Intelligent optimization method of household energy ecology driven by electricity meter data fusion

Abstract

The present invention discloses a household energy ecological intelligent optimization method driven by electric energy meter data fusion, including: multi-source electric energy data collection and preprocessing; dual-time dimension reactive feature extraction and characterization; dynamic self-update of feature fingerprints based on gradual change perception; equipment health status assessment based on reactive features; household energy decomposition under multi-device overlapping state; energy ecological intelligent optimization decision. The present invention realizes dynamic tracking of equipment features through the history-present dual model structure, and realizes accurate monitoring of household energy use, equipment status assessment and energy optimization.

Description

Technical Field

[0001] The present invention relates to power electronics technology, and in particular to a household energy ecological intelligent optimization method driven by electric energy meter data fusion. Background Art

[0002] With the rapid development of smart grid and Internet of Things technologies, home energy management systems are playing an increasingly important role in optimizing energy use, reducing electricity bills and extending equipment life. Intelligent optimization of home energy ecology driven by electricity meter data fusion can not only improve energy efficiency and reduce carbon emissions, but also extend the life of household appliances through predictive maintenance, creating a more economical, comfortable and sustainable home energy environment for users. Research in this field is of great significance for building green smart homes, promoting the development of energy Internet and supporting the national energy conservation and emission reduction strategy.

[0003] At present, home energy management systems mainly use non-intrusive load monitoring (NILM) technology based on active power characteristics to identify the operating status of equipment, and provide energy optimization suggestions in combination with simple power consumption behavior patterns and energy consumption data analysis. Existing systems usually use centralized cloud computing architecture to process power data, use batch processing to analyze historical data, and rely on fixed equipment feature models for identification. In terms of multi-device status identification, mathematical methods such as independent component analysis (ICA) or non-negative matrix factorization (NMF) are mainly used to separate mixed power signals, and energy optimization is performed through simple statistical models. Equipment health monitoring is mainly based on threshold detection and simple rule judgment, and lacks the ability to track gradual changes in equipment performance over the long term.

[0004] However, there are several key technical difficulties in the practical application of the existing technology. First, the traditional NILM system mainly focuses on the active power characteristics and ignores the rich device status information contained in the reactive power characteristics. In particular, when multiple devices have similar active power characteristics but different reactive characteristics, the recognition accuracy drops significantly. Secondly, most of the existing device feature models use static representation, which cannot adapt to the performance gradient of the device over time, resulting in a gradual decrease in recognition accuracy over time. Specifically, when household appliances are used for more than 2-3 years, their reactive power characteristics will shift by 5%-10%, and the static model cannot capture this subtle change, resulting in an increase in the misrecognition rate of 15%-25%. In addition, when multiple household appliances are running at the same time, their power characteristics are superimposed in a nonlinear manner, generating harmonic interference and mutual influence, causing the recognition accuracy of the traditional linear decomposition method to drop sharply from 85% to below 55% when the number of devices exceeds 3, especially in fast and continuous switching scenarios. These technical difficulties restrict the effect and reliability of intelligent optimization of household energy ecology. Summary of the invention

[0005] The purpose of the invention is to provide a household energy ecological intelligent optimization method driven by electricity meter data fusion, in order to solve at least one technical problem existing in the prior art.

[0006] Technical solution, a household energy ecological intelligent optimization method driven by electric energy meter data fusion, comprising the following steps:

[0007] Collect and preprocess multi-source power data, generate standardized power data sets, and decompose household energy in the state of multi-device overlap based on them to generate energy consumption behavior patterns;

[0008] Extract and characterize reactive power features in dual time dimensions on standardized electric energy data sets to generate a reactive power feature fingerprint library for equipment;

[0009] For the reactive power feature fingerprint library of the equipment, the historical-current dual model is used to dynamically update the feature fingerprint of gradual perception, generate a time-varying feature evolution model and an equipment aging feature library, and evaluate the health status of the equipment based on this, generate an equipment health score report and fault warning information;

[0010] Combine energy consumption behavior patterns, equipment health score reports and fault warning information to generate a home energy ecosystem optimization plan.

[0011] Beneficial effects: Dynamic tracking of device features is achieved through the history-present dual model structure, which enables accurate monitoring of household energy use, device status evaluation and energy optimization. A dual-threshold event detection mechanism is used to accurately capture power events, and an event-triggered multi-resolution sampling strategy is implemented to efficiently capture complete features. The three-domain feature extraction and fusion of time domain reactive features, frequency domain reactive features and energy domain reactive features enable the system to distinguish devices with similar active power but different reactive characteristics. The household energy decomposition technology under the overlapping state of multiple devices solves the identification problem of multiple devices running simultaneously. The dual model is updated with different learning rates through the EWMA algorithm, the model difference index is calculated to quantify the change of device status, and the feature change attribution analysis is applied to decompose the change into random fluctuations, periodic changes and systematic trends, accurately identifying the performance degradation of the device. The device health status assessment technology based on reactive features solves the problem of device performance degradation detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a flow chart of the present invention.

[0013] Figure 2 It is a flow chart of the invention for extracting and characterizing reactive features in dual time dimensions for a standardized electric energy data set.

[0014] Figure 3 It is a flow chart of the present invention for applying dual threshold event detection and classification to a standardized electric energy data set.

[0015] Figure 4 It is a flow chart of the event-triggered multi-resolution sampling mechanism implemented in the present invention.

[0016] Figure 5 It is a flow chart of the present invention for performing three-domain reactive power feature extraction on reactive power data in a multi-resolution electric power data set.

[0017] Figure 6 It is a flow chart of the present invention using the history-present dual model to dynamically self-update the feature fingerprint of gradual perception. DETAILED DESCRIPTION

[0018] like Figures 1 to 6 As shown, S1: Multi-source power data collection and preprocessing

[0019] S11: Read household smart meter data (including total active power, total reactive power, and power factor) and auxiliary device data (including smart socket data and IoT device status information), unify the sampling time standard through the timestamp alignment algorithm, and generate a time-aligned original data set.

[0020] S12: Analyze the time-aligned raw data set, use the wavelet threshold denoising method to remove high-frequency noise, use the modified Z-score algorithm to detect and mark outliers, apply the conditional interpolation algorithm to fill in the missing data, and generate the cleaned power data.

[0021] S13: The post-cleaning power data is integrated with the external environment data (including indoor and outdoor temperature, humidity, light and time information), and the correlation between environmental factors and power consumption patterns is quantified through the conditional mutual information calculation method to establish an environment-power consumption correlation model.

[0022] S14: Process the cleaned power data and environment-power consumption association model, use context-aware normalization method to standardize the data, extract first-order and second-order differential features through time series feature engineering, and generate a standardized power data set.

[0023] S2: Dual-time dimension reactive feature extraction and characterization

[0024] S21: Analyze the standardized electric energy dataset, apply the adaptive dual-threshold detection algorithm (high sensitivity threshold 0.5% for low-power devices and low sensitivity threshold 5% for high-power devices) to identify power events, perform preliminary classification of events through event feature analysis, and generate power event sequences and event type labels.

[0025] S22: According to the power event sequence and event type label, a dynamic sampling strategy is implemented: 200Hz high-frequency sampling is triggered for 0.5 seconds for the detected event to capture transient characteristics, and the sampling frequency is reduced to 2Hz in the steady-state stage. The circular buffer is used to store the data of the first 3 seconds of the event to generate a multi-resolution power data set.

[0026] S23: Process multi-resolution power data sets, analyze time-frequency characteristics through improved S-transform, calculate reactive power rate of change (RPC) index, apply discrete cosine transform (DCT) for sparse representation, and reconstruct compressed signals using orthogonal matching pursuit (OMP) algorithm to generate reactive power multi-domain feature set.

[0027] S24: Integrate reactive power multi-domain feature sets and event type labels, group similar features through unsupervised clustering methods, apply pattern matching algorithms to associate feature clusters with device types, and build a device reactive power feature fingerprint library. Each device fingerprint contains a steady-state feature vector and a transient feature sequence.

[0028] S3: Dynamic self-update of feature fingerprints based on gradient perception

[0029] S31: Based on the reactive power feature fingerprint library of the equipment, two parallel models are established for each type of equipment: a historical model (remains relatively stable as a reference benchmark) and a current model (dynamically updated to capture changes in equipment characteristics). Feature stability parameters and model update thresholds are set to form the initial state of the dual models.

[0030] S32: Monitor the newly collected reactive power multi-domain feature set, and update the current model through the exponentially weighted moving average (EWMA) algorithm: M_now=α*M_new+(1-α)*M_now, where α is the learning rate; update the historical model at a lower frequency: M_hist=β*M_now+(1-β)*M_hist, where β<<α; calculate the model difference index: Δ=||M_now-M_hist|| / ||M_hist||, and generate a model difference report.

[0031] S33: Analyze the model difference report, apply the feature decoupling algorithm to separate random fluctuations and systematic changes, identify the feature dimensions with the most significant changes through principal component analysis, determine the cause of the change in combination with the equipment domain knowledge base, and generate feature change attribution results and equipment status assessment.

[0032] S34: Based on the feature change attribution results and equipment status evaluation, a differentiated update strategy is implemented: when normal equipment aging is confirmed, the historical model is triggered to accelerate the update; when it is determined to be a potential fault, the historical model is kept stable as a reference; the learning rates α and β are dynamically adjusted to generate a time-varying feature evolution model and an updated equipment aging feature library.

[0033] S35: Monitor the scale and performance of the time-varying feature evolution model, identify and retain key features through the feature importance scoring mechanism, design a forgetting mechanism to remove obsolete or redundant features, achieve model lightweighting, and maintain the representativeness and computational efficiency of the device aging feature library.

[0034] S4: Equipment health status assessment based on reactive power characteristics

[0035] S41: Use the equipment reactive power feature fingerprint library and equipment aging feature library to establish a baseline value range of the Q / P ratio (reactive power to active power ratio) under normal operating conditions for each type of equipment, record the Q / P characteristic curves of different operating modes, and form an equipment Q / P baseline file.

[0036] S42: Analyze the Q / P change data in the time-varying feature evolution model, and implement three-level time scale monitoring: short-term monitoring (each operation cycle), medium-term monitoring (weekly), and long-term monitoring (monthly). Use the trend line extraction algorithm to filter out short-term fluctuations and generate a Q / P trend analysis report.

[0037] S43: Apply the Q / P trend analysis report and the equipment Q / P baseline file to execute the equipment judgment rules: for motor equipment, pay attention to the rising trend of the Q / P ratio (bearing wear characteristics); for resistor equipment, pay attention to the deterioration of the Q / P ratio stability (loose connection characteristics); for electronic equipment, pay attention to the falling trend of the Q / P ratio (capacitor aging characteristics), and generate an equipment health score report.

[0038] S44: Integrate equipment health score reports and feature change attribution results, apply Bayesian network reasoning models to calculate failure probability, evaluate potential failure types and urgency, predict failure time windows, and generate failure warning information and maintenance recommendations.

