Multi-dimensional edge matching load identification method based on electricity utilization acquisition and breadth search

Abstract

The invention discloses a multi-dimensional edge matching load identification method based on electricity utilization acquisition and breadth search, relates to the related field of load identification technologies, and aims to solve the core problem that the accuracy of power event detection and load state identification is seriously insufficient in the prior art. The method comprises the steps of firstly defining a load type, then dividing power slow change and sudden change by using first-order and second-order difference regions, extracting multi-dimensional edge features, designing a dynamic threshold adjustment mechanism, realizing multi-dimensional breadth-first search matching of a rising edge and a falling edge, and simultaneously extracting multi-dimensional load features to distinguish similar loads. And the load electricity consumption is planned and distributed in real time according to the breadth-first search result, and a set of full-flow technical scheme from power signal detection to load identification and energy consumption metering is formed. Accurate detection of a power event and effective identification of a full working state of a load are realized, the overall accuracy of non-intrusive load identification is improved from a core link, and a foundation is laid for subsequent load decomposition and accurate energy consumption metering.

Description

Technical Field

[0001] This application relates to the field of load identification technology, and in particular to a multidimensional edge matching load identification method based on power consumption data collection and breadth-first search. Background Technology

[0002] Non-intrusive load identification is a technology that identifies and decomposes the operating status and energy consumption of various electrical devices by analyzing electricity consumption data such as current and voltage at the user's main inlet, without installing sensors on each appliance. It has become an advanced analysis function in electricity information collection systems and is widely used in electricity consumption analysis, abnormal electricity consumption monitoring, and potential assessment of household demand response.

[0003] Existing non-invasive load identification methods can be divided into the following three categories:

[0004] The classic basic algorithm for traditional event detection is to perform time-domain analysis on the collected power sequence, set a threshold to detect power change points, extract the power change amplitude and edge features of the change points, and combine simple matching rules to achieve preliminary load decomposition and identification. However, this method has weak anti-interference ability, is easily affected by power grid noise and measurement errors, and is prone to false event detection when multiple devices are superimposed. Algorithms based on lightweight traditional machine learning: Extract steady-state shallow features of load through event detection, input the features into a lightweight machine learning model to achieve load type matching and identification. However, this method relies on shallow features extracted manually, and similar loads are prone to identification confusion. The accuracy decreases when multiple devices are superimposed. The algorithm based on lightweight deep learning performs simple feature encoding on the power sequence after event detection, inputs it into the lightweight deep learning model, and automatically extracts the time-domain features of the load, replacing manual feature extraction and realizing end-to-end real-time load identification. However, this method requires a small number of samples for training, and the model deployment requires certain hardware computing power, which cannot be adapted to pure microcontroller terminals. Summary of the Invention

[0005] To address the technical problems of the prior art, this application provides a multi-dimensional edge-matching load identification method based on power consumption data collection and breadth-based search. This method specifically solves the core technical problem of insufficient accuracy in power event detection and load status identification in existing non-intrusive load identification technologies. It overcomes the technical pain points of existing technologies, such as weak anti-interference capability in power surge detection, lack of effective identification methods for power changes during non-start-stop equipment operation, and the existence of natural blind spots in the status detection of certain types of loads. Specifically, the problems solved are as follows: To address the problem that existing power event detection methods are susceptible to interference from instantaneous grid noise and measurement errors, as well as the superposition and overlap of power changes caused by the simultaneous switching of multiple devices, which leads to false detections and missed detections of events, ultimately resulting in load decomposition failure; This addresses the problem that existing technologies focus solely on detecting equipment start-up and shutdown events, neglecting power step or smooth fluctuations caused by changes in the operating state of non-start-up and shutdown loads. This leads to misjudgments of the state (misjudging state transitions as new equipment start-up and shutdown) or a lack of state identification (treating multi-state operating equipment as a single power state). This addresses the problems of existing technologies failing to capture state changes of continuously varying loads without significant power abrupt changes, easily masking weak power signals of permanently operating low-power loads by high-power equipment, and misidentifying power changes of different components of multi-state loads as multiple independent devices, leading to inaccurate determination of equipment status and quantity. This application provides a multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search, achieving accurate detection of power events and effective identification of the full operating state of the load. It improves the overall accuracy of non-intrusive load identification from a core perspective, laying the foundation for subsequent load decomposition and accurate energy consumption metering. The method includes: Step S10: Based on the real-time acquisition of raw current and voltage data by the electricity information acquisition system, calculate the total active power, total reactive power and power factor, and after averaging and noise reduction processing, form a continuous power sequence as the input of the non-intrusive load identification algorithm; Step S20: Based on the characteristics of load power change, the electrical load is divided into three basic models in advance. For different types of electrical load, the first and second order differences are calculated on the continuous power sequence to distinguish between slow power change and power change, while filtering out false power change signals caused by noise. Step S30: For the detected power mutation point, extract the multi-dimensional features of its corresponding edge, including time dimension, power dimension and differential dimension, construct a normalized feature vector, and dynamically adjust the matching tolerance threshold of the multi-dimensional edge features according to the real-time operating conditions of the power grid, the degree of load superposition and the load type. Step S40: Taking multi-dimensional edge features as the core, a breadth-first search strategy is adopted. According to the principle of the highest similarity and the smallest matching cluster, the power rising and falling edges are matched, and the steady-state and dynamic features of the load are extracted from the successfully matched load operation segments to build a load feature library and distinguish similar loads. Step S50: Combining the results of load type definition, power mutation detection, edge matching and load feature identification, identify the full working state of the load, and split the total power sequence according to the time dimension based on the identification results, and plan and allocate the power consumption of various types of loads in real time. Step S60: Integrate the processing results of all the above steps and output visualized, standardized, non-intrusive load identification and energy consumption metering results, including basic information of each load, full working status information, energy consumption metering information, and power decomposition curves.

