Electricity consumption analysis method and system based on non-intrusive load identification technology

Abstract

The invention provides an electricity utilization analysis method and system based on a non-intrusive load identification technology. The method comprises the following steps: data acquisition: collecting load data through a non-intrusive measuring switch or an intelligent electric meter; constructing an electrical load digital twinborn model, and establishing a bidirectional mapping relation between the high-frequency real-time data and the digital twinborn model; preprocessing the load data, including denoising, time synchronization and normalization; key features are extracted from the preprocessed load data; a neural network model is fed back, dynamically trained, optimized and adjusted by using a digital twinborn model; using the trained model to identify the decomposition result of each load; evaluating a load decomposition result; and carrying out load identification on the collected real-time load data, and carrying out multi-dimensional analysis on the power consumption behavior of the user. According to the invention, a non-intrusive load identification technology and a digital twinborn technology are combined, the power utilization analysis capability is improved by using high-frequency data, the identification model is optimized, the identification precision is improved, and the power utilization behavior can be deeply analyzed.

Description

Technical Field

[0001] The present invention relates to the technical field of load identification, and particularly to a power consumption analysis method and system based on non-intrusive load identification technology. Background Art

[0002] At present, with the rapid development of smart grids, the accurate analysis of users' power consumption behaviors plays a crucial role in aspects such as the optimal allocation of power resources, demand-side management, and energy conservation and consumption reduction. Traditional power load monitoring methods require the installation of monitoring devices on each electrical appliance separately, which not only incurs high costs, but also has a cumbersome and complex installation and maintenance process, and will cause great interference to the original power consumption system of users.

[0003] The non-intrusive load identification technology (NILM) emerged as the times require. It only needs to install a small number of sensors at the user's power inlet to collect data such as total current and total voltage, and then realizes the decomposition and identification of the loads of different electrical appliances, greatly reducing the monitoring cost and implementation difficulty. However, most studies have shown that non-intrusive load identification schemes based on low-frequency data have many limitations. On the one hand, the method has a low upper limit and it is difficult to achieve high-precision load identification; on the other hand, transient data cannot be effectively obtained, resulting in an incomplete capture of the operating state of electrical appliances, and ultimately poor load identification effects.

[0004] In contrast, high-frequency data carries more abundant information, and the algorithm can more effectively extract the characteristic information of electrical appliances from high-frequency data, so as to achieve more accurate electrical appliance identification. Therefore, non-intrusive load identification schemes based on high-frequency data show better effects and broad application values. However, there are still some problems to be solved urgently in the existing NILM methods based on high-frequency data, such as insufficient identification ability for unknown loads, and low identification accuracy when facing electrical appliances with different categories but similar powers.

[0005] At the same time, in a complex power consumption environment, the existing non-intrusive load identification technology also has shortcomings in data processing accuracy, model generalization ability, and comprehensiveness of power consumption analysis. When encountering new types of electrical appliances, the model is difficult to adapt quickly, resulting in a decline in load identification accuracy; moreover, the power consumption analysis results cannot provide sufficient and effective guidance for power management and energy-saving optimization. The emergence of digital twin technology provides a new idea for solving these problems. However, the current method of deeply integrating digital twin with non-intrusive load identification technology for power consumption analysis is not perfect enough to fully utilize the advantages of the combination of the two. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a power consumption analysis method and system based on non-intrusive load identification technology, which combines non-intrusive load identification technology with digital twin technology, uses high-frequency data to improve power consumption analysis ability, optimizes the identification model, improves the identification accuracy, and can deeply analyze power consumption behavior.

[0007] To solve the above technical problem, the embodiments of the present invention provide the following technical solutions:

[0008] A power consumption analysis method based on non-intrusive load identification technology includes the following steps:

[0009] Data collection, collecting load data through non-intrusive measurement switches or smart meters; wherein, the load data includes three-phase voltage signals, three-phase current signals, active power, reactive power, frequency, power factor, angular difference, harmonic content.

[0010] Using professional 3D modeling software, based on the load data of the user's electrical equipment and the power system architecture, constructing a digital twin model of the power consumption load, and establishing a two-way mapping relationship between high-frequency real-time data and the digital twin model.

[0011] Data preprocessing, using a denoising algorithm for the characteristics of high-frequency real-time data to remove noise signals, ensuring the time consistency of high-frequency data from different sensors through a high-precision time synchronization algorithm, and using a normalization method suitable for high-frequency data to map the load data to a specific interval.

[0012] Feature extraction, extracting key features from the preprocessed load data, the key features including frequency domain features, time domain features, and statistical features, and converting the original data into a feature vector that can reflect the load characteristics.

