An Adaptive Heuristic Grid Clustering Method for Non-Intrusive Load Monitoring

Abstract

The present invention discloses an adaptive heuristic grid clustering method suitable for non-intrusive load monitoring, including the following steps: S1, selecting low-frequency data and preprocessing, converting the selected data into a four-dimensional tensor containing total load, hour, week and month; S2, using a time window to perform event detection on the total load measurement sequence, generating a four-dimensional matrix containing event amplitude and time reference, and using a heuristic grid clustering algorithm to group the events, and different clustering results represent different states of the equipment during operation; S3, constructing a state transfer matrix according to the event grouping results and the state changes of each device during operation; reconstructing the energy consumption sequence of the equipment according to the energy consumption cycle of the equipment and the state transfer matrix, completing the estimation of the energy consumption of the equipment, and calculating the frequency of use of the equipment. The present invention has high accuracy and efficiency when processing household data, and can effectively separate the power loads of various types of electrical appliances.

Description

Technical Field

[0001] The present invention relates to the technical fields of smart grid and energy management, and particularly to an adaptive heuristic grid clustering method applicable to non-intrusive load monitoring. Background Art

[0002] Non-intrusive load monitoring (NILM), as an important research direction in the current power field, is committed to accurately inferring the energy consumption of each electrical appliance based on the analysis of the main channel current. The accuracy of this technology has a profound impact on practical applications. In the power industry, accurate energy consumption inference can help users understand the contribution of each device to the total load, thereby optimizing the use of electrical appliances and effectively reducing energy waste.

[0003] In current practical research, various solutions have been proposed for the NILM problem. Researchers mainly use probability models or deterministic models (such as finite state machines) to explain the behavior of electrical appliances, and some also adopt heuristic-based technologies. In modeling energy usage, hidden Markov models (HMMs) are becoming increasingly common, and combined with supervised and unsupervised procedures and signal processing techniques to achieve the decomposition of electrical appliance characteristics. At the same time, the combination of fuzzy clustering and soft computing methods, and the use of specific algorithms for energy decomposition are also widely used. In addition, related research also involves the discussion of evaluation indicators and the hybrid device recognition scheme based on time series characteristics, etc.

[0004] However, these existing methods have certain limitations in practical applications. For example, the task complexity of HMMs increases significantly with the increase in the number of electrical appliances, the number of state combinations, and the number of sequences; some methods have limited adaptability to device types and are difficult to meet the accurate monitoring requirements of different types of devices; moreover, in dealing with complex situations such as measurement errors and multi-event concurrency, the accuracy and stability of inference need to be further improved. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide an adaptive heuristic grid clustering method applicable to non-intrusive load monitoring, aiming to identify and decompose the energy consumption of each electrical appliance by analyzing the total current signal in a household or commercial building, so as to achieve accurate monitoring of the electrical appliance usage and optimized management of energy consumption.

[0006] Technical Solution: An adaptive heuristic grid clustering method applicable to non-intrusive load monitoring includes the following steps:

[0007] S1, select the low-frequency data in the RAE and REDD datasets and perform preprocessing, and convert the selected data into a four-dimensional tensor including total load, hour, week, and month;

[0008] S2. Detect events for the total load measurement sequence using a time window, generate a four-dimensional matrix containing event amplitudes and time references, and group the events using a heuristic grid clustering algorithm. Different clustering results represent different states of the device during operation;

[0009] S3. Construct a state transition matrix based on the event grouping results and the state changes of each device during operation; adjust the power amplitude transfer within the period according to the zero loop and constraints to make the sum of the transfer amplitudes approach zero; reconstruct the energy consumption sequence of the device based on the energy consumption period and state transition matrix of the device, complete the estimation of the device energy consumption, and deduce the usage frequency of the device.