[0006] Furthermore, during the electricity consumption information collection process, the total active power, total reactive power, and power factor at the periodic level are converted into a continuous sequence at the second level through averaging. At the same time, the sliding window filtering method is used to reduce noise in the sequence and filter out small instantaneous noise from the power grid. The noise-reduced second-level total active power sequence, total reactive power sequence, and power factor sequence are used as the core input data for the subsequent non-intrusive load identification algorithm to ensure the continuity and effectiveness of the data.

[0007] Furthermore, based on the power change characteristics of the load, the electrical load is divided into three basic types in advance, and the power change characteristics of each type of load are clarified, laying the foundation for subsequent targeted detection and identification. The three types of loads include single-state loads, multi-state loads, and continuously changing loads.

[0008] Furthermore, the first-order difference and the second-order difference are calculated for the total active power sequence and the total reactive power sequence at the second level, respectively, to obtain the first-order difference sequence of active power and reactive power, and the second-order difference sequence of active power and reactive power. Set a basic judgment threshold for differential judgment, and judge power mutation through second-order differential judgment. When the absolute value of second-order differential exceeds the basic judgment threshold, it is judged as a power mutation, which corresponds to the start and stop of single-state load, and the gear or mode switching event of multi-state load, providing the core basis for edge detection. The power variation is determined by the first-order difference. When the second-order difference does not exceed the basic judgment threshold, but the first-order difference shows a continuous and smooth non-zero change, it is judged as a power variation, which corresponds to the power adjustment process of continuously changing load. By introducing reactive power differential characteristics for cross-validation, only when the slowly changing or abruptly changing signals detected by active power differential are consistent with the trend of reactive power differential characteristics are they determined to be effective power changes. This filters out spurious power change signals caused solely by grid noise and measurement errors, thus solving the problem of false detections caused by noise.

[0009] Furthermore, for effective power mutation points, multi-dimensional features corresponding to the rising and falling edges are extracted, breaking through the traditional limitation of extracting only time and power amplitude features. This provides feature support for subsequent accurate matching and similar load differentiation. Specifically, the extracted edge features consist of time dimension, power dimension, and difference dimension. The multi-dimensional edge features are normalized to construct multi-dimensional feature vectors for each rising and falling edge, which serve as the core basis for subsequent breadth-first search matching.

[0010] Based on the real-time operating conditions of the power grid, the degree of load superposition, and the load type, the matching tolerance threshold of multi-dimensional edge characteristics is dynamically adjusted to improve the edge matching accuracy under different operating conditions. For the switching of load levels and modes in multi-state loads, a matching threshold range for the power change amplitude is set separately to adapt to the characteristic law of its step change and solve the problem of missing identification of non-start-stop state transitions.

