Power consumption management method and system for electric energy metering box

Abstract

The invention relates to the technical field of electric energy metering, and discloses a power consumption management method and system of an electric energy metering box. The method comprises the following steps: acquiring an original electric signal containing a plurality of electric parameters, and generating a steady-state and transient-state data block through filtering noise reduction, variable-point detection and time domain segmentation processing; based on the data block, obtaining a standardized feature vector through feature extraction, feature sorting and normalization processing; carrying out load identification and time alignment on the feature vector by utilizing transfer learning and confidence judgment, and generating a load type identifier with a timestamp; the load access time is further analyzed and standardized, and structured intermediate data is generated through space-time alignment and data fusion; and finally, a power consumption management report is formed through visualization processing. According to the invention, the precision and reliability of non-intrusive load monitoring can be effectively improved, fine-grained and automatic analysis and visual feedback of power consumption behaviors of users are realized, and the method has the beneficial effects of high data processing efficiency, high identification accuracy and visual and convenient management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric energy metering, in particular to a power management method and system of an electric energy metering box. BACKGROUND

[0002] In the field of electric energy metering, the existing scheme related to the power management method and system of an electric energy metering box is generally based on steady-state power data to analyze load characteristics, which has limitations such as insufficient fusion of multi-dimensional power characteristics, low utilization rate of transient data, and low confidence of load type determination. The existing methods mostly rely on single feature extraction or static threshold determination, which is prone to load misidentification and access event omission in a dynamic power environment, and it is difficult to meet the stable implementation of accurate judgment of the type and environment of the power consumer. For the joint processing of steady-state and transient power data and the correlation analysis of load identification, the existing technology generally lacks a cooperative mechanism for feature priority sorting and change point detection, making it difficult to form a consistent process in the continuous link of power data collection, feature standardization, load label generation, and management report output, resulting in a lag in the determination of the access event of the power consumer and the accumulation of errors in the judgment of the power environment. Therefore, constructing a power management method and system of an electric energy metering box around the dynamic correlation of multi-dimensional power characteristic analysis and load identification, and establishing an operable process correlation between data preprocessing, model adaptation, and event response, has become a key research direction to improve the intelligent management level of the electric energy metering box. SUMMARY

[0003] To solve the above technical problems, the present application provides a power management method of an electric energy metering box, comprising:

[0004] Obtaining an original electric signal, performing filter denoising based on FIR and IIR cascading and combining an adaptive filtering algorithm and wavelet transform, change point detection using a sliding window mechanism and a multivariate change point detection algorithm based on Bayesian inference and CUSUM, and time domain segmentation processing based on the expansion of a preset time window before and after the event time point, to generate steady-state and transient data blocks; the original electric signal includes instantaneous current value, instantaneous voltage value, active power, reactive power, and apparent power;

[0005] Based on the steady-state and transient data blocks, performing feature extraction based on an improved multi-scale convolution kernel structure CNN model, feature sorting using fast Fourier transform and an improved self-attention mechanism, and normalization processing based on minimum-maximum normalization and Z-score standardization combined with zero padding or interpolation expansion, to generate a standardized feature vector;

[0006] From the standardized feature vector, load identification is performed using transfer learning fine-tuning and freezing part of the pre-training layer weight, confidence determination based on historical confidence mean and standard deviation dynamic threshold, and time alignment processing based on L1 norm time window sliding matching and interpolation filling, to generate load type label and timestamped load identification;

[0007] Based on the timestamped load identification, time analysis, duration calculation and unit standardization processing are performed to generate standard access duration;

[0008] Based on the load type label and standard access duration, spatio-temporal alignment, data fusion and structured coding processing are performed to generate structured intermediate data;

[0009] Obtain structured intermediate data, perform visual template matching, key indicator extraction, dynamic rendering and layout optimization processing, and generate electricity management report.

[0010] Further, the steady-state transient data block includes:

[0011] The steady-state transient data block contains high-resolution electrical parameter sampling data within a continuous timestamp range, covering multi-channel electrical parameters collected by the electric energy metering box. The sampling rate, data dimension and data quality indicators are explicitly recorded in the data structure.

[0012] Further, the process of generating standardized feature vectors also includes:

[0013] Spatio-temporal feature extraction is performed using an improved convolution kernel structure, and multi-scale convolution operations are defined:

[0014]

[0015] Wherein, is the number of input channels; is the weight matrix of the th convolution kernel of the th layer; is the convolution layer; is the convolution kernel index; is the th feature map of the previous layer; is the ReLU activation function; is the output feature map of the th layer; is the bias term of the th layer; is the convolution operator;

[0016] Evaluate feature saliency:

[0017]

[0018] Wherein, is the amplitude of the th time domain feature; is the time domain feature index; is the feature time span; is the decay coefficient; is the feature saliency entropy value; is the probability distribution function; is the feature number; is the natural exponential function;

[0019] Extracting the frequency domain component, calculating the feature weight, defining the frequency domain energy proportion:

[0020]

[0021] wherein, is the th high frequency component of the convolution feature; is the full frequency band feature; is the frequency domain energy proportion; is the number of high frequency components; is the high frequency component index; is the number of full frequency band components; is the full frequency band component index; is the fast Fourier transform function; is the absolute value operator;

[0022] Calculate the attention weight:

[0023]

[0024] wherein, is the query vector; is the key vector; is the value vector; is the dimension scaling factor; is the attention weight; is the softmax function; is the transpose operator;

[0025] Through the attention mechanism aggregation, generate the feature ranking sequence output to the input sequence interface;

[0026] Perform normalization processing on the feature ranking sequence, and construct a dynamic normalization function:

[0027]

[0028] wherein, is the original ranking score; is the sliding window mean; is the window variance; is a standardized feature vector; is a scaling parameter; is an offset parameter; is a decimal constant;

[0029] After normalization, a standardized feature vector is formed.

[0030] Further, the process of generating load type labels and time-stamped load identification also includes:

[0031] The standardized feature vector is input into a pre-trained load recognition model, and a transfer learning adaptation loss function is defined:

[0032]

[0033] wherein, is a real load class label; is a class prediction probability; is an adjustable parameter matrix; is a loss function value; is the total number of load classes; is the load class index; is a regularization coefficient; is the square of the L2 norm;

[0034] The item with the maximum probability is extracted and confidence judgment is performed, and a dynamic threshold calculation is defined:

[0035]

[0036] wherein, is the historical confidence mean; is the historical confidence standard deviation; is a sensitivity coefficient; is a dynamic threshold;

[0037] After threshold filtering, load type labels are generated and passed to the input label interface;

[0038] The load type labels are time series aligned, and a time window calibration function is defined:

[0039]

[0040] wherein, is the confidence of the th label at time ; is the label index; is the previous time label state; is the calibrated time point; is the number of labels; is the time variable; For the argmin function, it represents the parameters that minimize the objective function. It is an L1 norm;

[0041] After L1 norm optimization, a timestamped payload identifier is generated.

[0042] Furthermore, the process of feature extraction based on steady-state and transient data blocks using a CNN model with an improved multi-scale convolutional kernel structure also includes:

[0043] The steady-state and transient data blocks are input into a pre-configured feature extraction and ranking model, which is based on a deep convolutional neural network architecture.

[0044] The input data is processed through multiple convolutional layers, each containing multiple convolutional kernels. The kernel parameters are obtained through pre-training, enabling the automatic learning of local time-frequency features in electrical parameters.

[0045] Convolution operations, combined with batch normalization and activation functions, achieve nonlinear mapping;

[0046] The convolutional feature extraction process specifically covers capturing stable features of steady-state data and identifying dynamic features of transient data. The model structure design takes into account the multi-scale characteristics of electrical parameters and achieves multi-level feature extraction by combining convolutional kernels of different sizes.

[0047] Furthermore, the feature ranking process using Fast Fourier Transform and an improved self-attention mechanism also includes:

[0048] The fast Fourier transform is applied to map the time-domain features to the frequency domain to obtain the frequency components and their amplitude information;

[0049] The importance weights are calculated by combining the weight parameters learned during the training phase with the attention mechanism model.

[0050] A self-attention mechanism is used to weight the frequency domain features and calculate the contribution of each frequency component to the load identification task. The weight values ​​reflect the relative importance of each feature in the classification.

[0051] Furthermore, the normalization process based on min-max normalization and Z-score standardization, combined with zero-padding or interpolation extension, also includes:

[0052] The normalization method employs a strategy that combines min-max normalization and Z-score standardization;

[0053] The normalized feature sorting sequence is mapped to a fixed-length standardized feature vector structure, the length of which is configured according to the input requirements of the load identification model.

[0054] For sequences that are too short, zero-padding or interpolation is used to extend them; for very long sequences, a feature selection strategy is used to retain the first N important features.

[0055] Furthermore, the process of fine-tuning and freezing some pre-trained layer weights using transfer learning for load identification also includes:

[0056] It is built on a deep transfer learning framework and uses a multi-layer deep neural network.

[0057] During the transfer learning adaptation phase, the model introduces a fine-tuning mechanism to adjust the weights of some network layers based on the characteristics of the current power consumption environment.

[0058] By freezing the weights of some pre-trained layers and updating the parameters only for subsequent fully connected layers and classification layers, combined with a small amount of labeled data or unsupervised adaptive strategies, the model is adjusted to adapt to the new load distribution.

[0059] Furthermore, the time alignment process based on L1 norm time window sliding matching and interpolation filling also includes:

[0060] The input load type labels are aligned using a time series alignment algorithm with a time window sliding matching algorithm.

[0061] The time series alignment module maps the input labels to a unified time axis according to a preset time resolution and uses interpolation or padding methods to handle the problem of uneven timestamp intervals.