[0010] Furthermore, the preprocessing of the selected data is as follows: Extract the time attributes from the time series and transform them into a four-dimensional tensor containing the total load F(t), hour H(t), week W(t), and month M(t). Unify the data scale through a sine function and normalization processing. The expression is:

[0011] ,

[0012] where , , represent the total number of hours in a day, the total number of days in a week, and the total number of months in a year respectively; represents the total load, represents the hourly load, represents the weekly load, represents the monthly load.

[0013] Furthermore, in step S2, if any of the following conditions is met within the time window, it is determined as an event: The local change of the total load is greater than or equal to the threshold ; or the absolute value of the power difference between the input and output of the time window is greater than or equal to the threshold ;

[0014] ,

[0015] where represents the average power within the time window, represents the power measurement value at the i-th time point within the time window, and n is the number of measurement points within the time window; represents the n-th time measurement point;

[0016] When a certain electrical appliance is suddenly turned on, the power will fluctuate greatly. At this time, the value of will increase accordingly. At this time, it is necessary to calculate the local average power as the transitional amplitude:

[0017] ,

[0018] When , register as an event;

[0019] The absolute value of the power difference between the input and output of the time window , reflecting the change range of the power within the window, and the calculation formula is as follows:

[0020] ,

[0021] wherein, and are the power values at the start and end times of the time window respectively;

[0022] When , register as an event.

[0023] Furthermore, in step S2, the implementation process of grouping events by using the heuristic grid clustering algorithm is as follows:

[0024] S21, Initialize a grid structure containing N bins in the k-dimensional space, and fill the detected event data into the grid structure; Define the grid space filled with data as a block, then calculate the density and center point of each block, and divide the blocks into main blocks and ordinary blocks according to the density, and identify the main blocks, where the main blocks have high density and their directly adjacent blocks are not selected as main blocks;

[0025] S22, Use the middle range as the dissimilarity index to group events with similar characteristics into the same class; The calculation formula for the middle range is:

[0026] ,

[0027] wherein, Mid-Range represents the middle range, which is used to evaluate the difference degree of event characteristics. By comparing the middle range values of different event blocks, the similarity between events is judged; u represents an event in the block in the grid structure, represents the maximum value of the data contained in event u in the block in the grid structure, represents the minimum value of the data contained in event u in the block in the grid structure;

[0028] S23, Merge the blocks;

[0029] For example , then merge the ordinary block and the main block, represents the density after the merger of the main block and the ordinary block, represents the initial density of the main block;

[0030] such as , the ordinary block is regarded as a new independent cluster;

[0031] Repeat the above judgment process for the remaining blocks until there are no blocks that meet the merging conditions.

[0032] Further, in step S3, the clustering result is associated with the start and end markers of the device operation stage to construct a state transition matrix; the row and column indices of the state transition matrix represent the operation stages of the device, and the values in the matrix indicate whether a transition can occur between different stages;

[0033] Within a period of time, for the same device, signals with similar power amplitudes and similar change trends are grouped into one group and determined as a complete consumption cycle;

[0034] For a cycle containing s power amplitude transitions, according to the zero loop sum constraint, the following conditions are satisfied:

[0035] ,

[0036] ,

[0037] where is the s-th transition in the initial cycle, is the power amplitude transition value, is the threshold value defining the effective tolerance interval of the cycle.

[0038] Compared with the prior art, the remarkable effects of the present invention are as follows:

[0039] 1. The present invention adopts a heuristic grid clustering algorithm to achieve optimized event grouping. Compared with traditional clustering methods (such as Kmeans, KNN), this algorithm can process high-dimensional data (including the time dimension) more efficiently, reduce noise interference, and improve the accuracy of event grouping;

[0040] 2. The load reconstruction based on the state transition matrix of the present invention combines the clustering result to construct a steady-state transition probability matrix of the device, and identifies the complete cycle of the device operation mode (such as switch state, multi-mode switching); for multi-state devices (such as type II devices), by dynamically adjusting the transition matrix and error correction algorithm, it ensures that the total power change in each cycle approaches zero (zero loop sum constraint, ZLSC), solves the complexity problem of load decomposition of multi-mode devices, and improves the accuracy of energy consumption inference through mathematical constraints; it can help consumers understand the energy consumption of each electrical appliance, optimize the use and reduce energy waste;