[0011] Furthermore, based on the dynamically adjusted matching tolerance threshold, a multi-dimensional breadth-first search matching algorithm for rising and falling edges is designed to achieve accurate matching of power change edges, while solving the problems of edge overlap and matching failure caused by multiple devices stacking: The core rules of multi-dimensional breadth-first search matching are set as time logic rules and feature similarity rules. Taking multi-dimensional edge feature vectors as the core, the feature comprehensive similarity between all candidate rising edges and falling edges is calculated (using the calculation method of Euclidean distance and cosine similarity). The cost function of edge matching is constructed, and the breadth-first search strategy is used to solve the cost function. For edge overlap signals caused by multiple devices, decoupling analysis of multi-dimensional edge feature vectors is used to separate the edge features of different loads, achieve accurate matching of rising and falling edges of a single load, and solve the problem of load decomposition failure caused by edge overlap. After matching is completed, each valid matching cluster is output, and each matching cluster corresponds to the power change event of an independent load.

[0012] Furthermore, for each successfully matched edge cluster corresponding to the load operating segment, multi-dimensional steady-state and dynamic load features are extracted to construct a complete load feature library. Through comprehensive matching of multi-dimensional load features, the problem of confusion in identifying similar loads in existing technologies is solved. A standardized load template library is pre-built, which stores the benchmark values ​​and tolerance ranges of the above-mentioned multi-dimensional load characteristics of various common electrical equipment. The extracted multi-dimensional load characteristics to be identified are compared with the feature benchmark values ​​in the load template library to calculate the comprehensive similarity. The equipment type corresponding to the template with the highest similarity is selected as the identification result. Through cross-validation of multi-dimensional load characteristics, the accurate differentiation of similar loads is achieved.

[0013] Furthermore, by combining the results of load type definition, power change detection, edge matching, and load characteristic identification, the full operating states of different types of loads can be identified, overcoming the problems of existing technologies that only identify start-stop, ignore non-start-stop state transitions, and miss specific loads. Based on the edge multi-dimensional breadth-first search results and the full load working status identification results, the total power sequence is segmented by time dimension, and the power consumption of each load is planned and allocated in real time, solving the problems of power consumption calculation errors and allocation distortion in existing technologies; the power consumption calculation results of various loads are recorded in real time to ensure no duplication and no omission, and to achieve accurate metering of power consumption during single load and multi-load superposition periods.

[0014] This application discloses the following technical effects: This application proposes a multi-dimensional edge-matching load identification method based on power consumption data collection and breadth-first search. This method can be widely applied to load monitoring, energy consumption metering, and power safety management in various power consumption scenarios such as homes, residential communities, and commercial buildings. Compared with existing technologies, it has significant technical advantages and application value. Specific effects and advantages are as follows: 1. The anti-interference capability of power event detection has been greatly improved, and the false detection and false negative rates have been significantly reduced: By accurately distinguishing between gradual and abrupt power changes using first- and second-order differential methods, and introducing a cross-validation mechanism for differential features, the system effectively filters out spurious power change signals caused by instantaneous grid noise and measurement errors. Furthermore, addressing the edge overlap issue caused by multiple devices, the system achieves accurate decomposition of single load edge features through decoupling analysis of multi-dimensional feature vectors and breadth-first search matching. This completely solves the problems of false event detection, missed detection, and load decomposition failure in traditional methods, laying a precise signal foundation for subsequent load identification. 2. High accuracy in recognition under full workload, overcoming various recognition blind spots: The loads are pre-classified into three categories based on their power characteristics, and differentiated identification strategies are designed for different types of loads. This not only accurately identifies the start-stop state of single-state loads, but also effectively captures the non-start-stop state transitions of multi-state loads, as well as the power smoothing process of continuously changing loads. At the same time, by anchoring long-term stable power loads through multi-dimensional features, the method completely solves the problem of missed detection of this type of load in existing methods, and achieves blind-spot-free accurate identification of all load operating states and types. 3. Multi-dimensional feature matching enables accurate differentiation of similar loads, significantly improving uniqueness: Breaking through the limitations of traditional methods that rely solely on a single power feature, this method extracts both steady-state and dynamic features of the load and constructs a standardized multi-dimensional load template library for comprehensive similarity matching. Through multi-feature cross-validation, it effectively distinguishes similar loads, completely solving the problem of confusion in similar load identification in traditional methods and significantly improving the uniqueness and accuracy of load identification. 4. Accurate and real-time metering of load power consumption, resolving metering distortion issues: Based on the multi-dimensional breadth-first search results and the full load status identification results, the total power sequence is segmented according to the time dimension, realizing the real-time planning and allocation of load power consumption. This fundamentally solves the problems of energy consumption allocation errors, metering deviations caused by state transitions, and error accumulation in traditional methods. The load power consumption metering results are highly consistent with the actual power consumption, fully meeting the actual requirements of power consumption statistics and metering. 5. Strong engineering adaptability and scalability: The system can collect basic current and voltage data at the main incoming line using only a conventional electricity information collection system, and then combine this with a standardized algorithm program to achieve full-process identification. Both the data processing and algorithm inference processes are modularly designed and can be directly deployed on conventional hardware to achieve real-time inference at the edge without the need for high-performance cloud servers. At the same time, the algorithm's operation steps are standardized, eliminating the need for complex debugging by professional technicians. It is easy to implement and promote, and adaptable to the actual application needs of various residential and commercial electricity scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the multidimensional edge-matching load identification method based on power consumption data collection and breadth-first search provided in an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] This application provides a multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search, such as... Figure 1 As shown, the method includes: Step S10: Based on the real-time current and voltage raw data from the electricity information acquisition system, calculate the total active power, total reactive power, and power factor. After averaging and noise reduction, form a continuous power sequence as the input to the non-intrusive load identification algorithm.

