculating window features, typically including the current gradient value and voltage gradient value at each time point. During the statistical process, the system sets a uniform sample size threshold for each window; this threshold value is... (For example, setting it to 5) means that it will only be valid if the number of samples is ≥ Only windows with a certain number of data points are considered structural units suitable for subsequent feature analysis. The system performs a sample size evaluation operation on all windows sequentially. If the number of data points in a window is less than a certain value, the window is considered a structural unit suitable for subsequent feature analysis. If the window is invalid, it will be marked as invalid and removed from the data structure.

[0097] To prevent boundary effects from causing sparse window samples, the system further introduces a sliding calibration mechanism. After initial window segmentation, the system allows each window to expand or contract at one or two time points based on the sample situation of its preceding and following windows, thereby improving the effective window retention rate. After completing sample quantity judgment and window screening, the system incorporates all remaining qualified windows into the effective window set and reconstructs their numbering and time index to form a new data index structure. The data in the effective window set will serve as input for subsequent window feature summarization, ensuring that all windows participating in statistical analysis have sufficient sample support, avoiding the amplification of statistical errors caused by small sample windows, and thus improving the accuracy and stability of subsequent change pattern recognition.

[0098] The trajectory stitching unit is used to stitch together window feature vector groups in chronological order, determine the time sequence number of each window feature vector, and obtain the change trajectory dataset, specifically including:

[0099] The system first classifies the window feature vector groups, and divides them into subsets of the corresponding nodes based on the node identifier field attached to each feature vector. Then, under each node, the system sorts the window feature vector groups from earliest to latest based on the window start timestamp as the main sorting criterion. If two windows have overlapping times, the window with the larger time span is retained first, or the window coverage priority is adjusted according to the fluctuation variance when necessary to ensure the continuity and representativeness of the stitching.

[0100] After sorting, the system assigns a time series number to the sorted window feature vector within each node, starting from... Initially, the time series numbers are sequentially incremented to form a clear time mapping structure. Each time series number is simultaneously bound to the average gradient value and the variance of the window, forming a structure like... , The system generates a triplet time-series vector set. Furthermore, it adds trajectory metadata to each node trajectory, including start time, end time, total number of windows, and node number, facilitating rapid subsequent indexing.

[0101] During the stitching process, the system employs a stitching integrity verification mechanism to ensure that there are no invalid gaps in the start and end times of any two adjacent windows. If a time gap is detected exceeding a set threshold, the system will take corrective action. If there is one sampling period, the system automatically inserts placeholder vectors or interpolates in the original features to construct connecting windows to maintain the continuity of the trajectory time sequence. The final spliced ​​change trajectory dataset is organized according to the node number and structurally presents as a sequence of multiple feature vectors with time sequence labels.

[0102] In a preferred embodiment of the present invention, the clustering module includes:

[0103] The gradient feature extraction unit is used to extract the average gradient value and maximum gradient fluctuation amplitude of each change window based on the time series of each change window in the change trajectory dataset, and obtain the gradient feature group.

[0104] The persistent feature extraction unit is used to calculate the duration of each change window based on the start and end time points, and to count the number of gradient direction transformations within the window to obtain persistent feature groups.

[0105] The feature fusion unit is used to construct a multi-dimensional vector structure for each change window based on the gradient feature set and the persistent feature set, thereby obtaining the original feature vector set;

[0106] The normalization unit is used to normalize each vector based on the range of each dimension vector in the original feature vector set to obtain a standard feature vector set.

[0107] In this embodiment of the invention, the gradient feature extraction unit is used to extract the average gradient value and maximum gradient fluctuation amplitude of each change window based on the time series of each change window in the change trajectory dataset, thereby obtaining a gradient feature group. This enhances the system's ability to distinguish abnormal trends of different power nodes and provides a unified input format for subsequent multidimensional clustering, improving the structure of the data. The persistence feature extraction unit is used to calculate the duration of each change window based on its start and end times and count the number of gradient direction changes within the window, thereby obtaining a persistence feature group. This group can identify the stability and persistence characteristics of the change process. The feature fusion unit is used to construct a multidimensional vector structure for each change window based on the gradient feature group and the persistence feature group, thereby obtaining an original feature vector set. This enhances the system's feature expression capability and facilitates the formation of a decision region with clear boundaries. The normalization processing unit is used to normalize each vector based on the range of each dimension vector in the original feature vector set, thereby obtaining a standard feature vector set. This eliminates the dimensional differences between different dimensions, ensuring that each feature dimension has equivalent weights in subsequent distance calculations and density assessments, and preventing certain indicators from dominating the clustering process due to their large numerical range.

[0108] The gradient feature extraction unit is used to extract the average gradient value and maximum gradient fluctuation amplitude of each change window based on the time series of each change window in the change trajectory dataset, thereby obtaining a gradient feature group, which specifically includes:

[0109] For each variation window, the average gradient value within that window is first extracted based on its internal gradient sequence, specifically the normalized current or voltage gradient values ​​from the preceding module. More specifically, the average gradient value is obtained by summing the normalized gradient values ​​corresponding to all sampling points within the variation window and dividing by the total number of sampling points. This quantitatively reflects the overall increasing or decreasing trend of changes within the window. Next, the maximum and minimum values ​​of the normalized gradient within the window are extracted, and the difference between them is calculated. This value is defined as the maximum gradient fluctuation amplitude, used to characterize the severity of numerical fluctuations within the window. Through the above processing, each change window will form a binary feature set. This result constitutes the entire gradient feature set, providing the input basis for subsequent high-dimensional vector construction and cluster modeling.