[0062] For labels with overlapping or conflicting times, the system filters them based on confidence level priority, retaining the label with the highest confidence level.

[0063] Furthermore, an electricity metering box power management system, applied to any of the methods described above, includes:

[0064] Install the parameter configuration module to obtain and calibrate meter box component information;

[0065] The mixed signal acquisition module is used to capture steady-state and transient current and voltage waveforms;

[0066] The feature extraction and sorting module, connected to the mixed signal acquisition module, is used to generate standardized feature vectors.

[0067] The load identification and decision-making module, connected to the feature extraction and sorting module, is used to output load identifiers with timestamps.

[0068] The strategy arbitration module, connected to the load identification and decision module, is used to generate opening and closing control commands;

[0069] The instruction execution module, connected to the strategy arbitration module, is used to drive the circuit breaker mechanism to operate.

[0070] The data management module connects all functional modules and is used to store and update load characteristic data.

[0071] The key innovations of this invention include:

[0072] (1) Convolutional feature extraction is performed on steady-state and transient data blocks through a feature extraction and ranking model to generate original feature vectors for multi-dimensional electricity consumption feature analysis.

[0073] (2) Apply the pre-trained load identification model for transfer learning adaptation, extract the maximum probability term based on the initial load type probability, and generate load type labels to identify the types of electrical appliances.

[0074] (3) Using variable point detection technology combined with access duration, a load identification structure with timestamp and a multi-dimensional power management report are generated to determine the access event of electrical appliances and to make judgments on the power environment.

[0075] The following are its main beneficial effects:

[0076] (1) Through multi-dimensional electricity consumption characteristic analysis, important features can be effectively identified and extracted in the data preprocessing stage, thereby improving the accuracy of judging the type of electrical appliances and the electricity environment.

[0077] (2) The application of the load identification model makes the generation process of load type labels more accurate, optimizes the confidence judgment of load identification, and ensures the accuracy of load identification in dynamic power consumption environment.

[0078] (3) Combining the change point detection technology and the electricity management report generated by the access duration, it effectively supports the real-time judgment of electrical appliance access events, reduces false identification and missed detection, and enhances the intelligent management level of the electricity metering box. Attached Figure Description

[0079] Figure 1 A flowchart illustrating an electricity management method for an electricity metering box provided in this application embodiment;

[0080] Figure 2 This is a structural block diagram of an electricity metering box power management system provided in an embodiment of this application. Detailed Implementation

[0081] Example 1: Refer to Figure 1 This is a flowchart illustrating an electricity management method for an electricity metering box provided in an embodiment of the present invention. The process may include at least steps S100-S600:

[0082] S100: Acquire the original electrical signal, perform filtering and noise reduction based on FIR and IIR cascade combined with adaptive filtering algorithm and wavelet transform, perform change point detection using sliding window mechanism and multivariate change point detection algorithm based on Bayesian inference and CUSUM, and perform time-domain segmentation processing based on expanding the preset time window before and after the event time point to generate steady-state and transient data blocks; the original electrical signal includes instantaneous current value, instantaneous voltage value, active power, reactive power and apparent power;

[0083] S200: Based on steady-state and transient data blocks, feature extraction is performed using a CNN model with an improved multi-scale convolutional kernel structure. Feature sorting is performed using fast Fourier transform and an improved self-attention mechanism. Normalization processing based on min-max normalization and Z-score standardization combined with zero padding or interpolation expansion is performed to generate standardized feature vectors.

[0084] S300. From the standardized feature vector, load identification is performed by fine-tuning and freezing some pre-trained layer weights using transfer learning, confidence determination is performed based on dynamic thresholds of historical confidence mean and standard deviation, and time alignment processing is performed based on L1 norm time window sliding matching and interpolation filling to generate load type labels and timestamped load identifiers.

[0085] S400: Based on the timestamped load identifier, perform time parsing, duration calculation and unit standardization to generate standard access duration;

[0086] S500 performs spatiotemporal alignment, data fusion, and structured coding based on load type labels and standard access duration to generate structured intermediate data.

[0087] S600: Acquire structured intermediate data, perform visualization template matching, key indicator extraction, dynamic rendering and layout optimization, and generate electricity management reports.

[0088] Step S100 includes at least steps S110-S130:

[0089] S110. Obtain the original electrical signal, perform filtering and noise reduction processing to obtain the preprocessed electrical parameters;

[0090] This step uses the raw electrical signals collected by the current and voltage sensors built into the power metering box as the input source. The current sensor monitors current changes in the power lines, and the voltage sensor collects voltage fluctuations in the power lines. Together, they constitute a multi-dimensional electrical parameter data source. Specifically, the raw electrical signals include instantaneous current and voltage values, and further encompass electrical parameters such as active power (AP), reactive power (RP), and apparent power (AP). The electrical parameter data is continuously collected at a preset sampling frequency, forming a continuous time-series data stream. The acquisition process is controlled by the internal clock synchronization module of the metering box to ensure the accuracy of the data acquisition timestamps. Furthermore, the sampling rate and sampling accuracy of the sensor signals are dynamically adjusted according to configuration parameters to meet the sampling requirements under different power usage environments.

[0091] Furthermore, digital filtering techniques are used for signal preprocessing of the original electrical signal. This filtering includes combinations of low-pass, high-pass, and band-pass filters, specifically achieved through the cascading of finite impulse response (FIR) and infinite impulse response (IIR) filters to remove sampling noise and power grid interference, thereby improving the signal-to-noise ratio. The filtering process incorporates an adaptive filtering algorithm, dynamically adjusting the filter coefficients based on environmental noise characteristics to ensure real-time optimization of the filtering effect. After filtering, the signal undergoes detrending processing to eliminate long-term drift, and is further decomposed using wavelet transform (WT) to extract multi-scale time-frequency features, enhancing the sensitivity of subsequent change point detection.

[0092] For noise reduction, a Kalman filter is used to estimate the state of the current and voltage signals, filtering out random noise and abnormal spikes. The parameters of the Kalman filter are adaptively adjusted based on the statistical characteristics of historical data to ensure the robustness of the filtering. Abnormal data points are identified through a threshold detection mechanism. Abnormal data is temporarily stored in a buffer, triggering an anomaly recording module to generate an anomaly log for subsequent diagnosis and model training. This anomaly handling mechanism ensures the accuracy and continuity of subsequent data processing.

[0093] After filtering and noise reduction, the purified current and voltage signals, along with the calculated active and reactive power, are formatted and encapsulated to form a unified preprocessed electrical parameter data structure, including timestamps, sampled values, and data quality indicators. This preprocessed electrical parameter data structure supports multi-threaded concurrent access and asynchronous transmission, ensuring the real-time performance and integrity of data transmission. This preprocessed electrical parameter, as an output field of this step, is explicitly identified as "preprocessed electrical parameter" and passed to the "input data" interface of the next step, S120, for use in change point detection and event trigger signal generation. This preprocessed electrical parameter not only supports real-time processing in S120 but also provides a high-quality input foundation for multi-dimensional feature extraction and sorting in S200, ensuring the closed loop and consistency of the data link.

[0094] S120. Extract mutation features from preprocessed electrical parameters, perform real-time change point detection, and generate event trigger signals;

[0095] This step uses the preprocessed electrical parameters output by S110 as input data, specifically including multi-dimensional time-series signals such as current, voltage, and power after filtering and noise reduction. First, a sliding window mechanism is used to slice the data for these multi-dimensional electrical parameters. The window size and step size are set according to system configuration parameters to ensure continuous and complete data coverage. Statistical feature extraction is performed on the electrical parameter data within each sliding window, including time-domain features such as mean, variance, kurtosis, and skewness, as well as frequency-domain features obtained after Fast Fourier Transform (FFT), forming a multi-dimensional abrupt change feature vector. This abrupt change feature vector reflects the instantaneous change trend and abnormal fluctuations of the electrical parameters, providing basic data for change point detection.

[0096] Furthermore, real-time change point analysis is implemented by combining a multivariate change point detection algorithm. This algorithm includes Bayesian inference-based change point detection, the CUSUM (Cumulative Sum) method, and sliding window statistical tests, comprehensively considering the joint probability distribution of current, voltage, and power changes to determine the location of abrupt changes in the data stream. The change point detection module supports online operation and can quickly locate the time points when appliances are connected or disconnected in the data stream. To improve the robustness of change point detection, a multi-model fusion strategy is adopted, weighting and fusing the output results of multiple change point detection methods. The fusion weights are dynamically adjusted based on historical detection accuracy.

[0097] Upon detecting a change point, the system generates an event trigger signal through the event triggering logic module. This signal includes a change point timestamp, a change point type identifier, and a snapshot of relevant electrical parameters. The change point type identifier distinguishes between various event types, such as appliance connection, disconnection, and abnormal state switching. The event trigger signal is formatted into a standard data structure, supporting event logging and subsequent analysis. To prevent false triggers, the system employs a multi-level threshold judgment mechanism, combining event duration and amplitude changes to filter valid events. Abnormal events are reported through the exception handling module.

[0098] The event trigger signal in this step is output as an field, explicitly named "Event Trigger Signal," and transmitted to the "Trigger Identifier" input interface of the next step, S130. This event trigger signal not only drives the generation of steady-state and transient data blocks in S130 but also provides a crucial time anchor point for load type migration identification in S300, enabling precise data stream segmentation and semantic alignment. Simultaneously, the event trigger signal provides the necessary time reference for accurate calculation of S400 access duration, ensuring a complete capture and management loop for electricity events.