[0018] In this embodiment, the raw current and voltage data of the main power line are collected for each sampling period, and the sampling timestamp is recorded simultaneously: Install an electricity information collection terminal on the user's main power line (behind the household electricity meter) to collect the instantaneous voltage value sequence in real time. and instantaneous current value sequence To accurately capture load change characteristics, the sampling frequency should be no less than 6400 points per second (i.e., 128 points per power frequency cycle) to meet the accuracy requirements of subsequent power calculation and event detection. The total active power and total reactive power for each cycle are calculated using the collected current and voltage data, and the real-time power factor is derived. For the instantaneous voltage and current values ​​within each power frequency cycle, the total active power and total reactive power of that cycle are calculated using the discrete integration method. The calculation formula is as follows:

[0019] in, This indicates total active power or total reactive power. This represents the number of sampling points within one power frequency cycle. and These represent the sampling points. The corresponding instantaneous values ​​of voltage and current; Based on the calculated total active power and total reactive power, calculate the apparent power for that cycle, and use the ratio of total active power to apparent power as the power factor for that cycle. The total active power, total reactive power, and power factor at the periodic level are converted into a continuous sequence at the second level through averaging. The sequence is then denoised using a sliding window filtering method to filter out minute instantaneous noise from the power grid. The noise-reduced second-level total active power sequence, total reactive power sequence, and power factor sequence are used as the core input data for the subsequent non-intrusive load identification algorithm to ensure the continuity and effectiveness of the data.

[0020] Step S20: Based on the characteristics of load power change, the electrical load is divided into three basic models in advance. For different types of electrical load, the first and second order differences are calculated on the continuous power sequence to distinguish between slow power change and power change, while filtering out false power change signals caused by noise.

[0021] In this embodiment, the electrical loads are pre-classified into three basic types based on their power variation characteristics, clarifying the power variation characteristics of each type of load. This lays the foundation for subsequent targeted detection and identification. The three types of electrical loads are defined as follows: Single-state load: There are only two working states, on and off. During operation, the active and reactive power remain constant and there is no intermediate state, such as ordinary water dispensers and routers. Multi-state load: It has multiple discrete working states, multiple working levels and modes. When the working state is switched, the power consumption changes abruptly, such as when the hair dryer changes speed, the air conditioner switches operating modes, and the electric oven adjusts its speed. Continuously varying load: Power can be continuously and smoothly adjusted according to work requirements, without obvious discrete states, and the power exhibits a gradual variation characteristic, such as the temperature adjustment of variable frequency air conditioners and the brightness adjustment of dimming lights.