[0110] The feature fusion unit is used to construct a multi-dimensional vector structure for each change window based on the gradient feature set and the persistent feature set, thereby obtaining the original feature vector set, specifically including:

[0111] Gradient feature groups provide Two metrics, while the persistence feature group includes the duration of each window. and the number of gradient direction changes that occur within that window. The feature fusion process concatenates and combines these four scalar values ​​to form a four-dimensional feature vector. During the data stitching process, to ensure accurate data matching, the system uses the time index of the change window as the primary key to align the feature information from the two sources. This ensures that each high-dimensional feature vector accurately represents the behavioral characteristics of a specific change window. The complete set of feature vectors is collectively referred to as the original feature vector set, which plays a crucial role in subsequent normalization and clustering operations. Through the feature fusion process, joint modeling of the same window across multiple dimensions, such as amplitude characteristics, fluctuation degree, persistence, and directional complexity, can be achieved, providing high-dimensional data support for further behavioral pattern recognition.

[0112] The normalization unit is used to normalize each vector based on the range of each dimension vector in the original feature vector set, to obtain a standard feature vector set, specifically including:

[0113] The normalization operation employs the range normalization method, which maps each feature value to the [0,1] interval through a linear transformation. In practice, the system first iterates through each feature dimension of the original feature vector set and calculates the maximum value of that dimension across all samples. and minimum value Thus obtaining the range Subsequently, for each dimension of the eigenvalues ​​of each vector in the original vector... Normalization is performed using the following formula: After processing, the four components of each vector will be compressed into the [0,1] interval, forming a standard feature vector set.

[0114] In a preferred embodiment of the present invention, the clustering module further includes:

[0115] The density assessment unit is used to calculate the neighborhood density value of each vector based on the standard feature vector set, and obtain the density score sequence.

[0116] The category center identification unit is used to extract the top three positions of the neighborhood density value from the density score sequence and define them as the initial category centers.

[0117] The clustering partitioning unit is used to divide the standard feature vector set into multiple variable category clusters by calculating the distance between each vector and each initial category center and performing clustering processing, thereby obtaining the clustering assignment result;

[0118] The index generation unit is used to label each change category cluster with a category number based on the clustering results and align it with the corresponding time series to obtain a change index set.

[0119] In this embodiment of the invention, the density evaluation unit is used to calculate the neighborhood density value of each vector according to the standard feature vector set, and obtain a density score sequence, which effectively enhances the clustering recognition ability of the system under the condition of uneven data distribution and avoids misjudgment or omission of edge clusters; the category center identification unit is used to extract the top three positions of the neighborhood density value according to the density score sequence and define them as the initial category centers, using high-density points as initial centers to avoid the instability problem caused by random selection of centers; the clustering division unit is used to divide the standard feature vector set into multiple change category clusters by calculating the distance between each vector and each initial category center and performing clustering processing, and obtain the clustering assignment result, which gathers a large number of scattered node change features into a finite number of clusters, providing a foundation for subsequent judgment interval construction and abnormal behavior modeling; the index generation unit is used to label each change category cluster with a category number according to the clustering assignment result and align it with the corresponding time series to obtain a change index set, realize the mapping from the time series to the behavior label space, and establish the time-change category association.

[0120] The clustering partitioning unit is used to divide the standard feature vector set into multiple variable category clusters by calculating the distance between each vector and the centers of each initial category and performing clustering processing, thereby obtaining the clustering assignment result. Specifically, it includes:

[0121] The system first obtains the vector indices and numerical coordinates of all initial class centers in the standard feature vector set based on the outputs of the density evaluation unit and the class center identification unit. For any vector in the standard feature vector set, the Euclidean distance between it and all initial class center vectors is calculated using the following formula: Let the first vector be... The vectors are , No. The category center vectors are Then the distance between the two Expressed as: ,in For the number of dimensions, For the first A vector, For the first For each cluster center vector, the system calculates its distance to all cluster centers and selects the center with the smallest distance as the vector's cluster. This operation is performed across the entire set of standard feature vectors to obtain the cluster number for each vector. To improve the stability of the clustering results, the system can use a combination of initial clustering and multiple rounds of iterative optimization. After the initial clustering, the geometric mean centroid of all vectors within each cluster is recalculated as the new cluster center, and the clustering relationships of all vectors are redistributed based on the new centroids. This process is repeated iteratively until the clustering of each vector no longer changes or the maximum number of iterations is reached.

[0122] The index generation unit is used to assign a category number to each change category cluster based on the clustering results, and align it with the corresponding time series to obtain a change index set, specifically including:

[0123] Based on the clustering results, the system assigns a corresponding category number to each standard feature vector, denoted as the category label. Since each feature vector's change window number and start and end times are recorded during the generation phase, this information can be used to bind the category label to the time axis. Specifically, the system extracts the node identifier to which each feature vector belongs from the original trajectory data. and its corresponding start and end time window periods and assign category number Pair them together to generate triplet data items. This triple constitutes a single change index record.

[0124] The system sorts and organizes all triples chronologically, and then categorizes and aggregates them by node number, generating a complete change trajectory index table for each node. This index table marks the behavior categories of different stages with time slices as the dimension, which helps to build a node-level time-series behavior model and serves as the input basis for category assignment in the subsequent threshold extraction module. Simultaneously, the system can organize and store the change index set in a mapping structure using a key-value pair structure, where the key is the node number and the value is the time window corresponding to that node and its list of change categories, enabling rapid retrieval and access to the behavior tags of different nodes.