[0099] S130. Perform time-domain segmentation on the event trigger signal to generate steady-state and transient data blocks;

[0100] This step uses the event trigger signal output by S120 as input, specifically including key information such as the event timestamp and event type. Based on the event trigger signal, the system segments the continuous time series of preprocessed electrical parameters according to the timestamp, dividing the data into steady-state and transient phases. The steady-state phase is defined as the time interval where the electrical parameters are relatively stable and fluctuate little, while the transient phase is the time interval where the electrical parameters change rapidly and the waveform fluctuates violently. The time-domain segmentation uses a threshold-based interval division method, combined with change point detection results and statistical fluctuation indicators of the electrical parameters, to dynamically determine the steady-state and transient boundaries.

[0101] Specifically, for each event trigger signal, the system expands a preset time window forward and backward from the event time point, extracting electrical parameter data such as current, voltage, active power, and reactive power within the corresponding time period to form a time-series data block. The data block structure includes metadata such as timestamp range, data sampling rate, data dimensions, and data quality indicators to ensure the integrity and traceability of the data block. Steady-state data blocks are represented by continuous intervals where data fluctuations are below a set threshold, while transient data blocks correspond to intervals with significant amplitude changes. The data block partitioning results are processed by a data integrity verification module; abnormal or missing data is marked, triggering data compensation or resampling mechanisms.

[0102] Furthermore, the steady-state and transient data block structure supports the embedding of metadata, including event type labels, event trigger signal association identifiers, and data block generation time, facilitating subsequent feature extraction and input matching for the load identification model. The data block structure is stored in a standardized data format, supporting asynchronous access and batch processing to meet the requirements of efficient data transmission. This structure, as an output field of this step, is explicitly identified as a "steady-state and transient data block" and passed to the "input data" interface of the next step S210.

[0103] The steady-state and transient data blocks not only provide the basic data units for S200 multi-dimensional feature extraction and sorting, ensuring the temporal integrity and semantic accuracy of feature extraction, but also provide a basis for time segmentation for accurate calculation of S400 access duration, supporting the full lifecycle management of electricity events. Through this step, a key transformation from event triggering signals to structured time-domain data blocks is achieved, completing the temporal segmentation and data preparation of electricity events.

[0104] In summary, the technical effects of this step are as follows: By segmenting the preprocessed electrical parameters in the time domain based on the event trigger signal, a steady-state and transient data block structure is formed, which realizes accurate time-series segmentation of power consumption events. This provides a high-quality, structured input data foundation for subsequent multi-dimensional feature extraction and load identification, and ensures the continuity and consistency of the data processing link.

[0105] Step S200 includes at least steps S210-S230:

[0106] S210. Input the steady-state and transient data blocks into the feature extraction and ranking model, perform convolutional feature extraction, and obtain the original feature vector;

[0107] This step uses the steady-state transient data block output by S130 as input data, specifically including multi-dimensional electrical parameter sequences such as current, voltage, active power, and reactive power obtained by segmenting the event trigger signal in the time domain. The steady-state transient data block contains high-resolution electrical parameter sampling data within a continuous timestamp range, covering multi-channel electrical parameters collected by the power metering box. The data structure explicitly records the sampling rate, data dimensions, and data quality indicators. Specifically, the steady-state transient data block is input into a pre-configured feature extraction and ranking model. This model is based on a deep convolutional neural network (CNN) architecture and is specifically designed to process multi-dimensional time-series electrical parameter data.

[0108] During the model input phase, the data blocks are formatted and converted into tensors. The dimensions of the tensors correspond to the time step, number of channels, and feature dimensions, ensuring the temporal continuity and spatial distribution characteristics of the data are preserved. Furthermore, the input data is processed through multiple convolutional layers, each containing multiple convolutional kernels. The kernel parameters are obtained through pre-training, enabling the automatic learning of local time-frequency features in electrical parameters. Convolutional operations, combined with batch normalization (BN) and Rectified Linear Unit (ReLU) activation functions, achieve nonlinear mapping, enhancing the model's ability to express complex electrical signal patterns.

[0109] The convolutional feature extraction process specifically encompasses capturing stable features from steady-state data and identifying dynamic features from transient data. The model structure design considers the multi-scale characteristics of electrical parameters, achieving multi-level feature extraction through combinations of convolutional kernels of different sizes. Furthermore, the convolutional output undergoes dimensionality reduction and feature aggregation via pooling layers (PLs) to reduce data redundancy and enhance the expression of salient features. The pooling employs a strategy combining max pooling and average pooling, balancing the strength and overall trend of feature responses.

[0110] After convolutional feature extraction, the model converts the high-dimensional convolutional feature mapping into a one-dimensional feature vector, forming the original feature vector. This vector contains representative feature information of the steady-state and transient data blocks, reflecting the temporal and spatial distribution characteristics of electrical parameters. The original feature vector structure embeds time labels and data block identifiers to facilitate subsequent feature sorting and load identification model association processing. The original feature vector, as the output field of this step, is explicitly identified as the "original feature vector" and passed to the "input vector" interface of the next step S220 for use in frequency domain feature extraction and feature sorting sequence generation. Simultaneously, the original feature vector provides the basic input for load type migration identification in S300, running through the data processing chain of multiple modules.

[0111] S220. Extract frequency domain feature components from the original feature vector, calculate importance weights, and generate a feature ranking sequence;

[0112] This step uses the original feature vector output by S210 as input, specifically including multi-dimensional time-domain feature representation extracted through a convolutional neural network. First, for the original feature vector, a frequency domain analysis method is used for feature transformation. Specifically, a Fast Fourier Transform (FFT) is applied to map the time-domain features to the frequency domain, obtaining frequency components and their amplitude information. These frequency-domain feature components reveal the periodic variation law of electrical parameters and frequency response characteristics, reflecting the dynamic behavior of electrical loads and the operating mode of electrical appliances.

[0113] Furthermore, energy distribution, harmonic components, and phase information of key frequency bands are extracted from the frequency domain feature components to form a frequency domain feature set. This feature set includes the fundamental frequency component, even harmonics, odd harmonics, and higher-order harmonics, covering the characteristic spectral features of the electrical equipment. Combining power factor fluctuations and transient response characteristics, a multi-dimensional frequency domain feature matrix is ​​constructed to provide input for subsequent feature importance assessment.

[0114] For the aforementioned frequency domain features, importance weights are calculated. This calculation is based on weight parameters learned during the training phase and implemented using an attention mechanism model. Specifically, a self-attention mechanism is used to weight the frequency domain features, calculating the contribution of each frequency component to the load identification task. The weight values ​​reflect the relative importance of each feature in the classification. The weight calculation process includes feature mapping, weight normalization, and weight adjustment, dynamically responding to differences in the spectral characteristics of different load types.

[0115] Through the aforementioned weight calculations, a feature ranking sequence is generated. This sequence is arranged from highest to lowest feature importance, forming an ordered frequency domain feature index. The feature ranking sequence not only includes feature identifiers and corresponding weights but also embeds timestamps and data block identifiers, facilitating subsequent model calls and data tracking. The feature ranking sequence is stored in a standardized data format, supporting asynchronous access and batch processing.

[0116] The output field of this step is explicitly defined as the "feature sorting sequence," which is passed to the "input sequence" interface of the next step S230 for normalization processing and standardized feature vector generation. The feature sorting sequence provides key input to the load type migration identification module of S300, supports the dynamic adaptation of the transfer learning model to the importance of features, and runs through the multi-stage data processing flow of the power consumption management system.

[0117] S230. Normalize the feature sorting sequence to generate a standardized feature vector;

[0118] This step uses the feature ranking sequence output by S220 as input data, specifically including frequency domain feature indices and corresponding weight values ​​sorted by importance weight. For the feature ranking sequence, normalization is first performed. The normalization method combines Min-Max Normalization (MMN) and Z-score Standardization to ensure that feature weights are within a uniform dimension, eliminating the impact of differences in feature dimensions on subsequent model inference. During normalization, boundary thresholds are set for outlier weight values. Outliers are truncated or adjusted using a smoothing function to avoid extreme values ​​interfering with the model input.

[0119] Furthermore, the normalized feature ranking sequence is mapped to a fixed-length standardized feature vector structure, the length of which is configured according to the input requirements of the load identification model. For sequences with insufficient length, zero-padding (ZP) or interpolation expansion methods are used to supplement them; for extremely long sequences, a feature selection strategy is used to retain the top N important features, ensuring the semantic integrity and computational efficiency of the input features. The standardized feature vector structure contains meta-information such as feature indexes, normalized weights, timestamps, and data block identifiers, supporting the joint representation of multi-dimensional features.

[0120] The standardized feature vector structure is formatted through a data serialization module, employing a lightweight data exchange format that supports efficient transmission and model invocation. The normalization process incorporates a data integrity verification mechanism to check the continuity and consistency of the feature vectors. Abnormal data triggers logging and early warning modules for system maintenance and model optimization. The normalized feature vector structure meets the input specifications of the load identification model and is compatible with the dynamic adjustment requirements of the transfer learning framework.

[0121] The output field of this step is explicitly identified as "standardized feature vector" and passed to the "input feature" interface of the next step S310 for use by the load type migration identification model for reasoning and classification. The standardized feature vector not only serves as the core input of the S300 module, but also achieves data continuity and semantic alignment between the S200 and S300 modules, constructing a key data bridge for feature extraction and load identification in the power management system.

[0122] In summary, the technical effects of this step are as follows: By normalizing and standardizing the feature sorting sequence, a standardized feature vector structure with a unified format is formed, eliminating differences in feature dimensions, optimizing the expression of model input data, improving the input consistency and inference efficiency of the load identification model, and promoting seamless connection and collaborative operation of data among multiple modules.