[0041] 3. The present invention breaks through the limitation of traditional NILM that depends on the total load, decomposes the time attribute as the core feature and inputs it into the model; and conducts multi-dimensional feature fusion, combines the total load data (F(t)) with time features (hour (H(t)), week (D(t)), month (M(t)) to form a four-dimensional input tensor; the introduction of time features enables the model to capture the periodicity of electricity consumption behavior (such as differences between weekdays / weekends, seasonal changes); compared with traditional methods that it is difficult to distinguish the operation modes of different devices only through power changes (for example, air conditioners and electric water heaters may have similar power fluctuations), multi-dimensional feature fusion helps the model associate the typical usage periods of specific devices (such as air conditioners being frequently used during the day in summer), can solve the problem of device load mode confusion, and increases the adaptability of the devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is an energy decomposition diagram based on events;

[0043] Figure 2 is the overall flow chart of the present invention;

[0044] Figure 3 is the flow chart of event detection;

[0045] Figure 4 is a schematic diagram of the operation modes and state transitions of different devices, where (a) is a schematic diagram of the state transition of a device with two operation modes, (b) is a schematic diagram of the state transition of a device with three operation modes, and (c) is a schematic diagram of the state transition of a device with four operation modes;

[0046] Figure 5 is a performance diagram based on the correct allocation of energy percentage, where (a) is the performance diagram of the proposed model on the RAE dataset based on the correct allocation of energy percentage, and (b) is the performance diagram of the proposed model on the REDD dataset based on the correct allocation of energy percentage;

[0047] Figure 6 is a comparison diagram of the decomposition results and the real data within 24 hours, where (a) is the comparison diagram of the decomposition results and the real data within 24 hours of the RAE dataset, and (b) is the comparison diagram of the decomposition results and the real data within 24 hours of the REDD dataset. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The present invention will be further described in detail below in conjunction with the accompanying drawings of the specification and the specific embodiments.

[0049] The present invention proposes a NILM method based on heuristic grid clustering, aiming to infer the energy consumption of each electrical appliance by analyzing the household load power. This method detects the operation mode period of the electrical appliance through the steady-state transition probability matrix and conducts detailed energy consumption inference. It is carried out on low-frequency sampled household data, verifying that this data-driven model can efficiently and accurately estimate power consumption. This method constructs a unique non-intrusive analysis platform, introduces feature engineering on the basis of the typical NILM framework (Non-Intrusive Load Monitoring Framework) to increase the dimension of input data, and adopts an error correction algorithm to improve the credibility of the decomposition method. In terms of data, low-frequency data from the public RAE and REDD datasets are selected for experiments. In the data preprocessing stage, the time attributes in the time series are extracted and transformed into a four-dimensional tensor and normalized; in the event detection stage, the total load measurement sequence is processed to generate a four-dimensional matrix; the heuristic grid clustering algorithm is used to group the events; in the load reconstruction link, a device state transition matrix is created, the error is corrected, and the energy consumption is estimated. To solve the problems existing in the prior art and achieve more accurate and efficient non-intrusive load monitoring. The specific steps are as follows:

[0050] First step, perform data processing;

[0051] Select the RAE and REDD datasets, and divide them into a training set, a test set, and a validation set according to the data characteristics and experimental requirements. Extract the time attributes from the time series, transform them into a four-dimensional tensor containing the total load F(t), hour H(t), day of the week D(t), and month M(t), and make the data reach a unified scale through sine function and normalization processing, and segment by day.

[0052] Second step, execute event detection;

[0053] Use the time window to process the total load measurement sequence. If the local change of the total load within the window is greater than or equal to the threshold , or the absolute value of the power difference is greater than or equal to the threshold , then it is determined as an event, and a four-dimensional matrix containing the event amplitude and time reference is generated.