[0125] In a preferred embodiment of the present invention, the density assessment unit includes:

[0126] The neighborhood density value calculation unit is used to extract the fluctuation variance, average gradient value, duration length and time sequence number of each vector in the change window before normalization processing, based on the standard feature vector set.

[0127] Based on the standard eigenvector set vectors and their fluctuation variance values, the distance between different vectors is calculated to measure the spatial proximity between different vectors, resulting in a distance exponential decay term; based on the average gradient value of the vectors, the effect of gradient direction differences on density coupling between different vectors is calculated, resulting in a gradient similarity modulation term; based on the duration length of the vectors, a harmonic average is performed on vectors with similar duration lengths, resulting in a time-consistent weighting term; based on the time sequence number of the vectors, the influence of vectors in distant time periods on each other is controlled, resulting in a time proximity suppression term.

[0128] The neighborhood density value is obtained by fusing the distance exponential decay term, gradient similarity modulation term, time consistency weighting term, and time proximity suppression term.

[0129] In this embodiment of the invention, the neighborhood density value calculation unit is used to extract the fluctuation variance, average gradient value, duration, and time series number of each vector in the change window before normalization processing, based on the standard feature vector set. This visualizes the time-series energy consumption change characteristics in numerical space, providing an input basis for multi-dimensional coupled calculations. Based on the vectors and their fluctuation variance values ​​in the standard feature vector set, the unit calculates the distance between different vectors, measuring their spatial proximity and obtaining a distance exponential decay term. This term measures the proximity of two change windows in the energy consumption fluctuation space, achieving basic modeling of physical or behavioral similarity, suppressing the influence of long-distance interference, and highlighting local correlation. Based on the average gradient value of the vectors, the unit calculates the effect of gradient direction differences on density coupling between different vectors, obtaining a gradient similarity modulation term. This emphasizes the consistency of change direction and avoids mixing vectors with opposite positive and negative directions but similar amplitudes. The behavior is categorized into similar categories, thereby improving the directional accuracy of behavior recognition. Based on the duration of the vectors, a harmonic average is performed on vectors with similar durations to obtain a time-consistent weighting term, which enhances the sensitivity to continuous behavior recognition and avoids statistical bias caused by extremely short time windows, showing good recognition effect for long-term slowly changing behavior. Based on the temporal sequence number of the vectors, the influence of vectors between distant time periods on each other is controlled to obtain a time proximity suppression term, which suppresses erroneous associations between vectors between distant time periods, effectively maintaining the temporal continuity boundary and improving the accuracy of temporal recognition. By fusing the distance exponential decay term, gradient similarity modulation term, time-consistent weighting term, and time proximity suppression term, a neighborhood density value is obtained, which retains spatial geometric structure information and introduces dynamic features of behavior and temporal relationships, making the subsequent category center recognition more representative and significantly improving the clustering accuracy.

[0130] The formula for calculating the domain density value is as follows:

[0131] ,

[0132] in, For the first The neighborhood density value of each vector. and These are the indices of the vector. The total number of vectors in the standard eigenvector set. For the first A vector, For the first A vector, For the first The fluctuation variance of each vector. For the first The fluctuation variance of each vector. For the first The average gradient value of each vector. For the first The average gradient value of each vector. For the first The duration of each vector. For the first The duration of each vector. For the first The timing index of each vector, For the first The timing index of each vector, It is a constant.

[0133] In a preferred embodiment of the present invention, the threshold module includes:

[0134] The amplitude statistics unit is used to extract the maximum gradient value within the corresponding change window for each change category cluster in the change index set, and obtain the upper bound set of change amplitude.

[0135] The time extraction unit is used to calculate the time span and calculate the range based on the start and end times of each change window within each change category cluster, thereby obtaining the duration boundary set.

[0136] Boundary pairing unit is used to combine the upper bound set of change amplitude and the boundary set of duration according to the time order of the change window to obtain boundary pairing data;

[0137] The boundary integration unit is used to establish a mapping relationship between the pairing data of each boundary and the category number of the change category cluster to obtain the boundary combination data.

[0138] In this embodiment of the invention, the amplitude statistics unit is used to extract the maximum gradient value within the corresponding change window for each change category cluster in the change index set, thereby obtaining the upper bound set of change amplitudes to ensure that the system can accurately define the maximum abrupt change amplitude in each type of change behavior; the time extraction unit is used to calculate the time span and statistically analyze the range based on the start and end times of each change window within each change category cluster, thereby obtaining the duration boundary set to accurately characterize the dynamic features of different types of fluctuation behavior in the time dimension; the boundary pairing unit is used to combine the upper bound set of change amplitudes and the duration boundary set according to the time order of the change windows to obtain boundary pairing data. By jointly pairing amplitude and time within the same category cluster, a multi-dimensional boundary description of energy consumption change status can be achieved; the boundary integration unit is used to establish a mapping relationship between each boundary pairing data and the category number of the change category cluster to obtain boundary combination data. The category number, maximum change amplitude, and time boundary are integrated in an orderly manner, realizing multi-dimensional modeling of the system operation status with labeling and characterization.

[0139] The boundary pairing unit is used to combine the upper bound set of change amplitude and the boundary set of duration according to the time order of the change window to obtain boundary pairing data, specifically including:

[0140] First, the maximum gradient value of the window in each category cluster is extracted from the upper bound set of variation magnitudes as the representative upper bound parameter of that category. Then, the duration range of the windows in the same category cluster is extracted from the duration boundary set as the time boundary indicator. Since a variation category cluster usually contains multiple windows, to avoid mismatch caused by sample arrangement, the system first sorts all windows in ascending order within the category cluster according to the start time of each window, thereby unifying the order of the data structure in the time dimension.