[0123] In one embodiment, step S210 uses the steady-state transient data block output from S130 as input data. This data block contains filtered and denoised multi-dimensional time-series signals of current, voltage, and power. The steady-state transient data block is sliced ​​according to time windows and input into a feature extraction and ranking model. An improved convolutional kernel structure is used for spatiotemporal feature extraction. Formula ① defines the multi-scale convolution operation:

[0124]

[0125] in:

[0126] The number of input channels is derived from the multidimensional electrical parameter channels (such as current, voltage, etc.) of the steady-state and transient data blocks.

[0127] For the first Layer The weight matrix of each convolutional kernel is initialized using a pre-trained payload dataset;

[0128] It is a convolutional layer;

[0129] For convolution kernel index;

[0130] For the first layer above Each feature map originates from the active and reactive power components of the preprocessed electrical parameters;

[0131] The ReLU activation function is used to introduce nonlinearity.

[0132] For the first The output feature map of the layer represents the feature representation after the convolution operation. It is the output of this formula and serves as the input for subsequent processing.

[0133] For the first The layer's bias term is used to adjust the baseline value of the convolution output;

[0134] This is the convolution operator, representing the convolution operation.

[0135] Formula ② introduces an improved Shannon entropy assessment feature significance:

[0136]

[0137] in:

[0138] For the first The magnitude of each temporal feature is derived from the temporal activation value of the convolution output;

[0139] For time-domain feature indexing;

[0140] The characteristic time span is calculated from the timestamp range of the data blocks;

[0141] The attenuation coefficient is taken from the system configuration parameter table and is used to control the intensity of time decay.

[0142] The feature significance entropy value represents the degree of disorder or information content of the feature and is used for weighted fusion.

[0143] Let be the probability distribution function, representing The probability value;

[0144] The number of features represents the total number of temporal features, determined by the convolution output;

[0145] It is a natural exponential function used for time decay.

[0146] The current transient amplitude and voltage fluctuation rate index are generated by convolution using formula ①. Calculate the entropy value by combining the time decay factor. The original feature vector is generated by weighted fusion of the convolutional layer output and entropy value, and is consumed as the input vector of S220.

[0147] Step S220 extracts the frequency domain components from the original feature vector and calculates the feature weights using Fast Fourier Transform and an attention mechanism. Formula ③ defines the frequency domain energy proportion:

[0148]

[0149] in:

[0150] The first convolutional feature The high-frequency components originate from the high-frequency bandpass filtering result output by formula ①;

[0151] The full-band features are calculated from the full-band spectrum of the original feature vector;

[0152] Frequency domain energy percentage represents the proportion of high-frequency energy in the total frequency band energy, and is used in attention mechanisms;

[0153] The number of high-frequency components is determined by the bandpass filter settings;

[0154] Index for high-frequency components;

[0155] The number of components across the entire frequency band is determined by spectrum analysis.

[0156] Full-band component index;

[0157] This is the Fast Fourier Transform function, used to convert time-domain signals into frequency-domain signals;

[0158] This is the absolute value operator, used to calculate magnitude.

[0159] Formula ④ uses an improved attention weight calculation:

[0160]

[0161] in:

[0162] For the query vector, based on the frequency domain energy proportion Linear mapping generation;

[0163] The key vector is derived from the joint time-frequency distribution of the original feature vectors;

[0164] It is a value vector derived from the joint time-frequency distribution of the original feature vectors;

[0165] This is a dimension scaling factor, automatically configured based on the number of feature channels, used to stabilize gradients;

[0166] Attention weights represent feature importance scores and are used to generate ranked sequences.

[0167] The softmax function is used to convert scores into a probability distribution.

[0168] This is the transpose operator, used for matrix transpose.

[0169] Active power fluctuation patterns and reactive power change trend indicators are aggregated through an attention mechanism to generate a feature ranking sequence, which is then output to the input sequence interface of S230.

[0170] Step S230 performs normalization processing on the feature sorting sequence, and formula ⑤ constructs the dynamic normalization function:

[0171]

[0172] in:

[0173] The original ranking score is derived from the attention weight output of formula ④;

[0174] The mean of the sliding window is calculated from local statistics of the feature sorting sequence;

[0175] The variance is the window variance, taken from the data in the current batch being processed;

[0176] The standardized feature vector represents the normalized feature score, which is used for subsequent load identification.

[0177] This is a scaling parameter used to adjust the normalized scale;

[0178] This is the offset parameter, used to adjust the normalized offset;

[0179] It is a decimal constant used to prevent division by zero, and is usually set to a small value (such as 1e-8).

[0180] The voltage transient characteristic significance index and the current steady-state characteristic weight are normalized to form a standardized feature vector, which is then passed to the input feature interface of S310 to complete the cross-step closed loop.

[0181] Step S300 includes at least steps S310-S330:

[0182] S310. Input the standardized feature vector into the pre-trained load identification model, perform transfer learning adaptation, and obtain the initial load type probability.

[0183] This step uses the standardized feature vector output by S230 as input features, specifically including a normalized multi-dimensional frequency domain feature weight vector structure mapped to a fixed length. This standardized feature vector contains feature indices, normalized weights, and relevant timestamp information, satisfying the input specifications of the load identification model. The load identification model is a pre-trained load identification model built on a deep transfer learning (TL) framework. The model structure employs a multi-layer deep neural network, combining convolutional and fully connected layers to achieve complex pattern recognition of multi-dimensional electrical energy features. This model has been pre-trained on a previous large-scale electricity consumption dataset and possesses strong feature representation capabilities and classification generalization performance.

[0184] Specifically, the standardized feature vector is first mapped to the model's feature embedding space through the input layer. The embedding layer performs high-dimensional vector transformation on the input features to enhance their expressive power. Subsequently, the embedded vector undergoes deep feature extraction through multiple convolutional and pooling layers. The convolutional kernel parameters are initialized based on pre-trained weights, possessing prior knowledge for identifying load features. The model further implements nonlinear transformations through batch normalization (BN) and the Rectified Linear Unit (ReLU) activation function to enhance the model's adaptability to complex load signals.

[0185] In the transfer learning adaptation phase, the model introduces a fine-tuning mechanism to adjust the weights of some network layers based on the characteristics of the current electricity consumption environment. Specifically, it freezes some pre-trained layer weights and updates the parameters only for subsequent fully connected layers and classification layers. Combined with a small amount of labeled data or unsupervised adaptive strategies, the model is adjusted to adapt to the new load distribution. This process is implemented through the backpropagation algorithm, using cross-entropy loss (CEL) as the loss function, with the optimization objective being to maximize the accuracy of load category prediction.

[0186] Furthermore, the model integrates regularization techniques, including Dropout and weight decay, to prevent overfitting and improve its generalization ability under unknown power consumption environments. The standardized input feature vector undergoes layers of abstraction within the model, ultimately generating a probability distribution for multiple load types from the output layer. The probability values ​​reflect the confidence level corresponding to each load type. The output probability vector structure includes load type identifiers and corresponding probability values, along with timestamps and input feature association identifiers, facilitating subsequent data tracking and event matching.

[0187] The initial load type probability, explicitly identified as "Initial Load Type Probability," is output as an output field of this step and passed to the "Input Probability" interface of the next step, S320. This probability vector not only provides basic data for the generation of load type labels in S320 but also provides a basis for determining load status for the subsequent accurate calculation of S400 access duration, achieving close coupling between load identification and electricity event management. Through this step, the mapping transformation from standardized feature vectors to load category probabilities is completed, constructing the core reasoning link for intelligent identification of electricity loads.

[0188] S320. Extract the maximum probability term from the initial load type probability, determine the confidence threshold, and generate a load type label;

[0189] This step uses the initial load type probability output by S310 as input probability data, specifically including the probability distribution vector of multiple load categories. First, the system performs a maximum probability extraction operation on this probability vector to determine the load category with the highest probability value as the candidate load type. The load category identifier and its probability value corresponding to the maximum probability item are extracted and used as the basis for judgment. The load category identifier contains a predefined load type code, covering common appliance categories, such as lighting equipment, air conditioners, refrigerators, etc.

[0190] Furthermore, for the maximum probability value, the system performs a confidence threshold determination. The threshold is a preset confidence threshold parameter, determined based on the actual application scenario and model performance optimization. The determination logic includes comparing the maximum probability value with the threshold. If the maximum probability value is higher than the threshold, the load identification result is determined to be valid, and a corresponding load type label is generated. If the threshold is not reached, the identification result is determined to be uncertain, and the system can choose to output a default label or trigger an exception handling process, recording the low confidence event for subsequent analysis.

[0191] During the load type label generation process, the system combines timestamp information to synchronize the recognition results, ensuring that the label is aligned with the timeline of the input features. The load type label data structure includes load category code, confidence value, timestamp, and event identifier, supporting continuous tracking and management of subsequent electricity consumption events. The label generation module has a built-in log recording function to monitor abnormal situations and confidence fluctuations during the judgment process, generate diagnostic reports, and assist in model performance evaluation and optimization.

[0192] Furthermore, the system supports a dynamic threshold adjustment mechanism, automatically adjusting the confidence threshold based on historical recognition accuracy and user feedback data to improve the flexibility and adaptability of label generation. The confidence determination result and load type label are output in a standardized manner through a data interface to ensure seamless integration with subsequent modules. The load type label, as an output field of this step, is explicitly identified as "load type label" and passed to the "input label" interface of the next step S330.

[0193] The load type label not only provides direct input for the generation of timestamped load identifiers in the S330, but also provides a basis for load category determination for the S400 access duration precision calculation module, realizing semantic enrichment and time-series management of electricity events. Through this step, the conversion from probability distribution to explicit load category labels is realized, completing the semantic decoding and structured expression of load identification results.