[0054] Third step, initialize a grid structure with 100 bins in the k-dimensional space, calculate the density and center point of each block after filling with event data, identify the main block, use the middle range as the dissimilarity index, and perform iterative merging of the blocks according to the formula to achieve effective grouping of the events.

[0055] Fourth step, carry out load reconstruction;

[0056] Construct the state transition matrix of the device according to the clustering results, analyze the unclassified events to determine the energy consumption pattern of the device, and reconstruct the energy consumption sequence of the device based on the energy consumption pattern of the device and the corrected state transition matrix, so as to complete the estimation of the device energy consumption and the calculation of the usage frequency, and reserve a tolerance of 5W for matching. Adjust the power amplitude transfer within the period according to the zero loop and constraint (ZLSC) to make the sum of the transfer amplitudes approach zero, and discard the periods that cannot be balanced within the set interval. Finally, reconstruct the energy consumption sequence of the device based on the energy consumption pattern of the device and the corrected state transition matrix, so as to complete the estimation of the device energy consumption and the calculation of the usage frequency.

[0057] Step 5, conduct result evaluation;

[0058] Adopt evaluation indicators for energy decomposition such as precision, recall rate, decomposition accuracy rate, and Fmeasure, and combine error quantification methods to compare and analyze the estimation results with the true values measured actually, so as to comprehensively evaluate the performance of the scheme.

[0059] As Figure 1 shown, a typical event-based NILM framework for non-intrusive analysis platforms includes event recognition, establishing an operation mode conversion matrix, grouping highly similar conversions, identifying the complete usage cycle of the device, reconstructing the sequence corresponding to the device consumption profile, and then deriving the share of each device. The present invention introduces additional optimization steps on the Figure 1 basis of the framework, increases the dimension of the model input data through feature engineering, incorporates important parameters other than the total load, and adopts an error correction algorithm after identifying the device consumption cycle to enhance the credibility of the decomposition method. The overall scheme is as Figure 2 shown. The steps are as follows:

[0060] Step 1, collect data;

[0061] The RAE (Rainforest automation energy) dataset was released in 2018, contains data of two families, the sampling frequency is 1Hz, and in addition to power data, there are also environmental and sensor data. The REDD (Reference energy disaggregation data set) dataset is a commonly used energy decomposition dataset, providing high-frequency (15kHz) and low-frequency (1Hz) data. In this embodiment, mainly low-frequency data is used because it is more in line with the commonly used power measurement devices in households.

[0062] Step 2, preprocess the collected data;

[0063] The data processing part covers data preprocessing and feature engineering. In the data preprocessing stage, considering that household energy consumption is comprehensively affected by various factors (such as season, week, daily time period, geographical location, house size, and equipment used, etc.), the corresponding time attributes are extracted from the time series and transformed into a four-dimensional tensor containing total load, hour, week, and month. Subsequently, the sine function is used to transform these time data into a periodic sequence and normalized to have a unified scale.

[0064] 2.1) Data division;

[0065] The RAE dataset and the REDD dataset are selected as the experimental data sources. In actual data processing, although there is no clear similar time node division method, in actual application, the dataset can be divided into a training set, a test set, and a validation set according to the data characteristics and experimental requirements to serve model training and evaluation.

[0066] 2.2) Data preprocessing;

[0067] Time attributes are extracted from the time series. Considering that household energy consumption is affected by factors such as season, week, and daily time period, it is transformed into a four-dimensional tensor containing total load F(t), hour H(t), week D(t), and month M(t). The specific process is to transform the time data into a periodic sequence through the sine function and then normalize it to reach a unified scale. The expression is:

[0068] (1)

[0069] Among them, , , represent the total number of hours in a day, the total number of days in a week, and the total number of months in a year respectively; i represents the i-th time point, represents the total load, represents the hourly load, represents the weekly load, represents the monthly load. The processed data is segmented by day and analyzed based on the time axis.