[0039] S5: Home energy decomposition under multi-device overlap

[0040] S51: Process standardized power data sets, construct a three-layer dictionary structure (base layer: single device features; middle layer: common device combinations; top layer: complex patterns), apply the orthogonal matching pursuit algorithm to decompose mixed signals, identify combinations of simultaneously operating devices, and generate preliminary device state estimates.

[0041] S52: Combine the preliminary device state estimation and the device reactive power feature fingerprint library to implement the active device detection mechanism, build a subspace only for the currently active devices, use feature importance scoring to retain key features, achieve resource-efficient device state identification, and generate accurate device operation status.

[0042] S53: Analyze the precise equipment operation status, detect the harmonic correlation between equipment, build a group interference model to divide the equipment into mutual interference groups and independent groups, apply the hierarchical harmonic correction strategy (first deal with the intra-group interference, then deal with the inter-group interference), and generate the corrected equipment energy consumption data.

[0043] S54: Integrate the corrected equipment energy consumption data and environment-power consumption correlation model, apply the time series pattern mining algorithm to identify periodic energy consumption behaviors, quantify the contribution of various behaviors to total energy consumption through energy consumption attribution analysis, and generate equipment-level energy consumption decomposition results and energy consumption behavior patterns.

[0044] S6: Energy Ecosystem Intelligent Optimization Decision

[0045] S61: Integrate equipment health score reports, fault warning information, equipment-level energy consumption decomposition results and energy consumption behavior patterns, establish a multi-objective optimization model, balance energy efficiency, equipment life, user comfort and cost factors, define optimization objective functions and constraints, and form an energy optimization problem definition.

[0046] S62: Solve the energy optimization problem definition, use the Pareto optimization method to generate multiple optimization strategies, apply the user preference model to screen the strategies, formulate equipment usage time adjustment suggestions, energy efficiency setting optimization and alternative solutions, and form energy efficiency optimization suggestions.

[0047] S63: Based on fault warning information and equipment health score reports, design a maintenance plan based on the principle of minimum intervention, optimize the selection of maintenance time points, evaluate the balance between maintenance costs and fault risks, and generate an equipment maintenance plan.

[0048] S64: Comprehensive energy efficiency optimization recommendations and equipment maintenance plans, apply system dynamics models to evaluate the long-term impact of each plan, consider seasonal changes in household energy use, propose overall energy ecology optimization strategies, and generate household energy ecology optimization plans.

[0049] The traditional device feature modeling method has the following limitations: as new data is acquired, the model is constantly updated, the original feature information is gradually lost, and the performance changes of the device cannot be effectively compared. Using an unchanging baseline model cannot adapt to the normal aging process of the device, resulting in a high false alarm rate. Regularly replacing the model completely lacks continuity and cannot capture the gradual process.

[0050] To this end, this application uses the historical model as a relatively stable reference benchmark, but unlike a completely fixed model, it can slowly adapt to the normal evolution of the equipment. By quickly adapting the current model to new features, in contrast to the historical model, sensitive detection of gradual changes can be achieved. By analyzing the nature, rate and pattern of the differences between the two models, normal aging of equipment and abnormal failure signs can be distinguished. The dual-model structure achieves a balance between system adaptability and stability, neither overreacting to short-term fluctuations nor ignoring long-term gradual trends.

[0051] Specifically, the historical model has the same basic components as the current model, but the update mechanism and function are different. Each model contains a feature vector unit, a feature dimension weight unit, a state metadata unit, and an associated device operation mode parameter unit. The feature vector unit is usually a 25-dimensional vector, which contains the time domain, frequency domain, and energy domain features extracted from the three-domain reactive power features. In the feature dimension weight unit, each feature dimension has a corresponding weight value, reflecting the importance of the dimension to device identification and status monitoring. The weight value affects the difference calculation and model update process. The state metadata unit contains the model creation time, the last update time, an update counter that records the cumulative number of updates, and a stability indicator that measures the stability of the model features. The associated device operation mode parameter unit includes the associated normal operating parameter range and feature variants of different operation modes.

[0052] The differences between the two models lie in the update frequency and learning rate, response sensitivity, and state adaptive adjustment mechanism.

[0053] Now the model is updated every time it acquires new feature data. The initial learning rate α is usually set to 0.05-0.15, and the update formula is: M_now(t)=α*M_new(t)+(1-α)*M_now(t-1). The historical model is usually updated every 30 days or when the cumulative update count reaches a preset threshold (such as 100 times). The initial learning rate β is usually 1 / 10 of α, about 0.005-0.015, and the update formula is: M_hist(t)=β*M_now(t)+(1-β)*M_hist(t-1).

[0054] Modern models are highly sensitive to new features and quickly adapt to short-term changes in device characteristics, but may be affected by noise and temporary fluctuations. Historical models are relatively insensitive to new features, mainly capturing long-term, continuous trends, and are insensitive to noise and temporary fluctuations.

[0055] The model now has a dynamic learning rate adjustment mechanism: when significant changes are detected, the learning rate is increased, and when stable, the learning rate is reduced. It is usually adjusted dynamically between 0.02-0.3. The historical model adjusts the update strategy according to the result of feature change attribution. When it is confirmed to be normal aging: accelerate the update (increase the β value); when it is determined to be a potential fault: maintain stability (reduce or suspend the update); the decision of whether to update or not is more complicated, considering not only the time factor, but also the nature of the change.

[0056] Today, models are mainly used to identify and characterize the current state of equipment, as a basis for real-time decision-making, and to provide short-term trend information. Historical models are mainly used as a reference for comparison, representing the "normal" or "expected" state of equipment and providing a basis for long-term evaluation.

[0057] When model oscillation is detected (N consecutive updates in opposite directions), the current model triggers the stabilization process. It also has the function of filtering abnormal features, which can identify and reduce the impact of abnormal samples by the outlier detection algorithm. The historical model contains a state rollback mechanism, which can roll back to the previous stable state when necessary; it has a gradual threshold control to limit the maximum change range of a single update. Through this dual-model structure, it is possible to maintain the "memory" of the initial state of the device and the sensitivity to changes at the same time, without losing important reference benchmarks due to excessive updates, and without being unable to adapt to the natural evolution of the device due to model rigidity.

[0058] Existing technologies usually use fixed frequency sampling or simple threshold-based event detection, which results in either data redundancy or missing key transient information. This scheme designs an adaptive multi-resolution sampling strategy based on power events.

[0059] Combination of forward-looking buffering and real-time sampling: The ring buffer stores data 3 seconds before the event occurs, combined with 200Hz high-frequency sampling after the event is triggered, so as to capture the context before and after the complete event, rather than just focusing on the data after the event. This solves the limitation of traditional methods that cannot capture the state before the event. In addition, the time and frequency are adaptively adjusted. For example, variable-speed motor devices trigger longer high-frequency sampling cycles, while devices with obvious steady-state characteristics quickly switch to low-frequency sampling to achieve precise matching of resource usage. Multi-level sampling rate smooth transition: Different from traditional two-level sampling (high frequency / low frequency), this solution adopts a gradual decrease mechanism of sampling frequency: 200Hz at the moment of the event → 20Hz after the event stabilizes → 2Hz after it is completely stable, avoiding information discontinuity caused by sudden changes in sampling rate.

[0060] Traditional device feature modeling methods usually use a single model that is updated or completely replaced over time, which cannot simultaneously maintain the memory of the original state of the device and the perception of changes. The historical-current dual model structure solves this fundamental problem: Unlike the single model progressive update in the prior art, this solution simultaneously maintains two parallel feature models (historical model and current model), which can not only maintain the memory of the initial state of the device, but also keenly capture changes. The two models are updated using the EWMA algorithm with different learning rates to ensure that the current model quickly adapts to new features (high learning rate α), while the historical model remains relatively stable (low learning rate β), establishing a reliable reference benchmark. The model difference index Δ=||M_now-M_hist|| / ||M_hist|| is used as a quantitative indicator of device state changes to achieve accurate tracking of the gradual change process. According to the attribution results of feature changes, a differentiated update strategy is implemented: when it is confirmed to be normal aging, the historical model update is accelerated to adapt to the new normal; when it is determined to be a potential fault, the historical model is kept stable as a healthy reference point. This dynamic balance mechanism has not been reported in the prior art.

[0061] Existing methods usually only detect whether a change has occurred, without analyzing the nature and cause of the change. The feature change attribution analysis system achieves an accurate understanding of the nature of the change: the change in reactive power characteristics is decomposed into three independent components: random fluctuations, periodic changes, and systematic trends, and only the systematic trends are focused on to avoid overreaction to noise. PCA is used to identify the feature dimensions with the most significant changes, and the changes in different feature dimensions are associated with changes in equipment status (such as the association between increased Q / P ratio and motor bearing wear) in combination with the equipment knowledge base. A classification tree structure for equipment status changes is established, and the feature change patterns are mapped to predefined status change categories, achieving a leap from "detecting changes" to "understanding the nature of changes."

[0062] Traditional methods mainly focus on single-point anomalies or short-term changes, ignoring the long-term evolution of the health status of the equipment. This solution introduces multi-scale time analysis and equipment health rules: analyze the Q / P change trend at three time granularities (each operation cycle, weekly level, and monthly level) at the same time, capture the change process at different rates, and thus distinguish temporary fluctuations from systematic degradation. Special health assessment rules are formulated for different types of equipment. For example, motor equipment focuses on the rising trend of the Q / P ratio (bearing wear characteristics), and electronic equipment focuses on the falling trend of the Q / P ratio (capacitor aging characteristics), to achieve accurate health assessment. Through the trend line extraction method, short-term fluctuations and long-term trends are separated by wavelet decomposition and reconstruction technology, which improves the detection sensitivity of weak long-term changes.

[0063] In another embodiment of the present application, S13 is specifically:

[0064] S131: Read external environmental data (including indoor and outdoor temperature, humidity, and light) and time attribute data (including timestamp, date, weekend / weekday mark), split the continuous environmental data into time periods aligned with the power data sampling period through the time window segmentation algorithm, and apply the seasonal decomposition method to extract the trend, seasonality, and random components of environmental factors to generate a characterized environmental data set.

[0065] S132: Process the cleaned power data, apply sliding window statistical analysis to calculate the power consumption pattern characteristics (peak power, average power, standard deviation, peak-to-valley ratio, reactive power / active power ratio) of each time window, use the time series pattern mining algorithm to identify recurring power consumption sequence patterns, and generate a power consumption pattern feature set.

[0066] S133: Combine the characterized environmental data set and the electricity consumption pattern feature set, and calculate the conditional mutual information value I(X;Y|Z) between each pair of environmental factors and electricity consumption characteristics through the k-nearest neighbor conditional mutual information estimation method, where X is the environmental factor, Y is the electricity consumption feature, and Z is the set of known influencing factors. The permutation test method is used to evaluate the significance of the mutual information and generate a factor association strength matrix.

[0067] S134: Based on the factor association strength matrix, the cross-correlation analysis method is used to detect the lagged impact of environmental factors on electricity consumption patterns. Multiple time scales (immediate impact, hourly lag, daily lag, and weekly lag) are considered, and the maximum information coefficient (MIC) method is applied to capture nonlinear associations to generate a time-lag association feature table.