[0194] S330. Perform time series alignment on the load type labels to generate timestamped load identifiers;

[0195] This step uses the load type label output by S320 as the input label, specifically including the load category code, confidence value, and corresponding timestamp information. First, the system performs time series alignment processing on the input load type labels to ensure the continuity and consistency of the label data in the time dimension. Time series alignment employs a time window sliding matching algorithm, sorting load labels at adjacent time points according to timestamp order, handling potential timestamp discrepancies and omissions between labels, and forming a continuous load identification time series.

[0196] Specifically, the time series alignment module maps input labels to a unified time axis according to a preset time resolution, and uses interpolation or padding methods to handle uneven timestamp intervals. For labels with overlapping or conflicting times, the system filters them based on confidence level priority, retaining the labels with the highest confidence and removing low-confidence or anomalous labels to ensure the accuracy and reliability of the time series. This process is combined with an anomaly detection mechanism to identify and mark time series anomalies, triggering log recording and subsequent review.

[0197] Furthermore, the system structures the aligned load type labels into a timestamped load identifier structure, which includes a load category code, timestamp range, confidence interval, and event identifier. This load identifier structure supports multi-dimensional data description, facilitating subsequent electricity event management and duration calculation. The load identifier structure is stored in a standardized data format, supporting asynchronous access and batch processing, meeting the electricity management system's requirement for efficient data flow.

[0198] Furthermore, the timestamped load identifier structure matches the input requirements of the S400 access duration accurate calculation module, serving as the basic data for event start and end time parsing. This structure also provides semantically rich load identification information for subsequent S500 multi-dimensional data integration and processing, achieving spatiotemporal fusion of load identification results and electricity consumption events. The system records processing logs and statistical information during the generation process to assist in model performance monitoring and system maintenance.

[0199] The output field of this step is explicitly identified as "load identifier with timestamp" and passed to the "input identifier" interface of the next step S410. Through this step, the time series normalization and structured transformation of load type labels are realized, and a key data bridge is built between load identification results and power consumption event management.

[0200] In summary, the technical effects of this step are as follows: By aligning and structuring the time series of load type labels, a load identification structure with timestamps is generated, which realizes the temporal continuity and semantic enrichment of load identification results, supports the precise management of electricity consumption events and multi-module data collaboration, and forms an effective connection between load identification and electricity consumption duration calculation.

[0201] In one embodiment, step S310 inputs the standardized feature vector into the pre-trained load identification model, and formula ⑥ defines the transfer learning adaptation loss function:

[0202]

[0203] in:

[0204] The labels are the actual load categories, derived from historical labeled datasets;

[0205] The predicted probability for each category is output by the fully connected layer of the model.

[0206] This is an adjustable parameter matrix, corresponding to the network weights in the fine-tuning part of transfer learning;

[0207] The loss function value represents the difference between the model's predictions and the true labels, and is used to optimize the model.

[0208] The total number of load categories represents the number of categories in the classification.

[0209] For load category indexing;

[0210] is the regularization coefficient, used to control the strength of L2 regularization;

[0211] It is the square of the L2 norm, used to penalize large weights.

[0212] Current waveform characteristics and power factor indices are generated through model forward propagation. The parameters are optimized using L2 regularization. The initial load type probability is output to the input probability interface of the S320 through the softmax layer.

[0213] Step S320 extracts the highest probability term and performs confidence level determination; formula ⑦ defines the dynamic threshold calculation:

[0214]

[0215] in:

[0216] The historical confidence average is statistically derived from the model validation set output.

[0217] The historical confidence level standard deviation is calculated based on the model's volatility in the test environment.

[0218] This is a sensitivity coefficient, taken from the system policy table, used to adjust the strictness of the threshold.

[0219] This is a dynamic threshold used for confidence determination; predictions exceeding this threshold are accepted.

[0220] The voltage transient identification results and current steady-state classification indicators are filtered by threshold to generate load type labels, which are then transmitted to the input label interface of S330.

[0221] Step S330 performs time series alignment on the load type labels, and formula ⑧ defines the time window calibration function:

[0222]

[0223] in:

[0224] For the first A tag in time The confidence level is derived from the output probability of S320;

[0225] For tag indexing;

[0226] The label state from the previous time step is taken from the time series cache.

[0227] The calibrated time point represents the optimized timestamp, used to align the load identifier;

[0228] The number of labels represents the total number of labels in the time series.

[0229] This is a time variable, representing the time point to be optimized;

[0230] For the argmin function, it represents the parameters that minimize the objective function.

[0231] It is the L1 norm, used to calculate the sum of absolute differences.

[0232] The active power duration and reactive power fluctuation cycle indicators are optimized using the L1 norm to generate a timestamped load identifier, which is then output to the input identifier interface of S410 to form a cross-module closed loop.

[0233] This section summarizes the technical effects: Through transfer learning adaptation and dynamic confidence determination, high-precision identification of load types was achieved; combined with time series calibration optimization, a load identification structure with precise timestamps was constructed, providing a time-series alignment basis for subsequent power consumption event management.

[0234] Step S400 includes at least steps S410-S430:

[0235] S410. Obtain the load identifier with timestamp, parse the event start and end times, and obtain the access time pair and the input time pair;

[0236] This step uses the timestamped load identifier output by S330 as the input identifier, specifically including information such as load category code, timestamp range, confidence interval, and event identifier. The timestamped load identifier structure reflects the distribution of load type labels over time, encompassing the mapping relationship between load identification results and corresponding time intervals. Specifically, the system first parses the input timestamped load identifier, extracting the timestamp range field. This field clearly indicates the start and end times of the load event, using a unified high-precision time format and supporting millisecond-level timestamp representation.

[0237] Furthermore, the event start and end time parsing module, based on the timestamp range and combined with the event identifier of the load identifier, performs time boundary verification and merging processing. Specifically, this includes merging load identifiers for adjacent or overlapping time periods to avoid redundant time interval divisions; and marking abnormal timestamps, such as those with start times later than end times, and triggering log recording. This time parsing process, combined with the system-configured time tolerance parameters, handles timestamp deviations caused by clock drift or data transmission delays, ensuring the rationality and accuracy of event time boundaries.

[0238] Furthermore, the system filters valid event time periods based on the confidence range of load identifiers, excluding time segments with confidence levels below a preset threshold to reduce the impact of false positives. Event start and end time parsing also includes the splitting and reorganization of continuous load events, classifying them into independent electricity consumption events based on load category changes and time interval thresholds. The parsing results form a standardized access time pair data structure, including start time, end time, and a unique event identifier, supporting subsequent duration calculations and electricity consumption event management.

[0239] The access time pair from this step is explicitly identified as an "access time pair" as an output field and passed to the "input time pair" interface of the next step, S420. This access time pair not only provides basic time information for calculating the duration of access in S420, but also provides time dimension support for spatiotemporal alignment in the multidimensional data integration and processing of S500, achieving time synchronization between load identification and electricity event management. Through this step, the conversion of timestamped load identifiers to specific electricity event time intervals is completed, constructing the basic data unit for electricity event time management.

[0240] S420: Extract the time difference from the access time pair, perform second-level precision calculation, and generate the access duration;

[0241] This step uses the access time pair output by S410 as the input time pair, specifically including the start and end times of the electricity event. Both use a unified time format and have precision down to the second level or even higher. The system first verifies the time format of the input access time pair to ensure the validity and logical correctness of the start and end times. Subsequently, the duration calculation module calculates the access duration in seconds for each electricity event by subtracting the start time from the end time, based on the timestamp difference calculation method.

[0242] Specifically, the time difference calculation uses a high-precision time calculation library, supporting time spans across days, months, and even years to prevent time overflow or error accumulation. To ensure calculation accuracy, the system considers leap seconds adjustment and time zone differences, uniformly using Coordinated Universal Time (UTC) for time conversion. During the calculation process, abnormal time pairs, such as negative or excessively long durations, are detected and marked, triggering anomaly handling procedures, including logging and reporting of abnormal events.

[0243] Furthermore, for multiple consecutive access time pairs, the system supports batch calculation and cumulative duration statistics, enabling the merging of time differences between consecutive power consumption events to form the total access duration. This process incorporates an event interval threshold to determine whether they belong to the same continuous power consumption event, supporting flexible time merging strategies. The access duration data structure includes event identifiers, calculation durations, and time unit information, supporting subsequent data processing and management.

[0244] The access duration in this step is explicitly identified as "Access Duration" as an output field and passed to the "Input Duration" interface of the next step, S430. This access duration not only provides raw duration data for the unit standardization processing of S430 but also provides a quantitative basis for the duration dimension in the multi-dimensional data integration processing of S500, promoting refined management of electricity consumption events. This step achieves the quantitative conversion of electricity consumption event time intervals into duration, establishing a key link in electricity consumption duration calculation.

[0245] S430. Standardize the unit of access duration to generate a standard access duration;

[0246] This step uses the access duration output by S420 as the input duration, specifically the duration of the electricity event in seconds. The system first verifies the validity of the input access duration, including non-negativity checks and numerical range verification, eliminating abnormal or invalid duration data. Subsequently, the unit standardization module converts the second-level duration to a minute-level standard duration. The conversion method uses integer division or rounding strategies, and the conversion precision is determined based on system configuration parameters to ensure the uniformity and applicability of the duration representation.

[0247] Specifically, the standardized processing combines a unified standard for time units, dividing seconds by 60 to obtain minutes, retaining the integer part and necessary decimal places to form a standard access duration data structure. This structure includes the duration value, time unit identifier, and event-corresponding identifier, supporting multi-dimensional data association and subsequent processing. For extremely short-duration events, the system sets a minimum duration threshold; events below the threshold can be merged or marked as instantaneous events according to a strategy, assisting in electricity consumption event analysis.