[0022] Calculate the first-order and second-order differences for the total active power sequence and the total reactive power sequence at the second level, respectively, to obtain the first-order difference sequences of active and reactive power, and the second-order difference sequences of active and reactive power: The first-order difference represents the change in power between adjacent time points, used to capture the trend and rate of change. The second-order difference calculates the change based on the first-order difference, used to accurately locate the inflection points and edges of the power sequence. By analyzing historical data and power sequences during periods of no load variation in the power system, the distribution of the second difference of the total active power sequence is statistically analyzed, and a threshold is determined based on three times the standard deviation of the second difference of the total active power sequence. Power mutations are determined by second-order difference. When the absolute value of the second-order difference exceeds the basic judgment threshold, it is judged as a power mutation, which corresponds to the start-up and shutdown of single-state loads and the gear or mode switching events of multi-state loads, providing the core basis for edge detection. The power variation is determined by the first-order difference. When the second-order difference does not exceed the basic judgment threshold, but the first-order difference shows a continuous and smooth non-zero change, it is judged as a power variation, which corresponds to the power adjustment process of continuously changing load. Reactive power differential characteristics are introduced for cross-validation. Only when the slowly varying or abruptly changing signals detected by active power differential characteristics are consistent with the trend of reactive power differential characteristics are they determined to be effective power changes. Set a time tolerance window to check whether the peak values ​​of the second-order difference sequences of active and reactive power are within the window and have the same sign at each time point. This filters out spurious power change signals caused only by grid noise and measurement errors, thus solving the problem of false detections caused by noise.

[0023] Step S30: For the detected power mutation point, extract the multi-dimensional features of its corresponding edge, including time dimension, power dimension and differential dimension, and construct a normalized feature vector. Then, dynamically adjust the matching tolerance threshold of the multi-dimensional edge features according to the real-time operating conditions of the power grid, the degree of load superposition and the load type.

[0024] In this embodiment, for the detected effective power abrupt change point, a time window is extended forward and backward from this point as the center. This window is defined as the edge interval to be analyzed, and multi-dimensional features corresponding to the rising edge and falling edge are extracted from it. These multi-dimensional edge features include: Time dimension: edge start time, edge end time, edge duration; Power dimensions: active power change range, reactive power change range, power factor change value; Difference dimensions: first-order difference peak, second-order difference peak, difference rate of change; The above multi-dimensional edge features are normalized to construct multi-dimensional feature vectors for each rising edge and falling edge, which serve as the core basis for subsequent breadth-first search matching. Secondly, based on the real-time operating conditions of the power grid, the degree of load superposition, and the load type, the matching tolerance threshold of multi-dimensional edge characteristics is dynamically adjusted to improve the edge matching accuracy under different operating conditions. The detailed steps are as follows: The criteria for setting dynamic threshold adjustment are: grid noise intensity (calculated from the fluctuation variance of the differential sequence), load superposition degree (determined by the number and density of power mutation points), and load type (single-state, multi-state, and continuously changing loads). The fixed similarity matching threshold is replaced with a dynamic threshold, the value of which is determined by the above-mentioned judgment index and is expressed as a weighted sum of the basic similarity threshold, power grid noise intensity, load superposition degree and load type adjustment item; When the grid noise intensity exceeds the dynamic threshold or multiple devices overlap, causing power edge overlap, the matching tolerance thresholds for the active and reactive power variation amplitude and edge duration characteristics are expanded, while the thresholds for power factor change value and differential change rate remain unchanged to avoid missing matching due to complex operating conditions. When the grid noise intensity is less than the dynamic threshold and there is no load superposition, the matching tolerance threshold of all features is reduced to improve the accuracy of edge matching and avoid false matching. For multi-state loads switching between gears and modes, a matching threshold range for the power change amplitude is set separately to adapt to the characteristic law of its step change and solve the problem of missing identification of non-start-stop state transitions.

[0025] Step S40: Using multi-dimensional edge features as the core, a breadth-first search strategy is adopted to match power rising and falling edges according to the principle of highest similarity and smallest matching cluster. The steady-state and dynamic features of the load are extracted from the successfully matched load operation segments to build a load feature library and distinguish similar loads.