[0068] S135: Integrate the factor association intensity matrix and the time-lag association characteristic table, use the multivariate regression tree method to construct the mapping relationship between environmental factors and electricity consumption patterns, adjust the influence weights of different environmental factors through an adaptive weight allocation mechanism, apply the model compression algorithm to retain the most explanatory association rules, and generate an environment-electricity consumption association model.

[0069] In another embodiment of the present application, S21 is specifically:

[0070] S211: Read the power data in the standardized electric energy data set, calculate the first-order derivative (power change rate) and the second-order derivative (power acceleration), use the wavelet denoising method to smooth the derivative curve, determine the derivative feature distribution of different equipment types through statistical analysis, and generate a power derivative feature set.

[0071] S212: Analyze the power derivative feature set, divide the equipment into a high-power equipment group (peak power > 1000W) and a low-power equipment group (peak power ≤ 1000W) through cluster analysis, and use the statistical distribution of historical data to determine the initial threshold (the initial threshold of the high-power equipment group is 5% of the average power, and the initial threshold of the low-power equipment group is 0.5% of the average power). Apply the equipment operation status adaptation algorithm to dynamically adjust the threshold and generate a dual-threshold parameter set.

[0072] S213: Combine the standardized electric energy data set and the dual threshold parameter set to monitor the power change rate and the absolute value of power simultaneously. Event detection is triggered when the following conditions are met: (1) the power change rate exceeds the corresponding threshold; or (2) the absolute value of power changes by more than the corresponding threshold; or (3) the cumulative change exceeds the threshold within the sliding window. Edge detection technology is used to accurately locate the event boundary and generate preliminary event detection results.

[0073] S214: Process the preliminary event detection results and the standardized electric energy data set of the corresponding time period, extract the feature vector of each detected event, including power jump amount, rise / fall time, steady-state power, power factor change, etc., apply scaling invariant feature transformation to make the feature insensitive to the absolute value of power, and generate an event feature vector set.

[0074] S215: Analyze the event feature vector set, use an unsupervised learning algorithm (improved DBSCAN clustering) to preliminarily classify the events, associate event clusters with known equipment types through a pattern matching algorithm, apply a credibility scoring mechanism to assign a type probability distribution to each event, and generate power event sequences and event type labels.

[0075] In another embodiment of the present application, step S23 is specifically as follows:

[0076] S231: Read reactive power data from a multi-resolution power data set, calculate the reactive power change rate RPC(t)=(Q(t)-Q(t-Δt)) / Δt, where Q represents reactive power, extract time domain statistical features (mean, variance, skewness, kurtosis) through adaptive sliding window statistical analysis, and apply a denoising autoregressive model to estimate the time domain dynamic characteristics of reactive power, generating a reactive power time domain feature set.

[0077] S232: Processing reactive power data in multi-resolution power datasets, applying an improved S-transform algorithm (computational complexity increased from O(N 2 logN) is reduced to O(NlogN)) to analyze the time-frequency characteristics of reactive power. The algorithm realizes adaptive adjustment of time-frequency resolution through frequency-related window function, extracts frequency edge features and spectrum centroid, calculates energy distribution in different frequency bands, and generates frequency domain feature set of reactive power.

[0078] S233: Analyze the reactive power data in the multi-resolution power data set, decompose the signal into different frequency bands through wavelet packet decomposition, calculate the energy proportion and time-varying characteristics of each frequency band, extract the energy concentration and energy periodicity indicators of the reactive power fluctuation mode, apply multi-resolution entropy analysis to evaluate the signal complexity, and generate the reactive power energy domain feature set.

[0079] S234: Integrate the reactive power time domain feature set, reactive power frequency domain feature set and reactive power energy domain feature set, use principal component analysis to reduce the feature dimension, apply discrete cosine transform (DCT) to sparsely represent the fused features, retain the most discriminative features through the device category adaptive feature selection algorithm, dynamically adjust the compression ratio (5:1 to 20:1) according to the signal complexity, use the orthogonal matching pursuit (OMP) algorithm to reconstruct the compressed signal, and generate a reactive power multi-domain feature set.

[0080] In another embodiment of the present application, S24 is specifically:

[0081] S241: Read the reactive power multi-domain feature set and event type labels, apply the improved k-means++ clustering algorithm to group events with similar features, use the density peak detection method to determine the representative prototype of each cluster, identify and remove noise samples through the outlier detection algorithm, and generate a feature cluster prototype set.

[0082] S242: Combine the feature cluster prototype set and event type label, establish the association mapping between feature cluster and device type through multi-label propagation algorithm, calculate the feature space distribution of each device type, apply the probabilistic graph model to quantify the feature distribution differences under different operation modes, and generate the device type feature map.

[0083] S243: Based on the device type feature mapping, a dual fingerprint representation is constructed: the feature vector of the steady-state stage (continuous state feature fingerprint) and the feature sequence of the transient stage (time series feature fingerprint) are extracted, the length of the transient sequence is standardized using the dynamic time warping (DTW) algorithm, the key feature dimensions are highlighted through the feature weighting strategy of the device, and the original fingerprint of the device reactive feature is generated.

[0084] S244: Process the original fingerprint of the reactive power feature of the equipment, reduce the storage space requirement through lossless compression algorithm, design hierarchical index structure to accelerate fingerprint retrieval, use interpretable feature tagging method to improve fingerprint readability, apply hash coding technology to generate the unique identifier of the equipment fingerprint, and generate the final equipment reactive power feature fingerprint library.

[0085] In another embodiment of the present application, S32 is specifically:

[0086] S321: Read the newly collected reactive power multi-domain feature set and the current dual-model initial state, calculate the Mahalanobis distance between the new feature and the current model, set the abnormal distance threshold of the equipment type, standardize the distance metric through the Z-score method, apply the exponential sliding window technology to smooth the distance fluctuation, and generate a feature distance evaluation report.

[0087] S322: Based on the feature distance evaluation report, an adaptive learning rate adjustment algorithm is used: a smaller learning rate is used when the feature changes smoothly (the initial value of α is 0.05), and the learning rate is increased (up to 0.3) when significant changes are detected. A learning rate decay mechanism is designed to avoid over-response to short-term fluctuations. The upper and lower limits of the learning rate are adjusted considering the equipment operating status factors, and the updated learning rate parameters are generated.

[0088] S323: Combined with the updated learning rate parameters, the reactive power multi-domain feature set and the dual-model initial state, a dual-track update is performed: the current model is updated using the exponentially weighted moving average (EWMA) algorithm: M_now(t)=α*M_new(t)+ (1-α)*M_now(t-1), and the historical model is updated at a lower frequency: when the cumulative update count reaches the threshold N, M_hist(t)=β*M_now(t)+(1-β)*M_hist(t-1), where the β value is usually 1 / 10 of α, and the initial value is set to 0.005, to generate the updated dual-model state.

[0089] S324: Process the updated dual model status and calculate the model difference index: Δ=||M_now-M_hist|| / ||M_hist||, where ||*|| represents the weighted Frobenius norm. Consider the importance differences of different feature dimensions, establish a trend model of the difference index through time series analysis method, identify abnormal difference patterns, and generate a model difference report.

[0090] S325: Analyze the model difference report and the dual-model status after the update, evaluate the stability of the update process through model variance analysis, trigger the stabilization process when model oscillation is detected (N consecutive updates in opposite directions), adjust the learning rate or suspend the update, design a model rollback mechanism to deal with abnormal updates, and generate the updated dual-model status and update stability evaluation results for use in subsequent steps.

[0091] In another embodiment of the present application, S33 is specifically:

[0092] S331: Read the model difference report and the updated dual model status, apply the seasonal-trend decomposition method to separate the characteristic changes into three independent components: random fluctuations, periodic changes (such as load cycles), and systematic trends (equipment characteristic changes), evaluate the decomposition reliability through the Bootstrap method, assign confidence intervals to each component, and generate a characteristic change component set.

[0093] S332: Process the systematic trend components in the feature change component concentration, apply principal component analysis technology to identify the feature dimension with the largest contribution, calculate the contribution rate of each feature dimension to the total change, use the local linear embedding (LLE) algorithm to explore the low-dimensional change pattern in the high-dimensional feature space, apply the feature importance ranking algorithm to identify the most discriminative change features, and generate a change-sensitive dimension report.

[0094] S333: Integrate the equipment status feature model in the change-sensitive dimension report and the equipment domain knowledge base, build a mapping relationship between feature changes and equipment status changes, map the feature change pattern to the predefined status change category (such as "bearing wear", "capacitor aging", "loose connection") through the fuzzy reasoning system, calculate the probability distribution of each possible state, and generate the change pattern mapping result.

[0095] S334: Based on the change pattern mapping results, execute the cross-validation process: check whether the changes in different feature dimensions support the same state hypothesis, analyze the environmental factor data to eliminate the possibility of external interference, apply Bayesian posterior probability calculation to evaluate the confidence of the state change hypothesis, generate the final change attribution conclusion through the decision tree model, and output the feature change attribution result.

[0096] S335: Combine the feature change attribution results, change sensitive dimension reports and historical state evolution data, apply case-based reasoning (CBR) similar case matching, evaluate the change rate and severity, predict the state evolution trend, and generate an equipment state assessment report that includes state assessment, change cause, severity and recommended actions.

[0097] In another embodiment of the present application, S43 is specifically:

[0098] S431: Read the Q / P trend analysis report and the equipment Q / P baseline file, and select the corresponding health assessment indicator system according to the equipment type: motor equipment focuses on the upward trend of Q / P ratio and increased volatility; resistor equipment focuses on Q / P ratio stability and power factor fluctuations; electronic equipment focuses on Q / P ratio downward trend and high-frequency feature changes, establish an equipment scoring weight matrix, and generate an equipment evaluation indicator set.

[0099] S432: For motor equipment (such as refrigerator compressors, air-conditioning compressors, and fan motors), analyze their Q / P trend analysis reports, focus on calculating the Q / P ratio increase rate and steady-state Q / P fluctuation, apply the bearing wear feature recognition algorithm to detect changes in characteristic frequency components, evaluate the winding status through starting current analysis, and combine the equipment operation acoustic characteristics (if any) to assist in judgment and generate a health score for the motor equipment.

[0100] S433: For resistor-type equipment (such as electric water heaters, electric heaters, and electric furnaces), analyze their Q / P trend analysis reports, focus on calculating the Q / P ratio stability index and power fluctuation characteristics, apply the contact looseness detection algorithm to identify small power fluctuation patterns, evaluate the state of heating elements through thermal cycle response analysis, detect local aging characteristics of heating elements, and generate a health score for resistor-type equipment.

[0101] S434: Analyze the Q / P trend analysis report for electronic devices (such as TVs, computers, and chargers), focus on calculating the Q / P ratio decline trend and high-frequency feature changes, apply capacitor aging feature recognition algorithms to detect power filter capacitor performance, evaluate the switching power supply status through power quality analysis, monitor high-frequency harmonic change characteristics, and generate electronic equipment health scores.