[0070] In the third step, through the event detection and clustering link, using the intermediate range as the dissimilarity index, the grid blocks are iteratively merged according to specific conditions of density and intermediate range, so that events with similar characteristics are grouped into the same category, realizing the effective grouping of events;

[0071] 3.1) Event detection;

[0072] As Figure 3 shown, using the time window Process the total load measurement sequence. If any of the following conditions is met within the time window, it is determined as an event:

[0073] One is the local change of the power load greater than or equal to the threshold ;

[0074] The expression is as follows:

[0075] (2)

[0076] Among them, represents the average power within the time window, represents the power measurement value at the i-th time point within the time window, and n is the number of measurement points within the time window.

[0077] The local change of the total load reflects the power fluctuation within the time window: If the total load (i.e., the total power) fluctuates greatly within the time window, that is the value is large, it may mean that the operating state of an electrical appliance has changed, such as the electrical appliance being turned on or off.

[0078] When a certain electrical appliance is suddenly turned on, the power will fluctuate greatly. At this time the value will increase accordingly. At this time, it is necessary to calculate the local average power as the transition amplitude; calculate the local average power within the time window The expression is as follows:

[0079] (3)

[0080] When the value is greater than or equal to the threshold , the local average power of this time window is registered as an event.

[0081] The other is the absolute value of the power difference between the input and output of the time window greater than or equal to ;

[0082] The absolute value of the power difference reflects the power change amplitude within the window. The calculation formula is as follows:

[0083] (4)

[0084] Among them, , are the power values at the start and end times of the time window respectively.

[0085] When is greater than or equal to it will be registered as an event.

[0086] Figure 3 The process of event detection is shown in detail.

[0087] 3.2) Heuristic grid clustering algorithm;

[0088] Initialize a grid structure containing N bins in the k-dimensional space. In this embodiment, N = 100 is taken, and the detected event data is filled into the grid structure. The grid space filled with data is defined as a block. Then, calculate the density and center point of each block, and classify the blocks into main blocks and ordinary blocks according to the density. Identify the main blocks, that is, the blocks with high density and whose directly adjacent blocks are not selected as main blocks.

[0089] Take the mid-range as the dissimilarity index. The mid-range calculation formula is:

[0090] (5)

[0091] where Mid-Range represents the mid-range, which is used to evaluate the difference degree of event features. By comparing the mid-range values of different event blocks, judge the similarity between events, and then classify the events with similar features into the same class; u represents an event in a block in the grid structure, represents the maximum value of the data contained in event u in the block in the grid structure, represents the block in the grid structure the maximum value of the data contained in event u in the block, represents the block in the grid structure the minimum value of the data contained in event u in the block.

[0092] Through the block merging operation, select the eligible blocks, and finally complete the classification of events. When merging, if the density plus the mid-range Mid-Range after the merger of the main block and the ordinary block is not greater than the initial density of the main block, that is , then merge the ordinary block with the main block; otherwise, the ordinary block is regarded as a new independent cluster. Repeat the above judgment process for the remaining blocks until there are no blocks meeting the merging conditions.

[0093] Step four, identify the state by constructing a state transition matrix, reconstruct the load using the error correction state, complete the entire cycle of identifying the device operation mode, and estimate the device energy consumption in combination with the load time.

[0094] 4.1) Identify the state transition matrix and the complete consumption cycle;

[0095] In the heuristic grid clustering algorithm, different clustering results represent different states of the device during operation. One clustering result represents one state of one device. By associating these clustering results with the start and end markers of the device operation phase, the transition situation of each device between different states can be accurately determined, and then the state transition matrix can be constructed. The row and column indices of the state transition matrix represent the operation phases of the device, and the values in the state transition matrix indicate whether a transition can occur between different phases. Figure 4 Graphically and intuitively shows the operation modes and state transition situations of different devices. Among them, Figure (a) presents the state transition situation of a device with two operation modes, Figure (b) shows a finite state machine with three operation modes, and Figure (c) illustrates the state transition of a device with four operation modes. The state transition situation is the key basis for determining the elements of the state transition matrix.