[0026] In this embodiment, based on the dynamically adjusted matching tolerance threshold, a multi-dimensional breadth-first search matching algorithm for rising and falling edges is designed to achieve accurate matching of power change edges, while solving the problems of edge overlap and matching failure caused by multiple devices superimposed. The detailed steps are as follows: Define the core rules for multi-dimensional breadth-first search matching: Timing logic rule: The timing of the matching falling edge must be after the corresponding rising edge; Feature similarity rule: The similarity of the multi-dimensional feature vectors of the edge pairs to be matched must be higher than the similarity threshold set by the dynamic threshold; Using multi-dimensional edge feature vectors as the core, the feature comprehensive similarity of all candidate rising and falling edges is calculated. The cost function for edge matching is constructed by using a calculation method that combines Euclidean distance and cosine similarity. A breadth-first search strategy is adopted to solve the cost function: edge pairs with the highest feature similarity are matched first to minimize the number of edges in the constructed matching cluster. At the same time, intermediate calculation results are recorded and reused by the algorithm to avoid repeated matching attempts and improve matching efficiency. For edge overlap signals caused by multiple devices, decoupling analysis of multi-dimensional edge feature vectors is used to separate the edge features of different loads, achieve accurate matching of rising and falling edges of a single load, and solve the problem of load decomposition failure caused by edge overlap. After matching is completed, each valid matching cluster is output, and each matching cluster corresponds to an independent load power change event, i.e., start-up, shutdown, or state transition event. For the load operating segment corresponding to each successfully matched edge cluster, multi-dimensional load steady-state features and dynamic features are extracted to construct a complete load feature library. Through comprehensive matching of multi-dimensional load features, the problem of confusion in the identification of similar loads in existing technologies is solved, including the following detailed steps: Extract the steady-state characteristics of the load: average active power, average reactive power, rated power range, stable power factor, and equipment operating efficiency; Extract load dynamic characteristics: power change rate, state transition duration, power maintenance time of each operating state, and power gradual adjustment range (for continuously changing loads). A standardized load template library is pre-built, which stores the above-mentioned multi-dimensional edge characteristic benchmark values ​​and tolerance ranges of various common electrical equipment (including similar loads); The extracted multi-dimensional load features to be identified are compared with the feature benchmark values ​​in the load template library to calculate the comprehensive similarity. The device type corresponding to the template with the highest similarity is selected as the identification result. Through cross-validation of multi-dimensional load features, the accurate differentiation of similar loads is achieved.

[0027] Step S50: Combining the results of load type definition, power mutation detection, edge matching and load feature identification, identify the full working state of the load, and split the total power sequence according to the time dimension based on the identification results, and plan and allocate the power consumption of various types of loads in real time.

[0028] In this embodiment, by combining load type definition, power surge detection, edge matching, and load characteristic identification results, the full operating states of different types of loads are identified, overcoming the problems of existing technologies that only identify start-stop, ignore non-start-stop state transitions, and miss specific loads. For single-state loads, identify the start / stop status of their switches and the corresponding time points; For multi-state loads, it not only identifies their start-stop status, but also identifies their gear switching, mode adjustment and non-start-stop working state transition during operation, and records the switching time and corresponding power value of each state. For continuously changing loads, the system tracks the gradual change in power, identifies the start and end times of power regulation and the range of power change, and achieves full-process identification of the gradually changing power state. For loads that operate stably for a long time, feature anchoring is performed based on their unique characteristics of constant low power and stable power factor, so that their signals are separated from the power changes of the equipment and missed detection is avoided. Secondly, based on the edge-multidimensional breadth-first search results and the full-state load identification results, the total power sequence is segmented according to the time dimension, and the power consumption of each load is planned and allocated in real time, solving the problems of power consumption calculation errors and allocation distortion in existing technologies: For single-state loads, their power consumption is accurately calculated by multiplying the operating time of the load in the on state by the corresponding constant active power. For multi-state loads, the power consumption of each state is calculated by multiplying the running time of each state by the active power of the corresponding state, and then summed to obtain the total power consumption of the load. For continuously varying loads, the active power sequence during the power slow change process is integrated over time using the integral method to calculate the actual electricity consumption during the adjustment process. For multi-device overlay operation periods, based on the power decomposition results matched by multi-dimensional breadth-first search, the total power consumption is accurately divided and allocated in real time according to the real-time power contribution ratio of each load. It records the electricity consumption calculation results of various loads in real time to ensure no duplication or omission, and achieves accurate metering of electricity consumption during periods of single load and multiple load overlap.

[0029] Step S60: Integrate the processing results of all the above steps and output visualized, standardized, non-intrusive load identification and energy consumption metering results, including basic information of each load, full working status information, energy consumption metering information, and power decomposition curves.

[0030] In this embodiment, the processing results of all the above steps are integrated to output visualized, standardized, non-intrusive load identification and energy consumption metering results. The specific output content includes: Basic information for each load: load type, identification confidence level; Full operating status information for each load: start / stop time, gear and mode switching time, and power variation range; Energy consumption metering information for each load: electricity consumption in each operating state, total electricity consumption, and electricity consumption per unit time; Power decomposition curves: total power change curve, power change curves of each independent load, and power change curve of continuously changing loads.