[0102] S435: Integrate the health scores of motor equipment, resistor equipment and electronic equipment, apply the fuzzy comprehensive evaluation method to generate the overall health status score of the equipment (0-100 points), divide the health status level (excellent / good / medium / poor / critical), combine historical scoring data to establish a health trend model, predict future health status changes, and generate an equipment health score report.

[0103] In another embodiment of the present application, S51 is specifically:

[0104] S511: Read the standardized electric energy data set and the equipment reactive power feature fingerprint library, and build a hierarchical dictionary: the base dictionary D_base contains single-device feature atoms (directly extracted from the fingerprint library); the middle-level dictionary D_mid contains common equipment combination features (obtained or synthesized through controlled experiments); the top-level dictionary D_top contains complex patterns and abnormal patterns (obtained by clustering non-matching patterns), and applies atomic orthogonalization processing to improve the dictionary discrimination, generating a three-layer dictionary structure.

[0105] S512: Based on the three-layer dictionary structure, design hierarchical-aware sparsity parameters: base-layer sparsity λ_base (higher value, promotes basic device recognition), middle-layer sparsity λ_mid (medium value, balances common combination representation), top-layer sparsity λ_top (lower value, allows flexible representation of complex patterns), use cross-validation method to optimize sparsity parameters, and generate optimized sparse coding parameters.

[0106] S513: Processing standardized electric energy data sets, three-layer dictionary structures and optimized sparse coding parameters, implementing tracking orthogonal matching pursuit algorithm: first use the base-level dictionary to match basic devices, project the residuals to the middle-level dictionary space to identify device combinations, and finally use the top-level dictionary to capture complex patterns. The algorithm adaptively controls the number of iterations through residual energy, dynamically adjusts the dictionary selection strategy, and generates a hierarchical sparse representation result.

[0107] S514: Analyze the hierarchical sparse representation results, apply the physical constraint verification algorithm to check whether the decomposition results comply with energy conservation and equipment operation logic, use the timing consistency check to identify unreasonable equipment state jumps, evaluate the credibility of the decomposition results through the Bayesian model, adjust unreasonable decomposition results, and generate a verified sparse representation.

[0108] S515: Based on the verified sparse representation, apply the dictionary index to device mapping algorithm to convert the sparse coefficients into device switching states and power estimates, use a smoothing filter to process state transitions, optimize the state sequence through the Markov decision process model, consider device operation constraints (such as minimum on time), and generate preliminary device state estimates.

[0109] In another embodiment of the present application, S53 is specifically:

[0110] S531: Read precise equipment operating status and standardized power data sets, apply fast harmonic correlation detectors to analyze harmonic impacts between devices: calculate the harmonic interference coefficient of each pair of simultaneously operating devices, use graph clustering algorithms to divide devices into mutual interference groups (significant harmonic impacts) and independent groups (negligible harmonic impacts), build a harmonic correlation graph model, and generate device interference grouping results.

[0111] S532: Based on the device interference grouping results, a harmonic mutual interference matrix (HIM) is constructed for each mutual interference group: H_{i,j} represents the harmonic influence coefficient of device i on device j. The matrix elements are estimated through controlled experiments or historical data analysis. Sparse matrix technology is used to store only significant influence coefficients (>0.05). A matrix update mechanism is designed to adapt to changes in device characteristics and generate a set of intra-group mutual interference matrices.

[0112] S533: Process the precise equipment operation status, equipment interference grouping results and intra-group mutual interference matrix set, and implement two-level harmonic correction: first perform intra-group correction (use intra-group mutual interference matrix to correct nonlinear effects between devices in the same group), then perform inter-group correction (process weak interference effects between different groups), apply iterative least squares algorithm to solve the correction equation, and generate harmonic correction value.

[0113] S534: Reconstruct the actual energy consumption of each device by combining the precise device operating status and harmonic correction: E_true(i) =E_measured(i)±E_correction(i), where E_correction is calculated based on the harmonic correction, and the energy conservation constraint is applied to ensure the balance of total energy consumption. The reliability of the correction result is evaluated through confidence interval analysis to generate the corrected device energy consumption data.

[0114] S535: Analyze the corrected equipment energy consumption data and the standardized electric energy data set, calculate the total reconstruction error and the equipment-level reconstruction error, use the cross-validation method to evaluate the correction effect, fine-tune the mutual interference matrix parameters through the gradient descent algorithm to minimize the reconstruction error, record the correction effect and parameter update information for subsequent optimization, and generate the final corrected equipment energy consumption data and correction effect evaluation report.

[0115] According to one aspect of the present application, an implementation case is provided:

[0116] The test was conducted in a typical home environment, which was equipped with a smart meter and multiple smart sockets, covering a total of 8 common household appliances: refrigerator, air conditioner, washing machine, TV, electric water heater, rice cooker, desktop computer and LED lighting. By applying the "gradually perceived reactive power fingerprint self-update system" of the present invention, accurate monitoring of household energy use, equipment status evaluation and energy optimization were achieved.

[0117] During the implementation process, the total active power P_total(t) and total reactive power Q_total(t) data were first collected from the smart meter with a sampling frequency of 1Hz, and the sub-device power data of the main electrical appliances were collected from 5 smart sockets. The system also collected indoor temperature T_in(t), outdoor temperature T_out(t) and relative humidity RH(t) as environmental parameters. When preprocessing the original data, it was found that about 2.3% of the data points were noisy or missing. The system applied the wavelet threshold denoising method, selected the db4 wavelet basis function, decomposed the number of layers into 4 layers, and used the soft threshold method to process the high-frequency coefficients, removing more than 95% of the noise. For missing data, the conditional interpolation algorithm was used according to the data characteristics: linear interpolation was used for short-term missing (<30 seconds), spline interpolation was used for medium-term missing (30 seconds-5 minutes), and similar day pattern replacement was used for long-term missing (>5 minutes). Through conditional mutual information analysis, the correlation strength between environmental factors and power consumption patterns is quantified, and the calculation formula is: I(X;Y|Z)=∑∑∑p(x,y,z)log[p(x,y|z) / (p(x|z)p(y|z))]; where I(X;Y|Z) represents the conditional mutual information between X and Y under the condition of known Z; X represents environmental factors (such as temperature); Y represents power consumption characteristics (such as total power); Z represents a set of known influencing factors (such as time characteristics); and p(*) represents the probability distribution function. The analysis results show that the conditional mutual information value between outdoor temperature T_out(t) and air conditioning power consumption is the highest (0.83), and the conditional mutual information value between indoor temperature T_in(t) and electric water heater use is 0.64, indicating that these environmental factors have a significant impact on the power consumption pattern of the equipment.

[0118] During the implementation, two detection thresholds are adaptively set according to the power distribution characteristics:

[0119] For high-power devices (devices with power greater than the preset value, including refrigerators, air conditioners, washing machines, and electric water heaters), the power change detection threshold λ_high=5%×P_avg is set, where P_avg is the average power of this type of device; For low-power devices (devices with power less than the preset value, including TVs, LED lighting, and chargers), the power change detection threshold λ_low=0.5%×P_avg is set. To verify the effectiveness of the dual-threshold mechanism, the detection performance of the single threshold (3%) and dual-threshold methods was compared: the detection rate of the single threshold method for low-power devices was only 68%, while the dual-threshold method increased to 91%, while maintaining a high detection rate for high-power devices (97%).

[0120] After a power event is detected, an event-triggered multi-resolution sampling strategy is implemented: when an event is triggered, the sampling frequency is increased to 200 Hz for 0.5 seconds to capture complete transient characteristics; in the initial stable stage, the sampling frequency is reduced to 20 Hz for 5 seconds; in the fully stable stage, the sampling frequency is reduced to 2 Hz.

[0121] To achieve three-domain reactive power feature extraction, the following calculations are performed:

[0122] Time domain feature extraction, calculate the reactive power change rate RPC(t)=[Q(t)-Q(t-Δt)] / Δt; where Q(t) represents the reactive power at time t, and Δt is the time interval (Δt=0.005 seconds in the high-frequency sampling stage).

[0123] Frequency domain feature extraction, application of improved S transform analysis time-frequency characteristics S(τ,f)=∫x(t)*w(τ-t,f)*e -i2πft dt; where S(τ,f) represents the S-transform coefficient of the signal x(t) at time τ and frequency f; w(τ-t,f) is the frequency-dependent Gaussian window function, defined as: w(τ-t,f)=(|f| / √2π)*e -(τ-t)2f2 / 2 ; To improve computational efficiency, the fast S-transform algorithm was implemented, reducing the computational complexity from O(N 2 logN) is reduced to O(NlogN). Perform Fourier transform on the signal to obtain X(f); for each frequency f, multiply X(f+m) by the frequency domain Gaussian window W(m,f); and perform inverse Fourier transform on the product.

[0124] Energy domain feature extraction: Through 5-layer wavelet packet decomposition, the signal is decomposed into 32 frequency bands, and the energy ratio of each frequency band is calculated, E_ratio(i)=E(i) / E_total; where E(i) represents the energy of the i-th frequency band, and E_total is the total energy. Through feature extraction in these three domains, devices with similar active power but different reactive characteristics are successfully distinguished. For example, the active power of an 800W electric water heater and a 750W microwave oven are similar, but their reactive power characteristics are significantly different: the Q / P ratio of the electric water heater is stable at around 0.05, while the Q / P ratio of the microwave oven is 0.31 and has obvious high-frequency components.

[0125] Two parallel models are maintained for the refrigerator: the historical model M_hist and the current model M_now. Each model contains a 25-dimensional feature vector representing the reactive power characteristics of the equipment.

[0126] In the initial stage, the two models are the same: M_hist(0)=M_now(0)=M_initial, where M_initial is the initial feature vector of the refrigerator. When a new feature vector M_new(t) is obtained, the current model is updated by the exponentially weighted moving average (EWMA) algorithm: M_now(t)=α*M_new(t)+(1-α)*M_now(t-1); where α is the learning rate, initially set to 0.05. During the test, the α value is dynamically adjusted according to the feature change rate: when a significant change is detected (feature distance>2σ), the α value is increased to 0.15; when the feature is stable (feature distance<0.5σ), the α value is reduced to 0.02.

[0127] The historical model is updated at a lower frequency: an update is performed every 30 days or when the cumulative update count reaches the preset threshold (N=100): M_hist(t)=β*M_now(t)+(1-β)*M_hist(t-1); where β is the historical model learning rate, which is set to 1 / 10 of α and the initial value is 0.005.

[0128] Calculate the model difference index Δ, the formula is Δ=||M_now-M_hist||_W / ||M_hist||_W; where ||*||_W represents the weighted Frobenius norm, considering the importance of different feature dimensions ||M||_W=√(∑∑wij*mij 2 ); wij is the weight matrix, which is set according to the discriminative ability of the feature dimension. During the 6-month monitoring period, the model difference index Δ of the refrigerator increased significantly from the initial 0 to 0.082, exceeding the preset threshold of 0.07, triggering the feature change attribution analysis.