[0013] Selecting a neural network algorithm to construct a load identification model, collecting a large number of samples of known electrical equipment types and high-frequency load data for data annotation, then performing supervised training on the model, using the digital twin model to display the training process in real time, and dynamically optimizing and adjusting the neural network model according to the feedback of the digital twin model.

[0014] Using the trained neural network model to predict and decompose new unlabeled load data, and identifying the decomposition results of each load according to the output of the neural network model.

[0015] Comparing the decomposition results of each load with the pre-established actual load feature library, evaluating the results of the load decomposition, and adjusting the parameters of the load identification model or adopting other optimization strategies according to the evaluation results.

[0016] Taking the load identification model adjusted after comparison with the load feature library as the final load identification model, performing load identification on the collected real-time load data, and conducting multi-dimensional analysis of the user's power consumption behavior.

[0017] Optionally, for the constructed digital twin model of the electrical load, a two-way mapping relationship between the high-frequency real-time data and the digital twin model is established, specifically as follows:

[0018] Establish a mathematical model of the electrical load, where the electrical load includes resistive loads, motor loads, and power electronic loads;

[0019] Establish a physical simulation model of the digital twin, and establish a numerical model based on the principles and equations of the electrical load;

[0020] Integrate the data and the model, and integrate the actual operation data of the electrical load with the physical simulation model to form a complete digital twin model;

[0021] Verify the established digital twin model, and conduct scenario simulation and simulation tests.

[0022] Optionally, extract key features from the preprocessed load data. The key features include frequency-domain features, time-domain features, and statistical features. Specifically, combine the steady-state features of active power, reactive power, voltage, and current, and add the transient waveform of active power for feature extraction, where:

[0023] For frequency-domain features, the Fast Fourier Transform (FFT) captures the time-varying frequency components in the load data; in terms of time-domain feature extraction, calculation methods such as mean, variance, or crest factor basic statistics are used; for waveform feature extraction, a morphology-based algorithm is used; for statistical features, algorithms such as maximum value, minimum value, or standard deviation are used;

[0024] Segment the preprocessed continuous load data into frames of fixed length. Each frame contains a certain number of sampling points. The frame length is selected according to the change frequency of the data and the calculation efficiency to set the sampling points, and window processing is performed on the framed data;

[0025] Calculate the frequency-domain features, time-domain features, and statistical features respectively using corresponding algorithms.

[0026] Optionally, select a neural network algorithm to construct a load recognition model, specifically: use a Bidirectional Long Short-Term Memory (BiLSTM) network to construct a load recognition model.

[0027] Optionally, use the digital twin model to display the training process in real time, and dynamically optimize and adjust the neural network model according to the feedback of the digital twin model, specifically as follows:

[0028] Establish a data docking and communication mechanism, develop a dedicated data interface program, and realize data interaction between the digital twin model and the neural network training environment;

[0029] Display the training process in the digital twin model. In the 3D visualization interface of the digital twin model, design a dedicated area to display the neural network training process, and map the data during the neural network training process to the 3D scene of the digital twin model;

[0030] During the operation of the digital twin model, real-time monitor the state changes of power system components and simulate the occurrence of faults, integrate feedback information from different sources to form a unified feedback data set;

[0031] Compare the actual operating state of the power system fed back by the digital twin model with the prediction results of the neural network to evaluate the training effect of the neural network;

[0032] According to the analysis of the feedback information and training effect, dynamically adjust the parameters of the neural network, optimize the structure, increase the training samples, and optimize the feature engineering.

[0033] Optionally, use the trained neural network model to predict and decompose new unlabeled load data, and identify the decomposition results of each load according to the output of the neural network model. Specifically:

[0034] Continuously collect new unlabeled load data through non-intrusive measurement switches or smart meters at a pre-determined high-frequency sampling rate;

[0035] Perform preprocessing on the collected new unlabeled load data, including but not limited to data cleaning and data normalization;

[0036] According to the feature extraction method determined during model training, extract key features from the preprocessed new data, including frequency-domain features, time-domain features, and statistical features, and organize the extracted features into a format acceptable to the model;

[0037] The model uses the patterns and parameters learned during training to perform forward propagation calculations based on the input feature data, and outputs prediction results represented in the form of probability distributions or numerical values;

[0038] According to the prediction results output by the model, map them to the actual electrical equipment types and load parameters, and verify the initially identified load decomposition results, and make judgments in combination with the physical simulation results or the operating common sense of the actual power system.