[0141] Next, based on the time-sorting results, the system sequentially binds the maximum gradient upper bound value in each category cluster to its corresponding duration boundary value, forming boundary pair data pairs. Specifically, each pair is organized as a tuple or structure, containing: category cluster number, maximum gradient upper bound value, duration range value, and time-series index. The time-series index is derived from the sequential numbering of the window's start time, ensuring that each pair of boundary data can be quickly retrieved in chronological order in subsequent modules. Through this binding operation, a set of several boundary pair structures is generated. This set is the boundary pair data, which not only reflects the amplitude and time boundary characteristics under different categories but also possesses structured time-series numbers for locating its source window in subsequent logical judgments.

[0142] The boundary integration unit is used to establish a mapping relationship between the paired boundary data and the category numbers of the change category clusters to obtain boundary combination data, specifically including:

[0143] After constructing the boundary pairing data within all category clusters, the system aggregates them into a unified global data structure. This structure uses the cluster number as the index key and the corresponding boundary pairing dataset as the index value, forming a key-value mapping relationship. Based on this, the system further integrates the boundary pairing datasets for each category number, including normalizing the maximum gradient upper bound for the same category number in each data set, calculating the average amplitude range, and performing boundary standardization on the duration range to obtain a unified time fluctuation range for each category cluster.

[0144] Subsequently, the system encapsulates each category number, its normalized maximum amplitude value, time range boundaries, and corresponding time-series index into a "boundary combination data structure." This structure includes: category number, amplitude upper bound index, time boundary index, standardization factor, and optional reference statistical indicators such as the number of windows and the proportion of extreme values. This combination structure will be stored in the boundary determination library for quick access and retrieval by the subsequent threshold interval construction and anomaly identification modules. Furthermore, this boundary combination data supports version marking and timestamp recording, enabling traceability and historical comparison of model output results across different time periods, providing a data foundation for the long-term operation of the system.

[0145] In a preferred embodiment of the present invention, the threshold module further includes:

[0146] The amplitude interval unit is used to extract the upper bound of the change amplitude of each group based on the boundary combination data, and calculate the upper and lower limit intervals of the gradient of the change category cluster by combining the average gradient value of each window within the corresponding change category cluster, thus obtaining the amplitude determination interval group.

[0147] The time interval unit is used to extract the duration boundary of each group based on the boundary combination data, and expand its minimum duration period boundary by combining the standard deviation of the time length within the corresponding change category cluster to obtain the time determination interval group.

[0148] The dual-interval fusion unit is used to pair amplitude determination interval groups and time determination interval groups according to category number to construct a two-dimensional determination interval combination set containing gradient upper and lower limits and time upper and lower limits;

[0149] The node mapping unit is used to allocate the two-dimensional judgment interval combination set to the corresponding node identifier according to the node identifier in the change index set, so as to obtain the node judgment threshold set.

[0150] In this embodiment of the invention, the amplitude interval unit is used to extract the upper bound of each group of change amplitudes based on the boundary combination data, and calculate the upper and lower bound intervals of the gradient of the change category cluster by combining the average gradient value of each window within the corresponding change category cluster, thus obtaining the amplitude determination interval group. This effectively avoids misjudging critical changes or boundary disturbances as abnormalities and provides a stable gradient reference value for subsequent construction of time intervals and dual-interval fusion. The time interval unit is used to extract the duration boundary of each group of durations based on the boundary combination data, and expand its minimum duration period boundary by combining the standard deviation of the time length within the corresponding change category cluster, thus obtaining the time determination interval group. By comprehensively considering the extreme values ​​of the time span, this method effectively avoids misjudging critical changes or boundary disturbances as abnormalities and provides a stable gradient reference value for subsequent construction of time intervals and dual-interval fusion. An adaptive time tolerance mechanism was established to address the difference and fluctuation degree. A dual-interval fusion unit was used to pair amplitude judgment interval groups and time judgment interval groups according to category number to construct a two-dimensional judgment interval combination set containing gradient upper and lower limits and time upper and lower limits. This bound the two different dimensions of judgment criteria, amplitude and time, to construct a more discriminative two-dimensional anomaly judgment model. A node mapping unit was used to assign the two-dimensional judgment interval combination set to the corresponding node identifier according to the node identifier in the change index set to obtain the node judgment threshold set. By establishing a many-to-many association between nodes and judgment intervals, it was ensured that the system could still accurately match its historical evolution trend under polymorphic behavior.

[0151] The amplitude interval unit is used to extract the upper bound of the change amplitude for each group based on the boundary combination data. Combined with the average gradient value of each window within the corresponding change category cluster, the upper and lower bounds of the gradient for that change category cluster are calculated to obtain the amplitude determination interval group, specifically including:

[0152] First, the maximum gradient value corresponding to each change category cluster is extracted from the boundary combination data as the upper bound of the change amplitude. This value is obtained by statistically analyzing the maximum parameter change rate in each window during the boundary pairing process. Then, the system traverses all change windows classified under the same category number within the category cluster, extracts the average gradient value of each window, and combines them into a gradient value set. The set is then processed for mean and standard deviation to calculate the overall average gradient level and its fluctuation degree for the category cluster. To avoid single extreme values ​​causing interval shifts, the system adopts an interval calculation strategy based on standard deviation expansion. Specifically, the lower limit of the interval is set by multiplying the average gradient value by a preset standard deviation coefficient; the upper limit of the interval is taken from the corresponding maximum gradient value in the boundary combination data, thus forming an interval structure with a fixed upper limit and an adaptive lower limit. After calculation, the upper and lower limits of the gradient for each category cluster are stored as interval groups, ultimately forming an amplitude determination interval group.