[0096] After the event assignment of all input axes is completed (i.e., all data input is completed), the cycles are grouped according to similar amplitudes. The signals from the same device are concentrated together to obtain the complete consumption cycle of the device. For example, by analyzing the power change signals of a device over a period of time, the signals with similar power amplitudes and similar change trends are grouped into one group and determined as a complete consumption cycle. For a device with multiple operation phases, in-depth research is required to determine its energy consumption pattern. The device corresponds to different power levels in different operation phases, and the state transition matrix closely links the operation state conversion of the device with the power change. With the help of this state transition matrix, not only can we know the power magnitude of the device in each state, but also we can clarify the specific change value of the power during state conversion. The state transition matrix is an important reference index for the state of electrical equipment, and analyzing the state transition matrix is a key step in identifying the complete consumption cycle of the device. As Figure 2 in, [0W, 21W, 15W, 41W, 0W] represents that a certain electrical appliance has different power states in a consumption cycle.

[0097] 4.2) Error correction algorithm;

[0098] According to the zero loop sum constraint (ZLSC), the power transfer amplitude within the cycle is adjusted to make the sum of the power transfer amplitudes as close to zero as possible. When the sum of the power transfer amplitudes is zero, it can be considered that the power is stable and unchanged during this cycle. In actual electrical load measurement, due to various interference factors, the measurement results often have errors, resulting in the sum of the power amplitude transitions deviating from the zero value. The error correction algorithm aims to adjust the power transfer amplitude to make the sum of the power amplitude transitions as close to zero as possible. For a cycle containing s power transfers, according to the zero loop sum constraint, the following conditions are satisfied:

[0099] (6)

[0100] (7)

[0101] Wherein, is the s-th transfer in the initial period, is the power amplitude transition value, is the threshold value defining the effective tolerance interval of the period.

[0102] If some periods cannot be balanced within the set interval, they will be discarded. According to the error correction, the load reconstruction is completed, and the accurate device state and state transition matrix are obtained, which can further accurately identify the device consumption period.

[0103] 4.3) Estimate the device energy consumption;

[0104] Based on the energy consumption period state transition matrix of the device, reconstruct the energy consumption sequence of the device, complete the estimation of the device energy consumption, and calculate the usage frequency of the device accordingly.

[0105] Step Five, result evaluation;

[0106] Adopt evaluation indicators specifically for energy decomposition, including Precision (P), Recall (R), Decomposition Accuracy (Acc.), and F-measure. The expressions are as follows:

[0107] (8)

[0108] (9)

[0109] (10)

[0110] (11)

[0111] Wherein, TP is the true positive example, TN is the true negative example, FP is the false positive example, and FN is the false negative example.

[0112] At the same time, combined with the error quantization method, compare and analyze the estimation result with the true value measured actually to comprehensively evaluate the accuracy of the device deduced by the present invention.

[0113] Through Step 1 to Step 4, the present invention can realize non-intrusive load monitoring based on adaptive heuristic grid clustering and accurately infer the energy consumption of each electrical appliance.

[0114] In this embodiment, low-frequency data from the RAE and REDD public datasets are selected. First, the data is divided, preprocessed, and subjected to feature engineering. The time attribute is transformed into a four-dimensional tensor and normalized. Subsequently, the existing method is improved in the event detection stage. The time window is used to identify events and generate a four-dimensional matrix. Then, the heuristic grid clustering algorithm is adopted to group the events. In the load reconstruction stage, the device state transition matrix is constructed, the error is corrected, and the energy consumption is estimated. Finally, using specialized evaluation metrics such as precision and recall, combined with the error quantification method, the estimated energy consumption is compared and analyzed with the actual measured value.