[0031] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search, characterized in that, The method includes: Step S10: Based on the real-time acquisition of raw current and voltage data by the electricity information acquisition system, calculate the total active power, total reactive power and power factor, and after averaging and noise reduction processing, form a continuous power sequence as the input of the non-intrusive load identification algorithm; Step S20: Based on the characteristics of load power change, the electrical load is divided into three basic models in advance. For different types of electrical load, the first and second order differences are calculated on the continuous power sequence to distinguish between slow power change and power change, while filtering out false power change signals caused by noise. Step S30: For the detected power mutation point, extract the multi-dimensional features of its corresponding edge, including time dimension, power dimension and differential dimension, construct a normalized feature vector, and dynamically adjust the matching tolerance threshold of the multi-dimensional edge features according to the real-time operating conditions of the power grid, the degree of load superposition and the load type. Step S40: Taking multi-dimensional edge features as the core, a breadth-first search strategy is adopted. According to the principle of the highest similarity and the smallest matching cluster, the power rising and falling edges are matched, and the steady-state and dynamic features of the load are extracted from the successfully matched load operation segments to build a load feature library and distinguish similar loads. Step S50: Combining the results of load type definition, power mutation detection, edge matching and load feature identification, identify the full working state of the load, and split the total power sequence according to the time dimension based on the identification results, and plan and allocate the power consumption of various types of loads in real time. Step S60: Integrate the processing results of all the above steps and output visualized, standardized, non-intrusive load identification and energy consumption metering results, including basic information of each load, full working status information, energy consumption metering information, and power decomposition curves.

2. The multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S20, the three types of electrical loads are defined as follows: Single-state load: It has only two working states, on and off. During operation, the active and reactive power remain constant, and there is no intermediate state. Multi-state load: It has multiple discrete working states, multiple working levels and modes, and the power consumption changes abruptly when the working state is switched. Continuously varying load: Power is continuously and smoothly adjusted according to work requirements, without discrete states, and the power exhibits a gradual variation characteristic.

3. The multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S20, the first-order difference and the second-order difference are calculated for the total active power sequence and the total reactive power sequence at the second level, respectively: The first-order difference represents the change in power between adjacent time points and is used to capture the trend and rate of change. The second-order difference calculates the change based on the first-order difference and is used to accurately locate the inflection points and edges of the power sequence. By analyzing historical data and power sequences during periods of no load variation in the power system, the distribution of the second difference of the total active power sequence is statistically analyzed, and a threshold is determined based on three times the standard deviation of the second difference of the total active power sequence. Power mutations are determined by second-order difference. When the absolute value of the second-order difference exceeds the basic judgment threshold, it is judged as a power mutation, corresponding to the start-up and shutdown of single-state loads and the switching events of gears or modes of multi-state loads. The power variation is determined by the first-order difference. When the second-order difference does not exceed the basic judgment threshold, but the first-order difference shows a continuous and smooth non-zero change, it is judged as a power variation, which corresponds to the power adjustment process of continuously changing load. Reactive power differential characteristics are introduced for cross-validation. Only when the slowly changing or abruptly changing signals detected by active power differential characteristics are consistent with the trend of reactive power differential characteristics are they determined to be effective power changes.

4. The multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S30, for the detected effective power abrupt change point, a time window is extended forward and backward centered on that point. This window is defined as the edge interval to be analyzed, and multi-dimensional features corresponding to the rising and falling edges are extracted from it. These multi-dimensional edge features include: Time dimension: edge start time, edge end time, edge duration; Power dimensions: active power change range, reactive power change range, power factor change value; Difference dimensions: first-order difference peak, second-order difference peak, difference rate of change; The above multi-dimensional edge features are normalized to construct multi-dimensional feature vectors for each rising and falling edge, which serve as the core basis for subsequent breadth-first search matching.

5. The multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S30, the matching tolerance threshold of multi-dimensional edge characteristics is dynamically adjusted according to the real-time operating conditions of the power grid, the degree of load superposition, and the load type. The detailed steps are as follows: The criteria for setting dynamic threshold adjustment include the power grid noise intensity calculated from the fluctuation variance of the differential sequence, the degree of load superposition and the type of power load determined by the number and density of power mutation points; The fixed similarity matching threshold is replaced with a dynamic threshold, the value of which is determined by the above-mentioned judgment index and is expressed as a weighted sum of the basic similarity threshold, power grid noise intensity, load superposition degree and load type adjustment item; When the grid noise intensity exceeds the dynamic threshold or multiple devices overlap, causing power edge overlap, the matching tolerance thresholds for the active and reactive power variation amplitude and edge duration characteristics are expanded, while the thresholds for power factor change value and differential change rate remain unchanged to avoid missing matching due to complex operating conditions. When the grid noise intensity is less than the dynamic threshold and there is no load superposition, reduce the matching tolerance threshold of all features to avoid false matches; For multi-state loads switching between different load levels and modes, a matching threshold range for the power change amplitude is set separately to adapt to the characteristic pattern of its step change.