[0153] The time interval unit is used to extract the duration boundary of each group based on the boundary combination data, and expand its minimum duration period boundary by combining the standard deviation of the time length within the corresponding change category cluster, to obtain the time determination interval group, specifically including:

[0154] After receiving the boundary combination data, the time interval unit extracts the duration boundary information corresponding to each change category cluster, including the minimum and maximum duration of each category cluster, which is used to characterize the time range of the continuous operation of each change window within that category cluster. The system then obtains the duration sequence of all change windows within that category cluster and calculates the time standard deviation accordingly.

[0155] To enhance the fault tolerance of anomaly detection for weakly persistent signals at the boundary, the system expands the minimum duration value. First, it extracts the actual duration values ​​of all change windows within the category cluster and calculates the standard deviation of this time length sequence to measure the dispersion of the time distribution within the category cluster. Then, the system introduces an adjustment factor based on the original minimum duration boundary. This standard deviation is multiplied by a preset scaling factor and subtracted from the minimum duration value to obtain a downwardly expanded adjustment value. In this way, valid change windows that might have been mistakenly excluded by the system due to insufficient duration are included in the new lower limit interval, thereby enhancing the system's ability to identify actual anomaly persistence. To avoid the lower limit of the interval being less than the minimum value of the sampling period, the system automatically rounds all lower limits to an integer multiple not less than the sampling period, ensuring data interval alignment and the stability of subsequent judgments. Finally, each category cluster and its corresponding time judgment interval are grouped into a structure and uniformly summarized into a time judgment interval group.

[0156] The dual-interval fusion unit is used to pair amplitude determination interval groups and time determination interval groups according to category numbers, constructing a two-dimensional determination interval set containing gradient upper and lower limits and time upper and lower limits, specifically including:

[0157] After generating amplitude and time judgment interval groups, the system binds the two sets of intervals one-to-one according to the category number corresponding to each change category cluster. First, for each interval item in the amplitude interval group, its category number is extracted, and the time interval item with the same number is retrieved in the time interval group. To ensure the dimensional consistency of the combined data, the system pre-validates the two sets of data structures to ensure that there are valid amplitude and time interval pairs under the corresponding category number. Then, the two one-dimensional intervals are assembled into a two-dimensional judgment interval item, which is a composite structure consisting of gradient upper and lower limit intervals and duration upper and lower limit intervals. To improve the system's execution efficiency for composite judgment logic, each two-dimensional interval item is encapsulated as a structured parameter unit, including interval value, category number, and effective time window range, and stored according to category number, thereby generating a complete two-dimensional judgment interval combination set. This combination set is used as a template for node behavior judgment in subsequent identification processes and is the foundation for the system's two-dimensional screening mechanism for abnormal states.

[0158] The node mapping unit is used to allocate the two-dimensional judgment interval combination set to the corresponding node identifier according to the node identifier in the change index set, thereby obtaining the node judgment threshold set, specifically including:

[0159] First, the system iterates through each record in the change index set, extracting the node number and its matching category number, and inputting them as key-value pairs into the node mapping table. Then, based on the category number, the system extracts the corresponding decision interval combination from the two-dimensional decision interval combination set and binds this item to the mapping position corresponding to the node number, thus completing the interval allocation. During this process, if a node corresponds to multiple category numbers, the system maintains an interval queue for that node to support multiple matching judgments of different behavior patterns in subsequent recognition stages. Simultaneously, to avoid interval conflicts or redundant judgments, the system can merge or prune similar interval items to ensure that the judgment structure of each node is concise and effective. After mapping, the system stores the node and its corresponding decision interval set as a judgment template in the node threshold set data structure. This data structure serves as the basic reference for judging real-time data behavior in the recognition module, realizing personalized dynamic judgment capabilities at the node level.

[0160] In a preferred embodiment of the present invention, the identification module includes:

[0161] The normalization processing unit is used to standardize the current and voltage values ​​according to the real-time dataset of each node, so as to obtain a normalized real-time dataset.

[0162] The difference calculation unit is used to calculate the deviation between the current value and the upper bound at each time point based on the normalized real-time dataset and the node judgment threshold set, and obtain the threshold difference vector.

[0163] The over-limit monitoring unit is used to mark the time periods in the threshold difference vector that continuously exceed the upper bound of the gradient and meet the minimum duration as abnormal windows, thus obtaining a preliminary set of abnormal windows.

[0164] The anomaly integration unit is used to merge anomaly windows in the initial anomaly window set whose adjacent intervals are less than a preset recovery period, and integrate them according to node identifiers to obtain an anomaly identification result set.

[0165] In this embodiment of the invention, the normalization processing unit is used to standardize the current and voltage values ​​according to the real-time dataset of each node, so as to obtain a normalized real-time dataset, effectively eliminating the problem of inconsistent dimensions caused by differences in node physical properties, electrical appliance types, or acquisition accuracy; the difference calculation unit is used to calculate the deviation between the current value and the upper bound at each time point according to the normalized real-time dataset and the node judgment threshold set, so as to obtain the threshold difference vector, and compare the normalized electrical parameters with the node-specific dynamic upper bound threshold point by point, so as to accurately quantify the degree of energy consumption fluctuation at each moment; the over-limit monitoring unit is used to calculate the deviation between the current value and the upper bound threshold at each time point according to the threshold difference vector. The time intervals during which the quantity continuously exceeds the upper bound of the gradient and meets the minimum duration are marked as abnormal windows, resulting in a preliminary abnormal window set. By setting a dual constraint mechanism of duration and over-limit amplitude, false alarms caused by transient fluctuations and short-cycle load changes can be effectively shielded. The abnormal integration unit is used to merge abnormal windows with adjacent intervals less than the preset recovery period in the preliminary abnormal window set and integrate them according to node identifiers to obtain an abnormal identification result set. This avoids mistaking multiple abnormalities belonging to the same event as multiple independent events, providing a more structured and stable abnormal event basis for the subsequent instruction generation module, which helps to form key inputs in automated closed-loop control.