[0248] Furthermore, the unit standardization process, combined with data format conversion and serialization, generates a duration data structure that conforms to the system's data exchange standards, supporting asynchronous access and batch processing. The standard access duration data structure can embed metadata, such as data source, calculation time, and processing status, facilitating system maintenance and data tracking. Abnormal duration data triggers logging and early warning modules to assist in system diagnosis and model optimization.

[0249] The standard access duration in this step is explicitly identified as "Standard Access Duration" as an output field and passed to the "Input Duration Data" interface of the next step, S510. This standard access duration not only provides a unified quantitative indicator for the multi-dimensional data integration and processing of S510, but also provides basic data for the statistical analysis of electricity event durations in the entire electricity management system. Through this step, the unit of measurement and format standardization of electricity event duration are achieved, completing the standardized expression of electricity duration data.

[0250] In summary, the technical effects of this step are as follows: By converting and standardizing the unit of access duration, a unified minute-level standard duration structure is formed, the duration representation format of electricity consumption events is standardized, the availability and compatibility of data are improved, and the efficient integration and management of electricity consumption event duration and load identification results are promoted.

[0251] Step S500 includes at least steps S510-S530:

[0252] S510: Obtain the load type label and standard access duration, perform spatiotemporal dimension alignment and matching, and obtain the spatiotemporal related data group;

[0253] This step uses the standard access duration output by S430 and the timestamped load identifier output by S330 as input duration data and input identifier. Specifically, it includes a standardized minute-level duration data structure and a structured load category timestamped identifier. The load type label covers the load category code and its corresponding timestamp range, and the standard access duration reflects the electricity consumption duration within the corresponding time period. First, the system performs spatiotemporal alignment matching on the input load type label and standard access duration based on the timestamp information. This spatiotemporal alignment matching refers to synchronously verifying and associating the time interval of the load category identifier with the time range of the duration data, forming a multidimensional association between time and load category. Specifically, the system extracts the start and end times of the load identifier and the time interval of the duration through the timestamp parsing module, uses a time axis mapping algorithm to determine the overlap of the corresponding time periods, and filters out load label and duration data pairs with overlapping or continuous time intervals.

[0254] Furthermore, the system incorporates spatial dimension information. If the load type label contains spatial identifiers (such as meter box number, electricity area code, etc.), it performs spatial matching between the load label and the access duration based on the spatial identifiers to ensure spatial consistency of the associated data groups. Spatial matching is achieved through a spatial indexing mechanism. The spatial index contains preset metadata such as the geographical location of the meter box and the division of power distribution areas, supporting multi-level spatial matching strategies. For missing or incomplete spatial information, the system supplements it using a default spatial configuration or a spatial inference algorithm based on historical data.

[0255] During the spatiotemporal matching process, the system performs data integrity checks to detect any time mismatches, missing tags, or abnormal durations in the associated data groups. Abnormal data is marked by the anomaly handling module, generating an anomaly log for subsequent data correction and system maintenance. Anomaly handling includes methods such as timestamp resampling, duration correction, and tag completion to ensure the quality and consistency of the associated data groups. The system supports batch processing, employing a distributed computing framework to achieve efficient spatiotemporal alignment operations for large-scale electricity consumption event data.

[0256] The spatiotemporal correlated data group data structure includes fields such as load category code, timestamp range, standard access duration, spatial identifier, and data quality indicators, supporting the unified expression of multi-dimensional information. This data structure uses a standardized format for serialized storage, supporting asynchronous access and parallel processing. As an output field of this step, the spatiotemporal correlated data group is explicitly identified as "spatiotemporal correlated data group" and passed to the "input data group" interface of the next step, S520, for use in multi-dimensional data fusion and report base dataset generation. The spatiotemporal correlated data group not only provides basic data units for S520's multi-dimensional data fusion but also provides semantically rich electricity event correlation information for the S600 intelligent report visualization generation module, running through the spatiotemporal data link of the electricity management system.

[0257] S520. Extract key management dimensions from spatiotemporal correlated data groups, perform multi-dimensional data fusion, and generate the basic dataset for the report.

[0258] This step uses the spatiotemporal correlation data set output by S510 as the input data set, specifically including a multi-dimensional attribute set such as load type code, timestamp range, standard access duration, and spatial identifier. First, the system extracts key management dimensions from the spatiotemporal correlation data set. These key management dimensions include, but are not limited to, load category, electricity consumption time period, electricity consumption duration, spatial distribution characteristics, and data quality indicators. The extraction process is implemented through a dimension filtering module, which dynamically selects relevant dimension fields based on preset management objectives and analysis requirements, eliminating redundant or invalid data to form a dimension-optimized data subset.

[0259] Furthermore, for the aforementioned key management dimensions, the system performs multi-dimensional data fusion processing, encompassing operations such as data cleaning, data completion, data reduction, and data association. Data cleaning includes handling missing and outlier values, and using interpolation, statistical correction, and machine learning prediction methods to complete the data. Data reduction is achieved through feature selection and dimensionality reduction algorithms to reduce data dimensions and improve subsequent processing efficiency. Data association, based on spatiotemporal matching results and the logical relationship between load type and electricity consumption duration, constructs a multi-dimensional association matrix to reveal the inherent connections in electricity consumption behavior.

[0260] During the multidimensional data fusion process, the system incorporates time series analysis technology to extract features based on the continuity and periodicity of electricity consumption time periods, forming statistical indicators for the time dimension, such as peak periods, average load duration, and electricity consumption frequency. Spatial dimension fusion utilizes Geographic Information System (GIS) integration technology, combining spatial coordinates and regional divisions to achieve spatial aggregation and distribution feature analysis of electricity consumption data. Load category dimension fusion incorporates a hierarchical category structure, supporting grouped statistics and hierarchical management of different load categories.

[0261] Furthermore, the system performs data quality assessment on the fused multidimensional data, generating a data quality report based on indicators such as completeness, consistency, and accuracy. Data quality anomalies are recorded through the log module, triggering a data correction process. The fusion results are converted into a unified format by the data serialization module, supporting efficient use in subsequent processing. The multidimensional data fusion process supports both batch and real-time operation modes to adapt to the needs of different application scenarios.

[0262] The fused multidimensional data set is defined as the report foundation dataset, containing information such as the fused load category distribution, electricity consumption duration statistics, spatiotemporal distribution characteristics, and data quality indicators. This dataset is stored in a structured data format and supports cross-module calls and asynchronous access. As an output field of this step, the report foundation dataset is explicitly identified as "report foundation dataset" and passed to the "input dataset" interface of the next step, S530, for structured encoding and intelligent report generation. This report foundation dataset not only provides the data foundation for the structured encoding of S530 but also provides multi-dimensional management information support for the S600 intelligent report visualization module, forming the core data asset of the electricity management system.

[0263] S530. Perform structured coding on the report's basic dataset to generate structured intermediate data;

[0264] This step uses the report dataset output by the S520 as the input dataset, specifically including multi-dimensional structured information such as the fused load category distribution, electricity consumption duration statistics, spatiotemporal distribution characteristics, and data quality indicators. First, the system performs structured encoding processing on the input report dataset. This structured encoding refers to converting multi-dimensional data into a unified semantic web format, specifically using JSON-LD (JavaScript Object Notation for LinkedData) format. This format supports semantic description and linking of data, facilitating data scalability, interoperability, and machine readability.

[0265] The structured encoding process is implemented by the encoding engine module. First, semantic annotation is performed on each field in the report's base dataset, referencing predefined ontology vocabularies and namespaces to clarify the semantic meaning and data type of each data item. Then, according to the JSON-LD specification, the annotated data is converted into a JSON structure conforming to the Semantic Web standard, containing contextual information, data entities, attributes, and relationship descriptions. This encoding process supports nested structures and multi-level data representation, fully reflecting the complex relationships of the multidimensional dataset.

[0266] Furthermore, the encoding engine performs format validation on the structured data to ensure that the generated JSON-LD documents conform to syntactic specifications and semantic consistency. Validation includes field integrity, data type matching, and semantic coherence; abnormal data triggers error logging to assist system maintenance. The encoding process incorporates data compression and encryption technologies to support secure data transmission and storage, meeting the security and compliance requirements of the electricity management system.

[0267] The structured intermediate data supports asynchronous access and distributed storage, and uses standard API interfaces for data calls, facilitating integration with other system modules. This data structure not only contains multi-dimensional information on electricity load and duration, but also embeds data quality indicators and generation timestamps, enhancing data traceability and reliability. The structured encoding also supports version control mechanisms, recording data change history to meet the needs of system upgrades and model iterations.

[0268] The structured intermediate data in this step is explicitly identified as "structured intermediate data" as an output field and passed to the "input structured data" interface of the next step S610 for intelligent report visualization generation. This structured intermediate data enables the semantic expression and standardized storage of multi-dimensional electricity management information, constructing a foundational platform for data sharing and intelligent analysis in the electricity management system, and permeating the data integration and visualization process.

[0269] In summary, the technical effects of this step are as follows: By semantically and structurally encoding the basic dataset of the report, an intermediate data structure conforming to the JSON-LD specification is generated, which realizes the standardized expression and efficient storage of multi-dimensional electricity management information, promotes data interoperability between systems and the accurate generation of intelligent reports.