[0115] In the process of data processing and analysis of the present invention, metrics such as precision, recall, disaggregation accuracy, and F-measure are used to evaluate the present invention. From the experimental data (as shown in Table 1), relatively ideal results have been obtained in the monitoring of various electrical appliances for different datasets (RAE and REDD). For example, for the Lights & Plugs device in the RAE dataset, the precision reaches 0.990, the recall is 0.865, the F1 Measure is 0.923, and the disaggregation accuracy is 0.987. For the Oven device in the REDD dataset, the precision is 0.994. Although the recall is relatively low (0.554), the overall F1 Measure can still reach 0.711, and the disaggregation accuracy is 0.975. This indicates that in the energy consumption monitoring of most devices, the present invention can accurately identify the device state and reasonably allocate the energy consumption, with high reliability.

[0116] Table 1 Evaluation of the proposed solution based on common disaggregation metrics for 5 days of usage data

[0117]

[0118] Compared with the traditional methods (as shown in Table 2), the adaptive heuristic grid clustering method proposed by the present invention has obvious improvements in various metrics. Taking Precision as an example, in the traditional methods, K Means is 0.9241, K NN is 0.8954, etc., while the method of the present invention reaches 0.9565, with a significant improvement. This fully proves that the present invention has better performance in solving the problem of non-intrusive load monitoring compared with the traditional methods.

[0119] Table 2 Comparison of the results of the present invention with the technical performance of traditional methods

[0120]

[0121] The accuracy of the present invention is evaluated by comparing the consistency between the predicted energy consumption and the actual energy consumption. From Figure 5(Performance graph based on the correct allocation of energy percentage) It can be intuitively seen that on different devices in the REDD and RAE datasets, the present invention can correctly allocate most of the energy consumption to the corresponding devices. For example, for the dryer in the REDD dataset, the correct energy allocation percentage reaches a relatively high level, which means that the prediction of the energy consumption of this device by the present invention is highly consistent with the actual situation, effectively reducing the error of energy allocation.

[0122] As Figure 5 shown, it intuitively demonstrates the performance of the present invention in terms of the accuracy of energy consumption allocation. The abscissa represents different electrical appliances, and the ordinate represents the percentage of correctly allocated energy. Figure 5 In (a) and (b), the devices in different datasets (such as REDD and RAE) are distinguished by bar charts (or other graphical elements) of different colors or shapes. It can be clearly seen the accuracy of the present invention in allocating the energy consumption of each device under different datasets, intuitively demonstrating the performance of the present invention in terms of the accuracy of energy consumption allocation.

[0123] As Figure 6 shown is the comparison graph of the decomposition result and the real data within 24 hours. In the graph, the solid line and the dashed line represent the predicted result and the real result respectively, further showing the degree of fitting between the predicted curve of the device energy consumption and the actual energy consumption curve by the present invention within a 24-hour period. For example, for lamps and plugs, the predicted curve and the real curve are relatively close in trend and value, with only a certain deviation in a few time periods, which indicates that the present invention can accurately track the energy consumption change of the device in practical applications and has good real-time performance. The experimental results show that the present invention has high precision and efficiency in processing household data, can effectively separate the electrical energy loads of various types of electrical appliances, and has good generalization ability on different datasets.

[0124] As Figure 6 in (a) and (b), the abscissa is time (24 hours), and the ordinate is power (or energy consumption). Figure 6 In it, the solid line and the dashed line represent the predicted result (such as Predicted) and the real result (such as Actual) respectively. For different devices (such as lamps and plugs, dishwashers, etc.), the curves of the predicted and real results are plotted respectively to show the comparison of the device energy consumption prediction and the actual energy consumption by the present invention within 24 hours, helping readers intuitively understand the accuracy and stability of the present invention in real-time monitoring of device energy consumption.