[0166] The difference calculation unit is used to calculate the deviation between the current value and the upper bound at each time point based on the normalized real-time dataset and the node decision threshold set, to obtain the threshold difference vector, specifically including:

[0167] During real-time monitoring, the system first invokes the difference calculation unit in the identification module. Its input includes two core data structures: a normalized real-time dataset and a set of judgment thresholds matching the nodes. For each node, the difference calculation unit extracts the normalized current and voltage values ​​at each time point and compares them with the upper bound of the gradient in the judgment thresholds for that node at the corresponding time point. To ensure consistency in numerical comparisons, the system guarantees that the real-time data and historical datasets use the same normalization rules. Subsequently, the system uses the difference calculation function to perform point-by-point subtraction on the current and voltage values, obtaining the deviation of the current value from the upper bound, and forming a set of time series vectors representing the difference at each time point. This calculation process not only retains the deviation magnitude at each time point but also preserves the original time index information for subsequent continuous analysis in the time domain.

[0168] The over-limit monitoring unit is used to mark the time periods in the threshold difference vector that continuously exceed the upper bound of the gradient and meet the minimum duration as abnormal windows, thus obtaining a preliminary set of abnormal windows, specifically including:

[0169] After calculating the difference vector, the system inputs the results to the over-limit monitoring unit to analyze whether each node exhibits stable and continuous upper bound breach behavior during operation. This unit, based on a preset minimum duration parameter, scans the continuous time intervals in the threshold difference vector for each node and determines whether a continuous time interval exists within that interval where the deviation is consistently positive, i.e., consistently higher than the corresponding upper limit of the judgment threshold. The system employs a sliding window method, sliding sequentially through the difference vector of each node to check for any length of... The system checks whether all time points within a continuous segment meet the out-of-limit conditions. Once a continuous time period is found to meet the above requirements, it is marked as an abnormal window. During marking, the system simultaneously records the window's start time, end time, corresponding node number, maximum deviation, and average deviation amplitude, among other characteristic indicators. To ensure system performance, this process is performed in real time, and the abnormal windows generated in each monitoring cycle are sorted and numbered chronologically. The set of all windows identified as abnormal constitutes the initial abnormal window set.

[0170] The anomaly integration unit is used to merge anomaly windows in the initial anomaly window set whose adjacent intervals are less than a preset recovery period, and integrate them according to node identifiers to obtain an anomaly identification result set, specifically including:

[0171] After initial screening of abnormal windows, the system enters the anomaly integration phase, which addresses the fragmentation issues caused by monitoring granularity or short-term interference between abnormal windows. First, the abnormal window set is grouped by node number, and within each group, windows are sorted in ascending order based on their start time. Then, the system sequentially checks the time interval between adjacent abnormal windows. If the interval is less than the system's set recovery period threshold (e.g., less than 5 seconds), the system considers the two windows to belong to the same type of abnormal event and cannot be restored to normal operation; therefore, these two windows are merged. During merging, the start time of the previous window and the end time of the next window are retained, and the average deviation value and maximum deviation amplitude of the merged window are recalculated. This operation not only eliminates erroneous judgments caused by short-term recovery but also integrates multiple consecutive out-of-limit events into a complete abnormal behavior unit, ensuring that the system can accurately grasp the complete cycle of the fault when executing control strategies. After completing the integration of abnormal windows for all nodes, the system outputs the final anomaly identification result set for each node, including key information such as the abnormal period, deviation characteristic value, and node number, and transmits this result to the instruction module, providing a trigger basis for automatic power-off or alarm response.

[0172] In a preferred embodiment of the present invention, the instruction module includes:

[0173] The node localization unit is used to find the associated node identifier in the change index set according to the index position corresponding to each abnormal window in the abnormal identification result set, and obtain the abnormal node list.

[0174] The action determination unit is used to calculate the cumulative abnormal cycle value of each node based on the number of abnormal windows and time distribution of each node in the abnormal node list, and compare it with the preset power outage threshold to obtain the determination result data.

[0175] The parameter assembly unit is used to determine the node address, power-off method and execution time of each trigger node based on the trigger node number in the judgment result data, and obtain the execution parameter data.

[0176] The instruction generation unit is used to encapsulate the execution parameter data into a recognizable control instruction structure and output it to the execution interface to complete the power-off operation of the trigger node.

[0177] In this embodiment of the invention, the node positioning unit is used to search for the associated node identifier in the change index set according to the index position corresponding to each abnormal window in the abnormal identification result set, and obtain an abnormal node list, realizing the mapping from abstract abnormal data to specific power nodes, so that the system can clearly know which physical nodes are in an abnormal state; the action judgment unit is used to calculate the abnormal cumulative cycle value of each node according to the number of abnormal windows and time distribution of each node in the abnormal node list, and compare it with the preset power outage threshold to obtain judgment result data, and quantitatively judge whether each abnormal node needs to take power outage measures; the parameter assembly unit is used to determine the node address, power outage method and execution time of each trigger node according to the trigger node number in the judgment result data, and obtain execution parameter data, converting the node identifier into equipment operation parameters that can be directly called by the control module, solving the problem of execution failure caused by inconsistent information or missing parameters in traditional power outage commands; the instruction generation unit is used to encapsulate the execution parameter data into a recognizable control instruction structure and output it to the execution interface to complete the power outage operation of the trigger node, converting the identified abnormal state into a clear field control action, and improving the fault handling response efficiency.