[0270] Step S600 includes at least steps S610-S630:

[0271] S610. Obtain structured intermediate data, perform visualization template matching, and obtain chart mapping relationships;

[0272] This step uses the structured intermediate data output by the S530 as input structured data, specifically including multi-dimensional electricity management information encoded in JSON-LD (JavaScript Object Notation for Linked Data) format, covering multi-level structured data such as load category distribution, electricity consumption duration statistics, spatiotemporal distribution characteristics, and data quality indicators. The structured intermediate data contains rich semantic annotations and contextual information, supporting semantic association and machine readability of multi-dimensional data. First, the system loads a predefined set of visualization templates through the visualization template library access module. This template set includes various chart types and layout configurations, covering diverse presentation formats such as bar charts, line charts, pie charts, heatmaps, and geographic information maps. The template design is optimized based on data expression needs and user interaction experience.

[0273] Specifically, the aforementioned structured intermediate data is used as input. A semantic parsing engine parses the multidimensional data in JSON-LD format, extracting key management dimensions and their attribute values. These are then filtered using data quality indicators to form a candidate dataset for visualization data mapping. The semantic parsing engine identifies data entities, attributes, and their relationships based on preset ontology vocabulary and context rules, supporting the expansion and mapping of multi-level nested structures. Further, the system executes a template matching algorithm, comparing the candidate dataset with template features in the visualization template library. Based on data type, number of dimensions, and attribute features, the system selects the template with the highest matching degree. The template matching process comprehensively considers data dimension compatibility, chart expressiveness, and user customization preferences, supporting dynamic adjustment of weight parameters to adapt to different electricity management scenarios.

[0274] During template matching, the system triggers an anomaly detection mechanism to mark missing fields or abnormal values ​​in the structured intermediate data. Based on the template's fault tolerance capabilities, it selects a template version that supports default value substitution or automatic completion. The matching results form a chart mapping relationship data structure, including the selected template identifier, data field mapping rules, visual attribute configurations, and interactive behavior definitions. This chart mapping relationship supports multi-level mapping, enabling the mapping of a single data field to multiple chart elements, achieving multi-perspective display of complex data. The mapping relationship data structure is stored in a standardized format, supporting asynchronous calls and cross-platform transmission.

[0275] Furthermore, the system combines user permissions and role information to enforce access control policies, filtering data mapping items that do not meet permissions to ensure data security and privacy protection. The mapping relationship generation process records detailed logs, including template matching scores, anomaly handling records, and user operation trajectories, facilitating subsequent auditing and system optimization. The chart mapping relationship, as an output field of this step, is explicitly identified as "Chart Mapping Relationship" and passed to the "Input Mapping Relationship" interface of the next step, S620, for dynamic rendering processing. This chart mapping relationship not only provides the basic configuration for the generation of visualization components in S620 but also provides data support for layout optimization in S630, permeating the initial stage of intelligent report visualization generation.

[0276] S620: Extract key indicators from the chart mapping relationship, perform dynamic rendering processing, and generate a set of visualization components;

[0277] This step uses the chart mapping relationship output by S610 as the input mapping relationship, specifically including the selected visualization template identifier, data field mapping rules, visual attribute configuration, and interactive behavior definition. First, the system uses the data binding module to bind key indicators from the structured intermediate data to corresponding chart elements according to the mapping rules. These key indicators cover multi-dimensional information such as load category statistics, electricity consumption duration, spatiotemporal distribution characteristics, and data quality indicators. The data binding process combines data type checking and format conversion to ensure that the indicator data matches the attribute type of the chart elements, supporting multiple data formats such as numerical, categorical, time series, and geographic coordinates.

[0278] Furthermore, the system invokes the graphics rendering engine to perform dynamic rendering based on the visualization template and bound data. The rendering process includes drawing chart elements, color mapping, setting axis scales, generating legends, and configuring interactive controls, employing vector graphics technology to achieve high-resolution display and scaling adaptation. The rendering engine supports responsive design, dynamically adjusting chart size and element spacing according to the display device resolution and user interface layout. Dynamic rendering also incorporates animation effects and transition processing to enhance the user experience, supporting real-time updates of chart elements and data refresh.

[0279] During rendering, the system executes an exception data handling mechanism, specially marking missing or outlier values ​​in the bound data and using visual processing methods such as grayscale display, tooltip text, or data interpolation to ensure the integrity and readability of the charts. The interaction behavior definition module loads user interaction configurations to implement functions such as mouse hover tooltips, click filtering, zooming and panning, and linked responses for chart elements. The system supports multi-chart linkage; when a user operates on a single chart, related charts are triggered to update synchronously, improving the continuity of data analysis.

[0280] The dynamically rendered visualization component set contains instantiated objects of multiple chart components and their attribute configurations, supporting component combination, nesting, and hierarchical management. The component set adopts a modular design, supporting asynchronous loading and on-demand rendering to improve system performance and response speed. Metadata, such as generation time, data version, and user-customized parameters, is embedded in the component set's data structure for easy subsequent management and maintenance. The system monitors performance metrics during the rendering process in real time, recording rendering time, memory usage, and error logs to assist in system optimization.

[0281] The visualization component set in this step is explicitly identified as "Visualization Component Set" as an output field and passed to the "Input Component Set" interface of the next step S630 for layout optimization and final report structure generation. This visualization component set not only provides S630 with rich chart resources but also lays the foundation for multi-dimensional data display in the electricity management report, running throughout the dynamic generation stage of the intelligent report.

[0282] S630: Optimize the layout of the visualization component set and generate an electricity management report;

[0283] This step uses the visualization component set output by S620 as the input component set, specifically including instantiated objects of multiple chart components and their attribute configurations. First, the system optimizes the overall layout of the visualization component set through the layout management module. Based on preset layout templates and user interface design specifications, the layout management module automatically adjusts the arrangement order, size ratio, and spacing distribution of components, taking into account their size, priority, and interaction relationships. The layout optimization employs an algorithm combining a grid system and responsive design, supporting multi-terminal adaptation, including desktop, mobile, and tablet devices.

[0284] Specifically, the layout management module analyzes the spatial dependencies and visual hierarchy between components, performs conflict detection and spatial rearrangement to avoid component overlap and visual obstruction. The system integrates user-defined layout preferences, supporting drag-and-drop adjustments, partition management, and multi-page switching to enhance report personalization and interactivity. During layout optimization, the system schedules the loading order and rendering priority of components to ensure the priority display of critical data and a smooth user experience.

[0285] Furthermore, the system integrates auxiliary information such as text descriptions, titles, annotations, and legends to form a structured electricity management report document. The report structure includes modules such as a cover, table of contents, data summary, chart display area, and appendix, supporting multi-level navigation and content indexing. The report generation module, combined with a template engine, enables dynamic content filling and formatting, supporting multiple output formats, including HTML, PDF, and interactive web reports. Text content and chart components are synchronized through data binding, ensuring consistency between the report content and the underlying data.

[0286] This step combines access control and version control to generate the final electricity management report structure. The report structure embeds the generation time, version number, and user identifier, supporting auditing and tracking. The system records layout adjustment logs and user operation history to facilitate subsequent report updates and maintenance. The electricity management report, as an output field of this step, is explicitly identified as "Electricity Management Report" and passed to the "Configuration Parameters" interface of the next step S100 for system configuration and use in the next data collection cycle, achieving a closed loop between data processing and report generation.

[0287] In summary, the technical effects of this step are as follows: By automatically optimizing the layout and structuring the visualization component set, a final electricity management report structure that meets the needs of multi-terminal adaptation and user interaction is generated. This enables intuitive display of multi-dimensional data and efficient data transfer between systems, promoting the intelligent and closed-loop operation of the electricity management system.

[0288] Example 2: Figure 2 This diagram illustrates a structural block diagram of an electricity metering box-based electricity management system according to an embodiment of the present invention. Figure 2 As shown, the structure may include:

[0289] The installation parameter configuration module 01 is used to acquire and calibrate information about the metering box components. Specifically, it receives hardware component information from inside the electricity metering box and uses the preset calibration protocol and parameter table in the installation parameter configuration module to identify and calibrate each component of the metering box. The calibration process includes reading and recording the identification codes, model specifications, and calibration parameters of key components such as current sensors, voltage sensors, and circuit breakers to ensure the consistency of the status information of each component with the system configuration. The calibration results are stored in the parameter configuration database in structured data form and transmitted to the mixed signal acquisition module for subsequent use.

[0290] The mixed-signal acquisition module 02 is used to capture steady-state and transient current and voltage waveforms. Specifically, it receives component calibration information from the installation parameter configuration module and, in conjunction with the current and voltage sensors built into the power metering box, acquires current and voltage waveform data from the power lines in real time. The acquisition module uses a multi-channel data acquisition card to capture steady-state and transient signals in segments. It utilizes an embedded filtering algorithm to preprocess the acquired signals, removing noise and interference to ensure the accuracy and integrity of the waveform data. The acquired waveform data is stored in time-series format and transmitted to the feature extraction and sorting module for further analysis.

[0291] The feature extraction and sorting module 03, connected to the mixed signal acquisition module, is used to generate standardized feature vectors. Specifically, it receives steady-state and transient current and voltage waveform data from the mixed signal acquisition module and, combined with a preset feature extraction algorithm, extracts and sorts time-domain and frequency-domain features. The feature extraction process includes Fourier transform, power spectrum analysis, and wavelet transform of the waveform data to obtain multi-scale feature vectors. The extracted feature vectors are standardized to form a standardized feature vector data structure, ensuring the consistency and comparability of the feature data. This data structure is then transmitted to the load identification and decision module for load identification.

[0292] The load identification decision module 04, connected to the feature extraction and sorting module, is used to output load identifiers with timestamps. Specifically, it receives standardized feature vectors from the feature extraction and sorting module, combines them with various classification models in the load identification algorithm library, and performs load type identification and classification. The identification process determines the load type of the current electrical equipment based on pattern matching and similarity calculation of the feature vectors, and generates a load identifier with a timestamp. The load identifier result includes load category, identification confidence level, and timestamp information, ensuring the accuracy and timeliness of the identification results. The load identifier is then passed to the strategy arbitration module for decision processing.