[0129] The seasonal-trend decomposition method is applied to decompose the time series of the difference index into three components: random fluctuation component R(t), daily fluctuation; periodic component S(t), related to the usage cycle; trend component T(t), reflecting the systematic changes in equipment characteristics; the decomposition adopts the STL (Seasonal-Trend decomposition using LOESS) method, and the calculation formula is: Δ(t)=T(t)+S(t)+R(t); analyzing the trend component T(t), it is identified that the most significant changes in the characteristic dimensions are the reactive power peak at the moment of compressor startup (increased by 14%) and the Q / P ratio during steady-state operation (increased by 9%). These two characteristics are highly correlated with compressor bearing wear and refrigerant leakage.

[0130] To confirm the attribution results, cross-validation was performed: checking whether changes in other feature dimensions support the "compressor performance degradation" hypothesis. Analysis shows that the starting current characteristics, operating acoustic characteristics, and temperature reaching time have also changed accordingly, further supporting the hypothesis. Calculate the confidence of the state change hypothesis: P(H|E) =P(E|H)*P(H) / P(E); where H represents the "compressor performance degradation" hypothesis, E represents the observed evidence of feature changes, and P(H|E) is the posterior probability (confidence). The calculation results show that the confidence level is 83%, which exceeds the confidence threshold of 75%. The system confirms that the refrigerator compressor is in an early state of performance degradation.

[0131] For refrigerators confirmed to have early performance degradation, multi-scale Q / P trend monitoring and equipment health rule assessment were implemented.

[0132] Establish the Q / P ratio baseline value range under normal refrigerator operation. As a motor-type device, the normal Q / P ratio baseline of the refrigerator is 0.31±0.03, which is mainly due to the inductive load characteristics of the compressor.

[0133] Monitor changes in the Q / P ratio at three time scales: short-term monitoring, calculate the average Q / P value for each operating cycle (about 30 minutes); medium-term monitoring, calculate the Q / P trend every week; long-term monitoring, establish the Q / P trend curve every month;

[0134] The monitoring data was decomposed using wavelet method to extract trend lines and filter out the impact of short-term fluctuations: wavelet transform was applied to the Q / P time series to decompose it into multiple scale components; low-frequency approximate coefficients were retained and high-frequency detail coefficients were removed; trend lines were reconstructed through inverse wavelet transform; after 6 months of monitoring, it was detected that the refrigerator Q / P ratio showed a slow but steady upward trend, with an average monthly growth rate of 1.5% and a cumulative growth of 9%, exceeding the warning threshold (7%) for motor equipment.

[0135] Apply the evaluation rules for motor-type equipment, focusing on: the upward trend of the Q / P ratio, reflecting bearing wear or refrigerant problems; changes in starting current characteristics, reflecting compressor efficiency; changes in operating cycle patterns, reflecting overall refrigeration system performance.

[0136] The reactive power waveform at the moment of compressor startup is analyzed, and the transient feature similarity Sim(W_current, W_baseline)=1-DTW(W_current, W_baseline) / L is calculated; where DTW represents the dynamic time warping distance, L is the sequence length normalization factor, W_current is the current waveform, and W_baseline is the baseline waveform. The calculation results show that the similarity drops from the initial 0.95 to 0.83, further confirming the degradation of compressor performance.

[0137] Based on multiple indicators, the fuzzy comprehensive evaluation method is used to generate the equipment health score Score=∑(wi*vi); where wi is the weight of each indicator and vi is the normalized indicator value. The refrigerator's health score dropped from the initial 92 points (excellent) to 78 points (medium), which is considered to be in a "needs attention" state, but has not yet reached the severity of "needs immediate repair" (below 65 points).

[0138] In a real home environment, the overlapping power characteristics formed by multiple devices running simultaneously pose a challenge to energy decomposition. This system uses hierarchical sparse representation and nonlinear combination effect correction technology to solve this problem.

[0139] A three-layer dictionary structure is constructed: the base dictionary D_base contains 8 single-device feature atoms with a dimension of 25×8; the middle-layer dictionary D_mid contains 15 common device combination features obtained through controlled experiments; the top-layer dictionary D_top contains 10 complex patterns obtained by clustering non-matching patterns; in order to improve the dictionary discrimination, the Gram-Schmidt orthogonalization process v'i=vi-∑<vi,vj> vj / ||vj|| 2 ; Among them, vi represents the original eigenvector, v'i represents the orthogonalized eigenvector, and <*,*> represents the inner product operation.

[0140] Implement the improved tracking-OMP algorithm to decompose the mixed signal: Initialize the residual r 0 =y (observed signal); for iteration step k, select the atom that best matches the current residual: j k =argmax | <r k-1 ,d j >|, where d j For atoms in the dictionary; update the support set: Ω k =Ω k-1 ∪{j k}; Solve the least squares problem: x k =argmin||yD k x|| 2 , where D k is the matrix of the selected atoms; update the residual: r=yD k x k ; if || r k || 2 <ε or the maximum number of iterations is reached, the algorithm terminates.

[0141] Unlike traditional OMP, Tracking-OMP adds an atomic correlation tracking mechanism: when selecting new atoms, the timing correlation with the selected atoms is considered, and atoms with temporal coherence are given priority, so as to more accurately identify device state changes. In the test, when the air conditioner, refrigerator and washing machine were running at the same time, the recognition accuracy of the traditional method was only 64%, while this system reached 87%, significantly improving the decomposition accuracy in multi-device scenarios.

[0142] Detect and correct nonlinear interference between devices, especially the impact on the measurement of other devices when the air conditioner is started. Through harmonic correlation analysis, it was found that there was significant harmonic interference between the air conditioner and the TV (the harmonic interference coefficient was 0.17). Construct a harmonic mutual interference matrix (HIM), and the element H_{i,j} represents the harmonic influence coefficient of device i on device j. For the air conditioner (i=1) and the TV (j=5), H_{1,5}=0.17 means that the operation of the air conditioner will cause a deviation of about 17% in the measured power of the TV. Apply a hierarchical harmonic correction strategy, that is, first deal with the intra-group interference and then the inter-group interference. The corrected device power calculation formula is: P_true(i)=P_measured(i)-∑ j≠i H_{j,i}*P_measured(j); the total reconstruction error before and after correction dropped from 15.4% to 6.8%, significantly improving the accuracy of energy decomposition.

[0143] Based on precise analysis of equipment status and energy usage, a smart optimization plan for home energy ecology is generated, including: Maintenance recommendations for refrigerators: Due to early performance degradation of the compressor, it is recommended to perform preventive maintenance in the next 3-6 months, clean the radiator and check the refrigerant level, which is expected to improve energy efficiency by about 12%.

[0144] Energy usage optimization suggestions: Based on the analysis of user behavior patterns, it was found that washing machines and electric water heaters are frequently used during the peak electricity price period (18:00-20:00). It is recommended to adjust them to the off-peak period (22:00-6:00), which is expected to save 15% of electricity bills per month.

[0145] Long-term equipment replacement planning: According to the health score trend forecast, refrigerators may need to be replaced within 18-24 months, and a recommended list of replacement models with an energy efficiency ratio (EER)>3.5 is provided. By implementing the above optimization plan, the test family achieved a 11.5% reduction in total energy consumption and a 17.3% reduction in electricity bills within 3 months, while extending the service life of the equipment and improving user satisfaction.

[0146] The dual-threshold event detection mechanism improves the recognition rate of low-power devices from 68% to 91%, while maintaining a high recognition rate of high-power devices (97%). The history-present dual model structure successfully captures the 9% performance degradation of the refrigerator compressor and accurately attributes it to bearing wear and refrigerant problems with a confidence level of 83%. The three-domain reactive power feature extraction method improves the differentiation rate of devices with similar active power but different reactive characteristics from 72% to 94%. The hierarchical sparse representation and nonlinear combination effect correction technology improve the recognition accuracy of multiple devices running simultaneously from 64% to 87%.

[0147] In another embodiment of the present application, the cleaned electric energy data is context-aware and standardized as follows: first, the standardization parameters are determined according to the environment-electricity consumption association model, and the calculation process is as follows: the data subsets are divided according to different environmental conditions, such as the high temperature period (T_out>28°C), the medium temperature period (18°C≤T_out≤28°C) and the low temperature period (T_out<18°C); the standardization parameters are calculated for each data subset: the mean μ_c and the standard deviation σ_c; the context-aware Z-score standardization formula P_norm(t)=(P_raw(t)-μ_c) / σ_c is applied; wherein P_raw(t) is the original power data, μ_c and σ_c are the mean and standard deviation under the current context c, respectively.

[0148] Perform time series feature enhancement and calculate the following features: first-order difference feature ΔP(t)=P(t)-P(t-1); second-order difference feature Δ 2 P(t)=ΔP(t)-ΔP(t-1); moving average characteristic MA_P(t,w)=(1 / w)∑ i=0 w-1 P(ti), where w is the window size; peak factor PF(t,w)=max(|P(ti)|, i∈[0,w-1]) / RMS(P,t,w), where RMS is root mean square;

[0149] In the test households, standardized parameter sets were established for weekdays and weekends. During weekdays, the mean value of active power was μ_weekday = 1240W, and the standard deviation was σ_weekday = 980W; during weekends, the mean value was μ_weekend = 1560W, and the standard deviation was σ_weekend = 1150W.

[0150] Through context-aware normalization, the system successfully reduced time-related fluctuations by 78%, making device characteristics more consistent across time periods and improving the accuracy of subsequent recognition.

[0151] In another embodiment of the present application, in event-triggered sampling, the specific implementation process of the ring buffer is as follows: initialize a fixed-size ring buffer B with a capacity of 3 seconds × basic sampling rate (1 Hz) = 3 samples:

[0152] B = [b 0 , b 1 , b 2 ], the initial value is null;

[0153] In normal sampling mode (1Hz), the ring buffer is continuously updated: b_idx=t mod 3; B[b_idx]={P(t), Q(t), timestamp(t)}; where P(t) and Q(t) are the active and reactive powers at time t, respectively, and timestamp(t) is the timestamp.

[0154] When a power event is detected, the following operations are performed: immediately lock and copy the current ring buffer content B_event = B; reorder by timestamp to obtain the data sequence 3 seconds before the event; switch to high-frequency sampling mode (200Hz) and start collecting post-event data.

[0155] Continuous performance optimization: Use compressed storage technology to reduce memory usage and apply delta encoding to stable period data; implement adaptive buffer size and adjust it according to device type (use longer buffers for variable speed devices, up to 10 seconds).

[0156] In actual tests, the ring buffer technology successfully captured 42% of the pre-startup status information of electrical appliances that would have been missed. In particular, for devices with a preheating stage (such as electric water heaters), the pre-state capture increased the recognition accuracy by 18 percentage points.