Claims

1. An adaptive heuristic grid clustering method applicable to non-intrusive load monitoring, characterized in that, The steps include: S1, select low-frequency data from the RAE and REDD datasets and preprocess them, convert the selected data into a four-dimensional tensor containing total power, hour, week, and month, and obtain the total power measurement sequence; S2, using the time window to detect events in the total power measurement sequence, generating a four-dimensional matrix containing event amplitudes and time references, and grouping the events using a heuristic grid clustering algorithm. Different clustering results represent different states of the device during operation. The implementation process of grouping events using the heuristic grid clustering algorithm is as follows: S21, initialize a grid structure containing N boxes in a k-dimensional space, fill the detected event data into the grid structure; define the grid space filled with data as a block, then calculate the density and center point of each block, divide the blocks into main blocks and common blocks according to the density, and identify the main block, which has a high density and whose directly adjacent blocks are not selected as the main block; S22, using the middle range as the dissimilarity index, classifies events with similar characteristics into the same category; the calculation formula for the middle range is: Among them, Mid-Range represents the middle range, which is used to evaluate the degree of difference in event characteristics. By comparing the middle range values of different event blocks, the similarity between events is judged; u represents an event in block B in the grid structure i of, Max(u) represents the maximum value of the data contained in event u in block B in the grid structure i and Min(u) represents the minimum value of the data contained in event u in block B in the grid structure i ; S23, merging the blocks; Such as D union +Mid-Range ≤ D init , then merge the ordinary block with the main block, D union represents the density after the merger of the main block and the ordinary block, D init represents the initial density of the main block; Such as D union +Mid-Range > D init , the ordinary block is regarded as a new independent cluster; Repeat the above judgment process for the remaining blocks until there are no blocks that meet the merging conditions; S3, construct a state transfer matrix based on the event grouping results and the state changes of each device during operation; adjust the power amplitude transfer within the cycle based on the zero cycle and constraints so that the sum of the transfer amplitudes approaches zero; reconstruct the energy consumption sequence of the device based on the energy consumption cycle of the device and the state transfer matrix, complete the estimation of the energy consumption of the device, and calculate the usage frequency of the device; In step S1, the selected data is preprocessed as follows: the time attribute is extracted from the time series, converted into a four-dimensional tensor containing total power, hour, week and month, and the data scale is unified through sine function and normalization processing.

2. The adaptive heuristic grid clustering method applicable to non-intrusive load monitoring according to claim 1, wherein In step S2, if any of the following conditions is satisfied within the time window, it is determined as an event: the local change of the total power is greater than or equal to the threshold λ1; or the absolute value of the power difference ΔP(θ t ) between the input and output of the time window is greater than or equal to the threshold λ2; Among them, μ(θ t ) represents the average value of power within the time window, P(t i ) represents the power measurement value at the i-th time point within the time window, and n is the number of measurement points within the time window; t n represents the n-th time measurement point; When a certain electrical appliance is suddenly turned on, there will be a large power fluctuation. At this time, the value of will increase correspondingly. At this time, it is necessary to calculate the local average power ω(θ t ) as the transitional amplitude: When occurs, register ω(θ t ) as an event; The absolute value ΔP(θ t ) of the power difference between the input and output of the time window reflects the change range of the power within the window, and the calculation formula is as follows: ΔP(θ t ) = |P(t1) - P(t n )| Among them, P(t1) and P(t n ) are the power values at the start and end times of the time window, respectively; When ΔP(θ t ) ≥ λ2, register ΔP(θ t ) as an event.

3. The adaptive heuristic grid clustering method applicable to non-intrusive load monitoring according to claim 1, characterized in that In step S3, the clustering results are associated with the start and end marks of the device operation phase to construct a state transition matrix; the row and column indexes of the state transition matrix represent the operation phase of the device, and the values in the matrix indicate whether a transition can occur between different phases; For the same device within a period of time, signals with similar power amplitudes and similar change trends are grouped together to determine a complete consumption cycle; For a cycle containing s power amplitude transfers, according to the zero cycle and constraint, the following conditions are satisfied: Nd1, Nd2, …, Nd s = argmax(d1, d2, …, d s ) -λ3 < sum(Nd1 + Nd2, …, Nd s ) < λ3 where d s is the s-th transition in the initial period, Nd s is the power amplitude transition value, and λ3 is the threshold value defining the effective tolerance interval of the period.

Images