6. The multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S40, based on the dynamically adjusted matching tolerance threshold, a multi-dimensional breadth-first search matching algorithm for rising and falling edges is used to achieve accurate matching of power change edges, including the following detailed steps: Define the core rules for multi-dimensional breadth-first search matching: Timing logic rule: The timing of the matching falling edge must be after the corresponding rising edge; Feature similarity rule: The similarity of the multi-dimensional feature vectors of the edge pairs to be matched must be higher than the similarity threshold set by the dynamic threshold; Using multi-dimensional edge feature vectors as the core, the feature comprehensive similarity of all candidate rising and falling edges is calculated. The cost function for edge matching is constructed by using a calculation method that combines Euclidean distance and cosine similarity. A breadth-first search strategy is adopted to solve the cost function: edge pairs with the highest feature similarity are matched first to minimize the number of edges in the constructed matching cluster. At the same time, intermediate calculation results are recorded and reused by the algorithm to avoid repeated matching attempts.

7. The multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S40, for each successfully matched edge cluster corresponding to the load operating segment, multi-dimensional load steady-state features and dynamic features are extracted to construct a complete load feature library, achieving comprehensive matching of multi-dimensional load features. This includes the following detailed steps: Extract the steady-state characteristics of the load: average active power, average reactive power, rated power range, stable power factor, and equipment operating efficiency; Extract load dynamic characteristics: power change rate, state transition duration, power maintenance time of each operating state, and power gradual adjustment range; A standardized load template library is pre-built, which stores the benchmark values ​​and tolerance ranges of the above-mentioned multi-dimensional load characteristics of various common electrical equipment. The extracted multi-dimensional load characteristics to be identified are compared with the feature benchmark values ​​in the load template library to calculate the comprehensive similarity. The equipment type corresponding to the template with the highest similarity is selected as the identification result. Through cross-validation of multi-dimensional load characteristics, the accurate differentiation of similar loads is achieved.

8. The multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S50, by combining the results of load type definition, power surge detection, edge matching, and load characteristic identification, the full operating states of different types of loads are identified: For single-state loads, identify the start / stop status of their switches and the corresponding time points; For multi-state loads, it not only identifies their start-stop status, but also identifies their gear switching, mode adjustment and non-start-stop working state transition during operation, and records the switching time and corresponding power value of each state. For continuously changing loads, the system tracks the gradual change in power, identifies the start and end times of power regulation and the range of power change, and achieves full-process identification of the gradually changing power state. For loads that operate stably for a long period of time, feature anchoring is performed based on their unique characteristics of constant low power and stable power factor, so that their signals are separated from the power changes of the equipment and missed detections are avoided.

9. The multi-dimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S50, based on the edge multi-dimensional breadth-first search results and the load full operating state identification results, the total power sequence is segmented according to the time dimension, and the power consumption of each load is planned and allocated in real time: For single-state loads, their power consumption is accurately calculated by multiplying the operating time of the load in the on state by the corresponding constant active power. For multi-state loads, the power consumption of each state is calculated by multiplying the running time of each state by the active power of the corresponding state, and then summed to obtain the total power consumption of the load. For continuously varying loads, the active power sequence during the power slow change process is integrated over time using the integral method to calculate the actual electricity consumption during the adjustment process. For multi-device overlay operation periods, based on the power decomposition results matched by a multi-dimensional breadth-first search, the total power consumption is accurately divided and allocated in real time according to the real-time power contribution ratio of each load.

10. The multidimensional edge-matching load identification method based on power consumption data acquisition and breadth-first search as described in claim 1, characterized in that, In step S60, the processing results of all the above steps are integrated to output visualized, standardized, non-intrusive load identification and energy consumption metering results. The specific output content includes: Basic information for each load: load type, identification confidence level; Full operating status information for each load: start / stop time, gear and mode switching time, and power variation range; Energy consumption metering information for each load: electricity consumption in each operating state, total electricity consumption, and electricity consumption per unit time; Power decomposition curves: total power change curve, power change curves of each independent load, and power change curve of continuously changing loads.

Images