[0293] The strategy arbitration module 05, connected to the load identification and decision module, is used to generate circuit breaker opening and closing control commands. Specifically, based on the timestamped load identifier from the load identification and decision module, and combined with predefined power consumption strategies and safety rules, the circuit breaker opening and closing decision logic is executed. The decision-making process includes a comprehensive judgment of load type, power consumption duration, and safety thresholds to generate corresponding circuit breaker opening and closing control commands. The control commands include action type, execution time, and priority information to ensure the timeliness and effectiveness of the commands. The generated control commands are output to the command execution module to drive the actual operation of the circuit breaker.

[0294] The instruction execution module 06, connected to the strategy arbitration module, is used to drive the circuit breaker mechanism. Specifically, it receives opening and closing control commands from the strategy arbitration module and converts the control signals into circuit breaker operating actions through the instruction parsing and execution interface. The execution module, combined with the circuit breaker's action feedback mechanism, monitors the execution status and effect of the operation in real time to ensure the accuracy of the circuit breaker's actions. The execution result is sent back to the data management module in the form of status confirmation information, recording the operation log and updating the system status.

[0295] Data management module 07 connects all functional modules and is used to store and update load characteristic data. Specifically, it receives status confirmation information from the instruction execution module and load characteristic data from other modules, and performs data storage, update, and management operations. The data management module maintains a complete record of electricity consumption events and the historical trajectory of load characteristic data through a distributed database system. The data update process incorporates version control and indexing mechanisms to ensure data consistency and traceability. The updated load characteristic data is provided to the system's analysis and optimization module, supporting continuous improvement of electricity management strategies and enhancement of system performance.

Claims

1. A method for managing electricity consumption using an electricity metering box, characterized in that, include: The original electrical signal is acquired and subjected to filtering and noise reduction based on FIR and IIR cascade combined with adaptive filtering algorithm and wavelet transform; change point detection is performed using sliding window mechanism and multivariate change point detection algorithm based on Bayesian inference and CUSUM; and time-domain segmentation processing is performed based on expanding a preset time window before and after the event time point to generate steady-state and transient data blocks; the original electrical signal includes instantaneous current value, instantaneous voltage value, active power, reactive power and apparent power; Based on steady-state and transient data blocks, feature extraction is performed using a CNN model with an improved multi-scale convolutional kernel structure. Feature sorting is performed using fast Fourier transform and an improved self-attention mechanism. Normalization processing based on min-max normalization and Z-score standardization combined with zero padding or interpolation expansion is performed to generate standardized feature vectors. From the standardized feature vector, load identification is performed by fine-tuning and freezing some pre-trained layer weights using transfer learning, confidence determination is performed based on dynamic thresholds of historical confidence mean and standard deviation, and time alignment is performed based on L1 norm time window sliding matching and interpolation filling to generate load type labels and timestamped load identifiers. Based on the timestamped load identifier, time parsing, duration calculation and unit standardization are performed to generate standard access duration; Based on load type labels and standard access duration, spatiotemporal alignment, data fusion, and structured encoding are performed to generate structured intermediate data. Obtain structured intermediate data, perform visualization template matching, key indicator extraction, dynamic rendering and layout optimization, and generate electricity management reports.

2. The method according to claim 1, characterized in that, The steady-state and transient data blocks include: The steady-state and transient data blocks contain high-resolution electrical parameter sampling data within a continuous timestamp range, covering multi-channel electrical parameters collected by the power metering box. The data structure clearly records the sampling rate, data dimension, and data quality indicators.

3. The method according to claim 1, characterized in that, The process of generating standardized feature vectors also includes: An improved convolutional kernel structure is used for spatiotemporal feature extraction, and a multi-scale convolution operation is defined: in, Input the number of channels; For the first Layer The weight matrix of each convolutional kernel; It is a convolutional layer; For convolution kernel index; For the first layer above Each feature mapping; It is the ReLU activation function; For the first The layer's output feature mapping; For the first Layer bias terms; For convolution operators; Assess the significance of the features: in, For the first The amplitude of each time-domain feature; For time-domain feature indexing; The characteristic time span; The attenuation coefficient; The entropy value represents the significance of the feature. It is the probability distribution function; The number of features; It is a natural exponential function; Extract frequency domain components, calculate feature weights, and define frequency domain energy proportions: in, The first convolutional feature One high-frequency component; It features a full-band frequency response; Frequency domain energy percentage; The number of high-frequency components; Index for high-frequency components; The number of components across the entire frequency band; Full-band component index; It is the Fast Fourier Transform function; This is the absolute value operator; Calculate attention weights: in, For query vector; The key vector; It is a value vector; This is the dimension scaling factor; Attention weights; It is the softmax function; It is the transpose operator; The feature ranking sequence is generated by aggregation through an attention mechanism and output to the input sequence interface; Normalization is performed on the feature-sorted sequence to construct a dynamic normalization function: in, The original sort score; The mean of the sliding window; For window variance; The standardized feature vector; For scaling parameters; For offset parameters; It is a decimal constant; After normalization, a standardized feature vector is formed.

4. The method according to claim 1, characterized in that, The process of generating load type labels and timestamped load identifiers also includes: Input the standardized feature vectors into the pre-trained load identification model and define the transfer learning adaptation loss function: in, Labels for actual load categories; Predict probabilities for each category; This is an adjustable parameter matrix; The value of the loss function; Total number of load categories; Index for load categories; The regularization coefficient is used. The square of the L2 norm; Extract the term with the highest probability and perform confidence assessment, then define the dynamic threshold calculation: in, This represents the historical confidence level average. The standard deviation of the historical confidence level; This is the sensitivity coefficient; For dynamic thresholds; After threshold filtering, load type labels are generated and transmitted to the input label interface; Perform time series alignment on load type labels and define a time window calibration function: in, For the first A tag in time Confidence level; For tag indexing; The label status at the previous moment; This is the calibrated time point; Number of tags; It is a time variable; For the argmin function, it represents the parameters that minimize the objective function. It is an L1 norm; After L1 norm optimization, a timestamped payload identifier is generated.

5. The method according to claim 1, characterized in that, The process of feature extraction based on steady-state and transient data blocks and using a CNN model with an improved multi-scale convolutional kernel structure also includes: The steady-state and transient data blocks are input into a pre-configured feature extraction and ranking model, which is based on a deep convolutional neural network architecture. The input data is processed through multiple convolutional layers, each containing multiple convolutional kernels. The kernel parameters are obtained through pre-training, enabling the automatic learning of local time-frequency features in electrical parameters. Convolution operations, combined with batch normalization and activation functions, achieve nonlinear mapping; The convolutional feature extraction process specifically covers capturing stable features of steady-state data and identifying dynamic features of transient data. The model structure design takes into account the multi-scale characteristics of electrical parameters and achieves multi-level feature extraction by combining convolutional kernels of different sizes.

6. The method according to claim 1, characterized in that, The feature ranking process using Fast Fourier Transform and an improved self-attention mechanism also includes: The fast Fourier transform is applied to map the time-domain features to the frequency domain to obtain the frequency components and their amplitude information; The importance weights are calculated by combining the weight parameters learned during the training phase with the attention mechanism model. A self-attention mechanism is used to weight the frequency domain features and calculate the contribution of each frequency component to the load identification task. The weight values ​​reflect the relative importance of each feature in the classification.

7. The method according to claim 1, characterized in that, The normalization process based on min-max normalization and Z-score standardization, combined with zero-padding or interpolation extension, also includes: The normalization method employs a strategy that combines min-max normalization and Z-score standardization; The normalized feature sorting sequence is mapped to a fixed-length standardized feature vector structure, the length of which is configured according to the input requirements of the load identification model. For sequences that are too short, zero-padding or interpolation is used to extend them; for very long sequences, a feature selection strategy is used to retain the first N important features.

8. The method according to claim 1, characterized in that, The process of load identification using transfer learning to fine-tune and freeze some pre-trained layer weights also includes: It is built on a deep transfer learning framework and uses a multi-layer deep neural network. During the transfer learning adaptation phase, the model introduces a fine-tuning mechanism to adjust the weights of some network layers based on the characteristics of the current power consumption environment. By freezing the weights of some pre-trained layers and updating the parameters only for subsequent fully connected layers and classification layers, combined with a small amount of labeled data or unsupervised adaptive strategies, the model is adjusted to adapt to the new load distribution.

9. The method according to claim 1, characterized in that, The time alignment process based on L1 norm time window sliding matching and interpolation padding also includes: The input load type labels are aligned using a time series alignment algorithm with a time window sliding matching algorithm. The time series alignment module maps the input labels to a unified time axis according to a preset time resolution and uses interpolation or padding methods to handle the problem of uneven timestamp intervals. For labels with overlapping or conflicting times, the system filters them based on confidence level priority, retaining the label with the highest confidence level.

10. An electricity metering box power management system, applied to the method according to any one of claims 1-9, characterized in that, include: Install the parameter configuration module to obtain and calibrate meter box component information; The mixed signal acquisition module is used to capture steady-state and transient current and voltage waveforms; The feature extraction and sorting module, connected to the mixed signal acquisition module, is used to generate standardized feature vectors. The load identification and decision-making module, connected to the feature extraction and sorting module, is used to output load identifiers with timestamps. The strategy arbitration module, connected to the load identification and decision module, is used to generate opening and closing control commands; The instruction execution module, connected to the strategy arbitration module, is used to drive the circuit breaker mechanism to operate. The data management module connects all functional modules and is used to store and update load characteristic data.

Images