[0157] In another embodiment of the present application, during long-term operation, the growing size of the feature fingerprint library will lead to an increase in the computational burden. Implementing a model complexity adaptive management mechanism:

[0158] The importance of each feature dimension is evaluated regularly to obtain IS(f_i)=α*DS(f_i)+β·CS(f_i)+γ*TS(f_i); where: DS(f_i) is the discrimination score, which measures the ability of the feature to distinguish different devices, and the calculation formula is: DS(f_i) =σ_between(f_i) / σ_within(f_i); σ_between represents the standard deviation of the feature between different devices, and σ_within represents the standard deviation of the feature within the same device. CS(f_i) is the consistency score, which measures the stability of the feature under the same conditions. TS(f_i) is the timeliness score, which measures the time correlation of the feature. α, β, and γ are weight coefficients, which are set to 0.5, 0.3, and 0.2 in the implementation.

[0159] When the total number of features exceeds the preset threshold (N_max=100), the feature forgetting algorithm is applied: all features are sorted by importance score IS(f_i); the cumulative importance is calculated: CIS(k)=∑(k i =1) IS(f_i), features are arranged in descending order of importance; determine the minimum feature set K', so that CIS(K') / CIS(K) ≥ 0.95, that is, retain 95% of the information; remove the last 30% of the features ranked by importance, but retain at least K' features;

[0160] Apply sparse representation compression to the retained features: construct the feature autocorrelation matrix R = F*F T , where F is the characteristic matrix; perform eigenvalue decomposition R=V*Λ*V T ; retain the minimum set of feature vectors required to explain 90% of the variance; project the original features into a new low-dimensional space. During the 6-month test period, the feature dimension of the refrigerator was successfully compressed from the initial 25 dimensions to 16 dimensions, while maintaining the recognition accuracy above 93%, and the model calculation speed was increased by 37%, providing significant advantages for resource-constrained devices.

[0161] In another embodiment of the present application, the fault risk prediction model based on the Bayesian network is specifically calculated as follows: a Bayesian network with the following structure is constructed: root node, equipment type (refrigerator) and service life (3 years); intermediate node, Q / P ratio change, starting current characteristics, operation cycle mode, temperature reaching time; leaf node, possible fault type (compressor problem, refrigerant leakage, radiator blockage, etc.);

[0162] Conditional probability table (CPT) setting: Based on historical data and expert knowledge, fill in the conditional probability table, for example: P(compressor problem|Q / P ratio increase>7%, service life>2 years)=0.65; P(refrigerant leakage|temperature reaching time extension>15%, Q / P ratio increase>5%)=0.72;

[0163] Given the observed evidence E (Q / P ratio increased by 9%, starting current increased by 7%, and temperature reaching time extended by 12%), calculate the posterior probability of each fault type P(H_i|E)=P(E|H_i)*P(H_i) / ∑_j P(E|H_j)*P(H_j); where H_i represents the i-th fault hypothesis, P(H_i) is the prior probability, and P(E|H_i) is the likelihood probability.

[0164] Based on the fault progression model, the predicted fault development time window T_failure=T_current+(Threshold-CurrentValue) / ProgressRate, where Threshold is the fault threshold, CurrentValue is the current measurement value, and ProgressRate is the rate of change. The calculation results show that the most likely fault types of the refrigerator are "early wear of compressor bearing" (probability 0.58) and "minor refrigerant leakage" (probability 0.31). It is predicted that if no maintenance is performed, it will develop into a serious fault that requires replacement of the compressor within 14-20 months, and the fault risk level is determined to be "medium".

[0165] In another embodiment of the present application, in a multi-device environment, an active device detection mechanism is implemented to reduce computational complexity:

[0166] Maintain a device active state vector A=[a 1 ,a 2 ,...,a_n], where a_i∈{0,1} indicates whether the i-th device is active; device state changes are detected based on short-term power analysis: when the total power change ΔP_total exceeds the threshold θ and the duration exceeds the minimum operating time τ_min of the device, the active device re-evaluation is triggered;

[0167] Use the sliding window statistical model to determine the device activity probability P(a_i=1|data)=exp(S_i) / ∑_jexp(S_j); where S_i is the activity score of device i, calculated based on feature matching.

[0168] Subspace construction optimization: only build identification subspaces for currently active devices, not all devices; when only 3 out of 8 devices are detected to be active, only build identification subspaces for these 3 devices, reducing the computational complexity by 62%; the active device subset update adopts an inert strategy: the active device list is updated only when there is sufficient evidence to indicate that the device status has changed.

[0169] For each feature f_i, the information gain ratio IGR(f_i) = IG(f_i) / IV(f_i) is calculated; where IG(f_i) is the information gain and IV(f_i) is the intrinsic value used for normalization; the most discriminatory feature subset is selected according to the IGR value to reduce the feature dimension; in the test environment, dynamic complexity management reduces the system's computing resource requirements by 58% in a typical home scenario (an average of 3-4 devices are active at the same time), while the recognition accuracy only drops by 1.2 percentage points, achieving a good balance between accuracy and efficiency.

[0170] In another embodiment of the present application, a time series pattern mining algorithm is used to identify the energy consumption behavior pattern of the user: the device operation state is encoded as a symbol sequence S={s 1 ,s 2 ,...,s_T}; each symbol s_t represents the combination of equipment operating states at time t; define the pattern P as a continuous symbol subsequence P={s_i,s_{i+1},...,s_{i+k-1}} with a length of k;

[0171] Using the improved PrefixSpan algorithm to mine frequently occurring temporal patterns: Initialize the frequent 1-sequence set F 1 ={all frequent single symbols}; for each sequence s_α ∈ F 1 , recursively construct the projection database D|_s_α of s_α; mine local frequent patterns in D|_s_α to generate frequent supersequences of s_α; calculate the support sup(P) = count(P) / |D|, where count(P) is the number of times pattern P appears, and |D| is the total number of sequences;

[0172] Only patterns with support sup(P)≥ min_sup are retained. In the implementation, min_sup is set to 0.1. For the discovered frequent patterns P, the associated energy consumption E(P)=∑_t E(s_t), s_t ∈ P is calculated. The energy consumption ratio of the pattern R_E(P)= E(P) / E_total is calculated. A user behavior-energy consumption mapping map is constructed to identify high-energy consumption behavior patterns. The Shapley value method is used to quantify the contribution of different behaviors to total energy consumption. Five major energy consumption behavior patterns were identified in the test households, contributing 78% of the total energy consumption. Among them, the highest energy consumption pattern is "evening cooking + laundry + TV" (accounting for 23% of the total energy consumption). This pattern mainly occurs between 18:00-20:00 on weekdays, coinciding with the peak period of electricity prices, and has become the primary goal of energy optimization.

[0173] In another embodiment of the present application, a multi-objective optimization model for household energy is constructed to balance energy efficiency, equipment life, user comfort and cost factors:

[0174] Optimization objective function: Energy cost minimization objective f 1 (x)=∑_t∑_iP_i(t)*x_i(t)·price(t)*Δt; where P_i(t) is the power of device i at time t, x_i(t) is the operating state variable (0 or 1) of device i at time t, price(t) is the electricity price at time t, and Δt is the time interval;

[0175] Equipment life maximization objective f 2 (x)=-∑_i w_i*D_i(x); where w_i is the importance weight of device i, D_i(x) is the loss index of device i under scheme x, D_i(x)=∑_tα_i*x_i(t)+β_i*SW_i(x); α_i is the operating time loss coefficient, β_i is the switching number loss coefficient, and SW_i(x) is the number of switches;

[0176] User comfort maximization objective f 3 (x)=-∑_t∑_i u_i*|t - pref_i(t)|*x_i(t); where u_i is the comfort weight of device i, and pref_i(t) is the user's preferred usage time for device i at time t

[0177] Set constraints: total power constraint, ∑_i P_i(t)·x_i(t)≤P_max, device running time constraint, ∑_t x_i(t)≥T_min,i, device logic constraint, x_i(t+1)-x_i(t)≤y_i(t), where y_i(t) is the startup variable, minimum running time constraint, x_i(t)≥∑_{j=1}^{d_i} y_i(tj), where d_i is the minimum running time of device i, mutual exclusion constraint, x_i(t)+x_j(t)≤ 1, any (i,j)∈E_mutex, E_mutex is a set of mutually exclusive device pairs.

[0178] Using weighted sum method to integrate multiple objectives F(x)=w 1 *f 1 (x)+w 2 *f 2 (x)+w 3 *f 3 (x); where w 1 、w 2 、w 3 is the weight coefficient, which is set according to user preference; in the test home, the above model is used to construct a customized energy optimization problem, and the initial weight is set to w 1 =0.4, w 2 =0.3, w 3= 0.3, dynamically adjusted according to user feedback, and finally converged to w 1 =0.45, w 2 =0.25, w 3 =0.3, reflecting that the family pays more attention to energy cost factors.

[0179] In another embodiment of the present application, the seasonality-trend decomposition method STL separates the characteristic changes into three independent components, including:

[0180] In the refrigerator case, a time series of the difference index Δ {Δ 1 ,Δ 2 ,...,Δ 183} (one data point per day). The specific calculation process of STL decomposition is as follows:

[0181] First, set the decomposition parameters: cycle length n_p=7 (according to the weekly mode); trend filter length n_t=45 (about 1.5 months); seasonal filter length n_s=35; internal cycle number n_i=2; external cycle number n_o=1;

[0182] Calculate local regression smoothing (LOESS): For a given point i and window width q, calculate the distance weight d_j=|ji| / h, h=max(i,ni); w_j=(1-d_j 3 ) 3 , if d_j<1, otherwise 0;

[0183] Use these weights to perform local weighted regression and obtain smoothed values.

[0184] Initial trend estimate T (0) as a moving average;

[0185] Each inner loop k calculates: Detrended sequence D = Δ-T (k-1) ; Estimate initial seasonality and apply periodic LOESS smoothing to obtain S (k) ; Deseasonalized series R = Δ-S (k) ; Apply trend LOESS smoothing to get T (k) ;

[0186] In the outer loop, calculate the robust weight r_i=|Δ_i-(T_i+S_i)|; B=median(r)×6; ρ_i=B(r_i / B), where B(u) is the quadratic weight function; use ρ_i as an additional weight in the next inner loop;

[0187] Applying this decomposition method to the refrigerator data, we obtain the random fluctuation component R(t), with a mean of 0 and a standard deviation of 0.008; the seasonal component S(t), with an amplitude of 0.012 and a main period of 7 days (related to the weekend usage pattern); the trend component T(t), which increases monotonically from the initial 0 to 0.082, corresponding to the gradual degradation of refrigerator performance;

[0188] In another embodiment of the present application, a characteristic dimension sensitivity analysis is performed on the trend component T(t), and the specific calculation process is as follows:

[0189] Construct a 25×25 dimensional feature covariance matrix C=(1 / n)∑(x_i-μ)(x_i-μ)ᵀ; calculate eigenvalues ​​and eigenvectors: C*v_j=λ_j*v_j; sort the eigenvalues ​​in descending order: λ 1 ≥λ 2 ≥...≥λ 25 ; Calculate the variance contribution rate of each principal component: VCR_j=λ_j / ∑λ_i; Select the first k principal components whose cumulative contribution rate reaches 90%;

[0190] For each original feature dimension i, calculate its contribution to principal component j C_ij=|v_j[i]| / ||v_j|| 1 ;

[0191] Calculate the total contribution TC_i = ∑_j VCR_j * C_ij;

[0192] Find k nearest neighbors (k=8) for each data point x_i, calculate the reconstruction weight W, and minimize the error ε(W)=∑_i|x_i -∑_j W_ij*x_j| 2 ; The constraint condition is ∑_j W_ij=1, j∈N(i);

[0193] Compute the low-dimensional embedding Y, minimizing: Φ(Y)=∑_i|y_i-∑_j W_ij*y_j| 2 ;

[0194] Applying the above analysis to the refrigerator data, the system identified the five feature dimensions that contributed the most to the changes: the reactive power peak at the moment the compressor started (contribution rate 23.7%); the Q / P ratio during steady-state operation (contribution rate 18.5%); the reactive power change rate at the moment of starting (contribution rate 14.2%); the reactive power reduction feature at the moment the compressor stopped (contribution rate 9.8%); the reactive power fluctuation pattern during the operation cycle (contribution rate 8.3%); among them, the first two feature dimensions changed most significantly, with the reactive power peak at the moment the compressor started increasing by 14% and the Q / P ratio during steady-state operation increasing by 9%.