[0178] The node localization unit is used to search for the associated node identifier in the change index set based on the index position corresponding to each abnormal window in the abnormal identification result set, and obtain a list of abnormal nodes, specifically including:

[0179] After processing by the identification module, the system obtains an anomaly identification result set, which records information on all windows identified as anomalous. Each data entry includes a start time, end time, and the corresponding time series number. The system first receives this anomaly identification result set and extracts the time index range for each anomalous window data entry, using it as the base index value for subsequent matching. The system then retrieves the change index set previously generated by the clustering module. This index set pre-establishes a one-to-one mapping between all change windows according to their time series and node identifiers. The node localization unit compares the time index of the anomalous window with each record in the change index set, searching for change trajectory records that overlap or intersect with the time period of the anomalous window, and obtains the corresponding node identifier information accordingly.

[0180] When multiple abnormal windows match the same node, the system automatically deduplicates the extracted node identifiers to ensure that each node identifier in the abnormal node list is unique and represents an independent device abnormal event. Furthermore, to enhance the timeliness and accuracy of the results, the node localization unit can introduce a redundancy removal mechanism. That is, when a node experiences multiple abnormal windows consecutively within adjacent time periods, the system can merge them into a single persistent abnormality, thereby avoiding excessive redundant identification. The final generated abnormal node list includes: a unique node identifier, an abnormal trigger time period index, and corresponding abnormal window count information, serving as the basic data source for the action determination unit.

[0181] The action determination unit is used to calculate the cumulative abnormal period value of each node based on the number of abnormal windows and time distribution of each node in the abnormal node list, and compare it with a preset power outage threshold to obtain determination result data, specifically including:

[0182] First, each node in the list of abnormal nodes is processed one by one. The system reads all the abnormal windows corresponding to the current node and extracts the start and end times of these windows. Then, the abnormal windows are arranged in chronological order, and the time interval between adjacent abnormal windows is calculated sequentially to assess whether they constitute a continuous abnormality. If the interval between two abnormal windows is less than the system's preset minimum recovery period, then these two windows are determined to be a continuous abnormal time period and are merged into a single continuous abnormal period.

[0183] After merging all abnormal windows, the system calculates the total duration of the period from the first abnormality to the last abnormality for each node, denoted as the cumulative abnormality period value. The action determination unit then compares this period value with a preset power-off trigger threshold in the system. This threshold can be flexibly set according to the application scenario; for example, a short period can be set for high-security scenarios, while a longer period can be set for scenarios with strong fault tolerance. If the cumulative abnormality period value exceeds the power-off threshold, the system generates a power-off trigger flag for that node, indicating that the node has met the automatic power-off conditions. Finally, the action determination unit organizes the judgment results for all nodes into judgment result data, including each node's number, cumulative abnormality period value, number of abnormalities, and whether a power-off flag has been triggered, and transmits this data to the parameter assembly unit for further processing.

[0184] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A comprehensive energy consumption monitoring and management system, characterized in that, The system includes: The data acquisition module is used to collect the current and voltage values ​​of different nodes in the power distribution network, and arrange and align them in chronological order to obtain a node dataset. The feature module is used to analyze the parameter change trends of different nodes based on the node dataset, calculate the change gradient sequence and fluctuation variance of each node, and divide them into continuous change windows to obtain the change trajectory dataset. The clustering module is used to analyze the gradient distribution and fluctuation persistence of each change window based on the change trajectory dataset and perform clustering processing to divide the change trajectory dataset into multiple change category clusters and obtain a change index set; The threshold module is used to calculate the upper bound of the change magnitude and the duration boundary of each change category cluster based on the change index set, obtain boundary combination data, and construct a dynamic numerical judgment interval based on it to obtain the node judgment threshold set. The identification module is used to collect the real-time dataset of each node, compare it with the node judgment threshold set, mark the running windows that exceed the threshold, and obtain the anomaly identification result set. The instruction module is used to determine the node identifier and running window position based on the anomaly identification result set, generate operation instructions, and execute power-off. The threshold module includes: The amplitude statistics unit is used to extract the maximum gradient value within the corresponding change window for each change category cluster in the change index set, and obtain the upper bound set of change amplitude. The time extraction unit is used to calculate the time span and calculate the range based on the start and end times of each change window within each change category cluster, thereby obtaining the duration boundary set. Boundary pairing unit is used to combine the upper bound set of change amplitude and the boundary set of duration according to the time order of the change window to obtain boundary pairing data; The boundary integration unit is used to establish a mapping relationship between the pairing data of each boundary and the category number of the change category cluster to obtain the boundary combination data; The amplitude interval unit is used to extract the upper bound of the change amplitude of each group based on the boundary combination data, and calculate the upper and lower limit intervals of the gradient of the change category cluster by combining the average gradient value of each window within the corresponding change category cluster, thus obtaining the amplitude determination interval group. The time interval unit is used to extract the duration boundary of each group based on the boundary combination data, and expand its minimum duration period boundary by combining the standard deviation of the time length within the corresponding change category cluster to obtain the time determination interval group. The dual-interval fusion unit is used to pair amplitude determination interval groups and time determination interval groups according to category number to construct a two-dimensional determination interval combination set containing gradient upper and lower limits and time upper and lower limits; The node mapping unit is used to allocate the two-dimensional judgment interval combination set to the corresponding node identifier according to the node identifier in the change index set, so as to obtain the node judgment threshold set.