[0195] In another embodiment of the present application, it is necessary to process the nonlinear combination effect when identifying multiple devices running simultaneously. The specific calculation process is as follows: collect the power spectrum data of the devices when running alone and in combination; calculate the harmonic component H 1 To H 13 (1 to 13 times the fundamental frequency) amplitude; for devices i and j, calculate the harmonic interference coefficient HIJ_ij = ||H_combined-(H_i+H_j)|| 2 / ||H_i+H_j|| 2 ; Where H_combined is the harmonic vector when devices i and j are running simultaneously, and H_i and H_j are the harmonic vectors when they are running separately.

[0196] Construct the device harmonic association graph G=(V,E), where V is the device set and E is the edge set; when HIJ_ij>τ_HIJ (τ_HIJ=0.05, indicating a 5% harmonic interference threshold), add an edge between devices i and j;

[0197] Construct the association matrix W, W_ij=HIJ_ij; if (i,j)∈E, 0. Calculate the Laplace matrix L=DW, where D is the degree matrix (the diagonal elements are the sums of each row); calculate the eigenvalues ​​and eigenvectors of L, and select the eigenvectors corresponding to the k smallest non-zero eigenvalues; apply k-means clustering to these k eigenvectors to divide the devices into k groups.

[0198] In the test households, the harmonic relationship of 8 major household appliances was systematically analyzed and two obvious mutual interference groups were found: Group 1, air conditioners, televisions and computers (digital electronic equipment group, with significant harmonic interference); Group 2, refrigerators, washing machines and electric water heaters (high-power electrical appliances group, with moderate harmonic interference); independent devices, rice cookers and LED lighting (harmonic interference is negligible).

[0199] In another embodiment of the present application, for the identified mutual interference group, the system constructs a harmonic mutual interference matrix (HIM), and the specific calculation is as follows: the devices in group 1 are tested in pairs, and each combination is repeated 5 times; the operating power P_i of a single device and the combined operating power P_combined are recorded; the power sum P_ideal=P_i+P_j under ideal conditions is calculated; the actual power difference ΔP=P_combined-P_ideal is measured;

[0200] Through multivariate regression analysis, estimate the harmonic influence coefficient H_{i,j} of device i on device j: ΔP_j=H_{i,j}* P_i+ε; where ε is the error term; sparse matrix representation is used, and only significant influence coefficients (|H_{i,j}|>0.05) are stored;

[0201] Matrix update mechanism design: Set the initial update weight w=0.3;

[0202] When new measurement data is available, update the matrix element H_{i,j}^new=(1-w)*H_{i,j}^old+w* H_{i,j}^measured;

[0203] As data accumulates, the update weight is gradually reduced: w_new=0.9*w_old;

[0204] For group 1 (air conditioner, TV and computer), the harmonic interference matrix constructed by the system is:

[0205] HIM_1=[ 0.00 0.17 0.12 ]; [ 0.06 0.00 0.08 ]; [ 0.09 0.10 0.00 ];

[0206] In the matrix, H_{1,2}=0.17 means that the operation of the air conditioner will cause the measured power of the TV to increase by about 17%, and H_{3,1}=0.09 means that the operation of the computer will cause the measured power of the air conditioner to increase by about 9%.

[0207] In another embodiment of the present application, a two-stage harmonic correction strategy is implemented, and the specific calculation process is as follows:

[0208] For the device set D_k in the mutual interference group k, a linear equation system is constructed:

[0209] P_true=P_measured-HIM_k*P_true;

[0210] Where P_true is the true power vector to be solved, and P_measured is the measured power vector;

[0211] Transformation equations: (I+HIM_k)*P_true=P_measured;

[0212] Use iterative least squares method to solve: P_true^(t+1)=P_measured-HIM_k*P_true^(t);

[0213] Initial value P_true^(0)=P_measured, iterate until convergence (|P_true^(t+1)-P_true^(t)|<ε);

[0214] Inter-group correction algorithm: Calculate the interference effect of group k on group l:

[0215] I_{k→l}=∑ i属于D_k ∑_{j∈D_l} P_true(i)*H_{i,j}^group;

[0216] Where H_{i,j}^group is the inter-group harmonic influence coefficient;

[0217] For each device j in group l, apply the inter-group correction: P_true^final(j)=P_true(j)-I_{k→l}*(P_true(j) / ∑_{j'∈D_l}P_true(j'));

[0218] Corrected energy consumption reconstruction: For each device i, calculate the correction value E_correction(i)=∑_t [P_measured(i,t)-P_true(i,t)]*Δt;

[0219] Reconstruct the actual energy consumption E_true(i)=E_measured(i)-E_correction(i);

[0220] Apply energy conservation constraint verification: ∑_i E_true(i)=E_total±ε;

[0221] In the test scenario, when the air conditioner, TV and computer were running at the same time, harmonic interference caused the total power measurement to be 12% higher than the actual value. After applying the hierarchical harmonic correction, the system successfully reduced this error to 2.3%. In particular, the power measurement of the TV was corrected from the original +19% deviation to within +3%, significantly improving the accuracy of energy decomposition.

[0222] Specific implementation of multi-domain feature fusion and compression

[0223] During the implementation process, the system needs to fuse and compress the reactive power features extracted from the time domain, frequency domain, and energy domain to improve the computational efficiency and retain the most discriminative features. The specific implementation method is as follows:

[0224] For the refrigerator in the test home, the system first obtains feature sets in three domains: time domain feature set F_time (including 8 features, such as mean, variance, skewness, etc.), frequency domain feature set F_freq (including 10 features, such as frequency band energy distribution, spectrum peak position, etc.) and energy domain feature set F_energy (including 7 features, such as energy concentration, energy periodicity index, etc.). These feature sets are combined to form the initial 25-dimensional fusion feature vector F_combined.

[0225] The system implements a device category adaptive feature selection algorithm, the specific steps are as follows:

[0226] First, the system divides household appliances into three categories according to their types: resistive devices (such as electric water heaters and rice cookers); inductive devices (such as refrigerators, air conditioners, and washing machines); electronic devices (such as televisions, computers, and LED lighting);

[0227] For each type of device, the intra-class discrimination of each feature dimension is calculated as CD(f_i,c)=σ_between(f_i,c) / σ_within(f_i,c); where σ_between(f_i,c) represents the standard deviation of feature f_i between different devices in category c, and σ_within(f_i,c) represents the standard deviation of feature f_i within the same device in category c. The calculation results show that resistive devices are more sensitive to frequency domain features (the average CD value is 3.7), inductive devices are more sensitive to time domain and energy domain features (the average CD values ​​are 4.2 and 3.9 respectively), and electronic devices are sensitive to all three domain features but mainly in the frequency domain (the average CD value is 3.5).

[0228] For the refrigerator (an inductive device), the information gain ratio IGR(f_i) of each feature dimension is calculated:

[0229] IGR(f_i)=IG(f_i) / IV(f_i); where IG(f_i) is information gain, IG(f_i)=H(S)-∑_v |S_v| / |S| * H(S_v); H(S) is entropy, S_v is the subset when feature f_i takes value v, IV(f_i) is the intrinsic value, IV(f_i)=-∑_v|S_v| / |S|*log 2 (|S_v| / |S|).

[0230] For refrigerators, the system calculates the top five features as follows: reactive power peak at startup (IGR=0.351), steady-state Q / P ratio (IGR=0.324), reactive power change rate at startup (IGR=0.287), energy proportion in the 3-5 Hz frequency band (IGR=0.248), and energy periodicity index (IGR=0.231).

[0231] Device type adaptive feature selection: The system designs an adaptive selection strategy based on the device type:

[0232] F_selected(c)={f_i|IGR(f_i)≥τ_c}; where τ_c is the feature selection threshold for device type c, determined by optimization on the validation set. For inductive devices (such as refrigerators), τ_c is set to 0.15; for resistive devices, τ_c is set to 0.12; for electronic devices, τ_c is set to 0.18. For refrigerators, this strategy selects the 16 most discriminative features with a retention rate of 64%.

[0233] Discrete cosine transform (DCT) sparse representation: Apply DCT transform to the selected feature subset F_selected:

[0234] X_dct=D * X; where D is the DCT transformation matrix and X is the original eigenvector. The i,jth element of DCT is calculated as: D_i,j=α_i*cos[(π*(2j+1)*i) / (2N)];

[0235] Among them, α_i=√(1 / N) when i=0; α_i=√(2 / N) when i≠0;

[0236] After applying DCT to the refrigerator feature vector, the system found that the first 9 DCT coefficients contained 92.5% of the information.

[0237] The compression ratio CR(X)=min(5,max(20,10+5*C(X))) is dynamically determined according to the signal complexity; where C(X) is a normalized signal complexity measure, calculated based on sample entropy. For the steady-state operation mode of the refrigerator, the system determines the compression ratio to be 15:1; for the variable frequency regulation mode, the compression ratio is reduced to 8:1 to retain more details.

[0238] The OMP algorithm is used to reconstruct the original features from the compressed measurements. In the implementation, the average error of the 16-dimensional feature vector of the refrigerator after reconstruction using the OMP algorithm is 3.7%, which meets the recognition accuracy requirements.

[0239] Through the above device category adaptive feature selection and compression algorithm, the system compresses the original 25-dimensional feature vector to a dimension dynamically determined according to the device characteristics (9 dimensions for refrigerators), significantly reducing the computational complexity (4.6 times faster) while maintaining a 93.5% recognition accuracy. This method is particularly suitable for resource-constrained edge computing environments, improving system response speed while ensuring performance.

[0240] In another embodiment of the present application, when constructing the reactive power characteristic fingerprint of the device, the system adopts a steady-state-transient dual fingerprint representation method to fully capture the device characteristics. For the refrigerator test device, the specific implementation process is as follows:

[0241] The system divides the characteristic fingerprint of the refrigerator into two parts: the steady-state characteristic fingerprint F_steady consists of 16 steady-state operating parameters, including average reactive power, power factor, harmonic distribution, etc. The transient characteristic fingerprint F_transient consists of the reactive power time series during the start and stop process, and the number of sampling points is not fixed (about 300-500 points for the start sequence and about 100-200 points for the stop sequence).

[0242] Since the lengths of transient sequences vary, the system uses the dynamic time warping (DTW) algorithm for standardization. The specific implementation of the DTW algorithm is as follows: Given two time series P=(p 1 ,p 2 ,...,p_n) and Q=(q 1 ,q 2 ,...,q_m), the goal is to find a regular path W=(w 1 ,w 2 ,...,w_ K ), where w_k=(i,j) means that the i-th point of P corresponds to the j-th point of Q, so that the regular distance is minimized. First, construct an n×m distance matrix D, where D(i,j) represents the distance between points p_i and q_j D(i,j)=(p_i-q_j) 2 ;

[0243] Calculate the cumulative distance matrix M: M(1,1)=D(1,1); M(i,1)=D(i,1)+M(i-1,1), i=2,3,...,n; M(1,j)=D(1,j)+M(1,j-1), j=2,3,...,m; M(i,j)=D(i,j)+min(M(i-1,j),M(i,j-1), M(i-1,j-1)), i>1, j>1; start from M(n,m) and trace back, each time selecting the smallest value among M(i-1,j), M(i,j-1) and M(i-1,j-1), until returning to M(1,1), and obtaining the optimal regularized path W. According to the path W, resample the sequence P to a sequence P' of the same length as the standard template Q.

[0244] In the refrigerator implementation case, a typical refrigerator startup sequence (400 points long) was selected as a standard template, and all collected refrigerator startup sequences were regularized to 400 points in length.

[0245] To improve the efficiency of the DTW algorithm, the following optimizations are implemented:

[0246] Restrict the path to a band of width r around the main diagonal (r is set to min(n,m) / 10);

[0247] Use the LB_Keogh lower bound to quickly eliminate impossible matches;

[0248] First, downsample the long sequence by 4:1 for coarse matching, and then accurately match it at a fine granularity;

[0249] According to the characteristics of the refrigerator, the feature weighting strategy w_i=β_device*γ_i*δ_i is designed; where β_device is the device type weight (1.2 for inductive devices); γ_i is the feature reliability weight, γ_i=SNR_i / max(SNR), where SNR_i=μ_i 2 / σ_i 2 ; δ_i is the feature discrimination weight, calculated based on the information gain ratio. For the refrigerator, the system assigns the highest weight (1.85) to the reactive power peak feature at the startup moment and a lower weight (0.65) to the reactive power fluctuation feature.

[0250] DTW similarity calculation optimization: When calculating the similarity of transient feature sequences, the weighted DTW algorithm DTW_w(P,Q)=∑_(i,j)∈W w_phase(i,j)*D(i,j) is used; where w_phase(i,j) is the phase weight function, giving a higher weight to the key phase (such as the startup moment) w_phase(i,j)=1.5, if i,j is within the startup key phase, it is 1, otherwise it is 0.

[0251] For refrigerators, the first 80 sampling points (about 0.4 seconds) of startup are defined as the critical phase and given a 1.5-fold weight in the DTW calculation.

[0252] Fingerprint integrity verification: The constructed dual fingerprint has undergone verification testing, and the recognition accuracy is evaluated using 10-fold cross validation: using only steady-state fingerprints: recognition accuracy is 82.5%; using only transient fingerprints: recognition accuracy is 88.3%; combining dual fingerprints: recognition accuracy is 96.7%; especially in distinguishing similar devices (such as refrigerators and air-conditioning compressors), the error rate of dual fingerprints is reduced from 15% of a single fingerprint to 3.4%.

[0253] By implementing the above DTW algorithm and dual fingerprint construction method, the problem of inconsistent length of transient feature sequences was successfully solved, and the feature information of the device at different operating stages was captured, which significantly improved the accuracy and robustness of device identification. It is suitable for household appliances with obvious transient characteristics, such as refrigerators, air conditioners and washing machines, and provides a reliable foundation for subsequent equipment status monitoring and energy optimization.

[0254] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A household energy ecological intelligent optimization method driven by electric energy meter data fusion, characterized in that: The following steps are involved: Collect and preprocess multi-source power data, generate standardized power data sets, and decompose household energy in the state of multi-device overlap based on them to generate energy consumption behavior patterns; Extract and characterize reactive power features in dual time dimensions on standardized electric energy data sets to generate a reactive power feature fingerprint library for equipment; For the reactive power feature fingerprint library of the equipment, the historical-current dual model is used to dynamically update the feature fingerprint of gradual perception, generate a time-varying feature evolution model and an equipment aging feature library, and evaluate the health status of the equipment based on this, generate an equipment health score report and fault warning information; Combine energy consumption behavior patterns, equipment health score reports and fault warning information to generate a home energy ecosystem optimization plan; The steps of extracting and characterizing the reactive features of the standardized electric energy data set in dual time dimensions include: Apply dual-threshold event detection and classification to a standardized electric energy dataset to generate power event sequences and event type labels; According to the power event sequence and event type labels, an event-triggered multi-resolution sampling mechanism is implemented to generate a multi-resolution power data set, and three-domain reactive power feature extraction is performed on it to generate a reactive power multi-domain feature set; Integrate reactive power multi-domain feature sets and event type labels to build a reactive power feature fingerprint library for equipment; The steps to apply dual threshold event detection and classification to a standardized electric energy dataset include: Calculate the derivative features of power data in the standardized electric energy data set to generate a power derivative feature set; Analyze the power derivative feature set, determine the adaptive threshold parameters of high-power devices and low-power devices, and generate a dual-threshold parameter set; Combined with the standardized power data set and dual threshold parameter set, multi-feature event detection is performed to generate preliminary event detection results; Process the preliminary event detection results and the standardized electric energy data set of the corresponding time period, extract event features, and generate event feature vector sets; Analyze the event feature vector set, perform preliminary classification of events, and generate power event sequences and event type labels.

2. The method according to claim 1, characterized in that The steps to implement an event-triggered multi-resolution sampling mechanism include: According to the power event sequence, when an event trigger is detected, the sampling frequency is increased to M Hz, and the data is continuously collected for N seconds to obtain transient high-frequency sampling data; M and N are positive numbers; Based on transient high-frequency sampling data and event type labels, a multi-level dynamic adjustment of the sampling frequency is implemented, which gradually decreases from the high-frequency stage to the steady-state stage to generate multi-level sampling rate data; Extract pre-event context data from a pre-built look-ahead ring buffer; Integrate transient high-frequency sampling data, multi-level sampling rate data and pre-event context data to generate a multi-resolution power dataset.

3. The method according to claim 1, characterized in that The steps of performing three-domain reactive power feature extraction on reactive power data in a multi-resolution power data set include: Calculate the reactive power change rate and statistical characteristics, and generate a reactive power time domain feature set; Apply S-transform algorithm to analyze time-frequency characteristics and generate reactive power frequency domain feature set; Perform wavelet packet decomposition, calculate the energy distribution and time-varying characteristics of each frequency band, and generate a reactive power energy domain feature set; The reactive power time domain feature set, reactive power frequency domain feature set and reactive power energy domain feature set are integrated, and the reactive power multi-domain feature set is generated through feature dimensionality reduction and sparse representation.

4. The method according to claim 1, characterized in that The steps of dynamically self-updating feature fingerprints using the history-present dual model for gradual change perception include: Monitor the newly collected reactive power multi-domain feature set, update the initial state of the history-present dual model through an asymmetric update strategy, calculate the model difference index, and generate a model difference report; Analyze model difference reports, apply feature decoupling algorithms to separate random fluctuations and systematic changes, perform feature change attribution analysis, and generate feature change attribution results and equipment status assessment; According to the feature change attribution results and equipment status evaluation, a differentiated update strategy is implemented, the balance of the dual models is dynamically adjusted, and a time-varying feature evolution model and an equipment aging feature library are generated.

5. The method according to claim 4, characterized in that The steps of updating the initial state of the dual models through the asymmetric update strategy include: Process the feature data in the reactive power feature fingerprint library of the equipment, extract the initial feature vector, establish the feature weight matrix, and initialize the same historical model and current model for each type of equipment, including feature vector units of the same dimension, feature dimension weight units, state metadata units, and associated equipment operation mode parameter layer units; Configure the first learning rate and the second learning rate for the current model and the historical model respectively. The second learning rate is one K times the first learning rate, and K is greater than 1. When a new reactive power multi-domain feature set is obtained, the exponentially weighted moving average algorithm is applied to update the current model. When the cumulative number of updates reaches a preset threshold, the historical model is updated to generate the updated dual model state; Calculate the difference index between the updated dual model states, apply the weighted norm to calculate the distance between models, and generate a model difference report.

6. The method according to claim 4, characterized in that The steps to perform feature change attribution analysis include: Apply the seasonal-trend decomposition method to the model difference report to separate the characteristic changes into three independent components: random fluctuations, cyclical changes, and systematic trends, and generate a set of characteristic change components; For the systematic trend components, PCA is used to identify the characteristic dimensions with the greatest contribution and generate a report on the dimensions that are sensitive to changes; Integrate the change-sensitive dimension report and the pre-configured device state feature model, build a mapping relationship between feature changes and device state changes, generate change pattern mapping results, and perform a cross-validation process based on this to check whether changes in different feature dimensions support the same state hypothesis and generate feature change attribution results; Combine the feature change attribution results and change sensitivity dimension reports to evaluate the change rate and severity, predict the state evolution trend, and generate equipment status assessment.

7. The method according to claim 1, characterized in that The steps of evaluating the equipment health status based on the time-varying feature evolution model and the equipment aging feature library include: Using the equipment reactive power feature fingerprint library and equipment aging feature library, establish the Q / P ratio baseline value range under normal operating conditions for each type of equipment and generate the equipment Q / P baseline file; Analyze Q / P change data in the time-varying characteristic evolution model, monitor in the short-term, medium-term and long-term time scales, and generate Q / P trend analysis reports; Apply Q / P trend analysis reports and equipment Q / P baseline profiles to perform health rule assessments for equipment types and generate equipment health score reports; Integrate equipment health score reports and feature change attribution results, apply Bayesian network reasoning models to assess fault risks, and generate fault warning information and maintenance recommendations.

8. The method according to claim 1, characterized in that The steps to perform home energy decomposition with multiple devices overlapping include: Process the standardized electric energy data set and construct a three-layer dictionary structure consisting of a base-level dictionary, a middle-level dictionary, and a top-level dictionary; Based on the three-layer dictionary structure, we design the level-aware sparsity parameters and generate optimized sparse coding parameters. Based on the optimization of sparse coding parameters, the orthogonal matching pursuit algorithm is performed on the three-layer dictionary structure to generate a hierarchical sparse representation result; Analyze the hierarchical sparse representation results, apply the physical constraint verification algorithm to check the decomposition results, and generate a verified sparse representation; Based on the verified sparse representation, the sparse coefficients are converted into device switching states and power estimates to generate preliminary device state estimates.

Images