2. The integrated energy consumption monitoring and management system according to claim 1, characterized in that, The feature module includes: The gradient calculation unit is used to calculate the parameter difference value based on the current value and voltage value at adjacent time points in the node dataset, and obtain the initial change sequence; The normalization transformation unit is used to perform linear normalization on each parameter based on the maximum and minimum values ​​in the initial change sequence to obtain the change gradient sequence; The trend recognition unit is used to identify the time points of positive and negative changes in parameter difference values ​​based on the changing gradient sequence, and to mark the corresponding gradient direction inflection points to obtain a set of gradient inflection points. The fluctuation assessment unit is used to divide the changing gradient sequence into different sub-regions based on the gradient inflection point set, and calculate the variance value within each sub-region to obtain the fluctuation variance set.

3. The integrated energy consumption monitoring and management system according to claim 2, characterized in that, The feature module also includes: The window partitioning unit is used to divide the data of each node into multiple continuously changing windows according to the inflection point position in the gradient change sequence, so as to obtain the initial window structure. The window filtering unit is used to remove windows with low sample counts based on the minimum data length of each window in the initial window structure, thereby obtaining a set of effective windows. The window feature aggregation unit is used to obtain a set of window feature vectors based on the average gradient value and fluctuation variance value of each window in the effective window set. The trajectory stitching unit is used to stitch together window feature vector groups in chronological order, determine the time sequence number of each window feature vector, and obtain the change trajectory dataset.

4. The integrated energy consumption monitoring and management system according to claim 3, characterized in that, The clustering module includes: The gradient feature extraction unit is used to extract the average gradient value and maximum gradient fluctuation amplitude of each change window based on the time series of each change window in the change trajectory dataset, and obtain the gradient feature group. The persistent feature extraction unit is used to calculate the duration of each change window based on the start and end time points, and to count the number of gradient direction transformations within the window to obtain persistent feature groups. The feature fusion unit is used to construct a multi-dimensional vector structure for each change window based on the gradient feature set and the persistent feature set, thereby obtaining the original feature vector set; The normalization unit is used to normalize each vector based on the range of each dimension vector in the original feature vector set to obtain a standard feature vector set.

5. The integrated energy consumption monitoring and management system according to claim 4, characterized in that, The clustering module also includes: The density assessment unit is used to calculate the neighborhood density value of each vector based on the standard feature vector set, and obtain the density score sequence. The category center identification unit is used to extract the top three positions of the neighborhood density value from the density score sequence and define them as the initial category centers. The clustering partitioning unit is used to divide the standard feature vector set into multiple variable category clusters by calculating the distance between each vector and each initial category center and performing clustering processing, thereby obtaining the clustering assignment result; The index generation unit is used to label each change category cluster with a category number based on the clustering results and align it with the corresponding time series to obtain a change index set.

6. The integrated energy consumption monitoring and management system according to claim 5, characterized in that, The density assessment unit includes: The neighborhood density value calculation unit is used to extract the fluctuation variance, average gradient value, duration length and time sequence number of each vector in the change window before normalization processing, based on the standard feature vector set. Based on the standard eigenvector set vectors and their fluctuation variance values, the distance between different vectors is calculated to measure the spatial proximity between different vectors, resulting in a distance exponential decay term; based on the average gradient value of the vectors, the effect of gradient direction differences on density coupling between different vectors is calculated, resulting in a gradient similarity modulation term; based on the duration length of the vectors, a harmonic average is performed on vectors with similar duration lengths, resulting in a time-consistent weighting term; based on the time sequence number of the vectors, the influence of vectors in distant time periods on each other is controlled, resulting in a time proximity suppression term. The neighborhood density value is obtained by fusing the distance exponential decay term, gradient similarity modulation term, time consistency weighting term, and time proximity suppression term.

7. The integrated energy consumption monitoring and management system according to claim 6, characterized in that, The recognition module includes: The normalization processing unit is used to standardize the current and voltage values ​​according to the real-time dataset of each node, so as to obtain a normalized real-time dataset. The difference calculation unit is used to calculate the deviation between the current value and the upper bound at each time point based on the normalized real-time dataset and the node judgment threshold set, and obtain the threshold difference vector. The over-limit monitoring unit is used to mark the time periods in the threshold difference vector that continuously exceed the upper bound of the gradient and meet the minimum duration as abnormal windows, thus obtaining a preliminary set of abnormal windows. The anomaly integration unit is used to merge anomaly windows in the initial anomaly window set whose adjacent intervals are less than a preset recovery period, and integrate them according to node identifiers to obtain an anomaly identification result set.

8. The integrated energy consumption monitoring and management system according to claim 7, characterized in that, The instruction module includes: The node localization unit is used to find the associated node identifier in the change index set according to the index position corresponding to each abnormal window in the abnormal identification result set, and obtain the abnormal node list. The action determination unit is used to calculate the cumulative abnormal cycle value of each node based on the number of abnormal windows and time distribution of each node in the abnormal node list, and compare it with the preset power outage threshold to obtain the determination result data. The parameter assembly unit is used to determine the node address, power-off method and execution time of each trigger node based on the trigger node number in the judgment result data, and obtain the execution parameter data. The instruction generation unit is used to encapsulate the execution parameter data into a recognizable control instruction structure and output it to the execution interface to complete the power-off operation of the trigger node.

Images