[0039] Optionally, compare the decomposition results of each load with a pre-established actual load feature library to evaluate the load decomposition results. Specifically:

[0040] Extract the decomposition result data of each load from the output of the load decomposition model. The decomposition result data includes information such as the electrical equipment type corresponding to the load, relevant electrical parameters, and the confidence level predicted by the model;

[0041] Access the pre - established actual load characteristic library, which is stored using a relational database or a non - relational database. According to the types of electrical equipment involved in the decomposition result, read the load characteristic data of the corresponding equipment from the characteristic library, including the rated parameters of the equipment, the typical operating parameter ranges, and the characteristic curves under different operating conditions;

[0042] Match the types of electrical equipment in the load decomposition result with the equipment types in the actual load characteristic library, including electrical parameter comparison and characteristic curve comparison;

[0043] According to the results of electrical parameter comparison and characteristic curve comparison, calculate various error indicators. According to the results of equipment type matching, calculate the accuracy rate of load decomposition, and evaluate the load decomposition result in combination with operation reliability.

[0044] Optionally, according to the evaluation results, adjust the parameters of the load recognition model or adopt other optimization strategies. Specifically: adjusting the parameters of the load recognition model includes adjusting the learning rate, weight decay coefficient, and the number of neurons in the hidden layer; adopting other optimization strategies includes increasing training samples, data augmentation, and feature engineering optimization.

[0045] Optionally, perform load recognition on the collected real - time load data and conduct multi - dimensional analysis of user electricity consumption behavior, specifically including: time - dimension analysis, equipment - dimension analysis, and user - group - dimension analysis; among them, time - dimension analysis includes daily electricity consumption patterns, weekly electricity consumption differences, monthly and annual electricity consumption trends; equipment - dimension analysis includes the electricity consumption ratio of equipment, equipment usage frequency and duration, and equipment combined usage patterns; user - group - dimension analysis includes electricity consumption differences among users of different occupations, household size and electricity consumption differences, and regional electricity consumption differences.

[0046] The present invention also proposes an electricity consumption analysis system based on non - intrusive load recognition technology, including:

[0047] A data acquisition module that collects load data through a non - intrusive measurement switch or a smart meter; among them, the load data includes three - phase voltage signals, three - phase current signals, active power, reactive power, frequency, power factor, phase difference, and harmonic content;

[0048] A data twin module that constructs a digital twin model of the electricity load based on the load data of the user's electrical equipment and the power system architecture, and establishes a two - way mapping relationship between high - frequency real - time data and the digital twin model;

[0049] A data pre - processing module that uses a denoising algorithm for high - frequency real - time data characteristics to remove noise signals, ensures the time consistency of high - frequency data from different sensors through a high - precision time synchronization algorithm, and maps the load data to a specific interval using a normalization method suitable for high - frequency data;

[0050] A feature extraction module extracts key features from the preprocessed load data. The key features include frequency-domain features, time-domain features, and statistical features, and transforms the original data into a feature vector that can reflect the load characteristics.

[0051] A load recognition model construction module collects a large number of samples of known electrical equipment types and high-frequency load data for data annotation, then conducts supervised training on the model, uses a digital twin model to display the training process in real time, and dynamically optimizes and adjusts the neural network model according to the feedback of the digital twin model.

[0052] A load decomposition and recognition module uses the trained neural network model to predict and decompose new unlabeled load data, and identifies the decomposition results of each load according to the output of the neural network model.

[0053] A load recognition evaluation module compares the decomposition results of each load with a pre-established actual load feature library to evaluate the results of load decomposition, and adjusts the parameters of the load recognition model or adopts other optimization strategies according to the evaluation results.

[0054] An electricity consumption behavior analysis module uses the load recognition model adjusted after comparison with the load feature library as the final load recognition model to perform load recognition on the collected real-time load data and conduct multi-dimensional analysis of the user's electricity consumption behavior.

[0055] Compared with the prior art, the above technical solutions of the present invention have at least the following beneficial effects:

[0056] In the above solution, a non-intrusive load recognition technology is adopted. Only a small number of sensors are installed at the user's power inlet to collect data such as total current and total voltage to achieve load decomposition and recognition. Compared with the traditional method that requires monitoring devices to be installed on each electrical equipment, the monitoring cost is greatly reduced, and the installation and maintenance processes are simple, reducing the interference to the user's original power consumption system.

[0057] Load recognition is based on high-frequency data. High-frequency data carries rich information, and the algorithm can more effectively extract electrical appliance feature information. Combining the bidirectional mapping relationship of the digital twin model, the operation state feedback of the actual power system is obtained in real time to optimize the neural network model, thereby achieving more accurate electrical appliance recognition and improving the accuracy of load recognition.

[0058] By establishing a digital twin model, verifying and conducting scenario simulation tests on it, it is ensured that the model accurately reflects the actual power system. During the model training process, the digital twin model is used to display the training situation in real time, and the parameters and structure of the neural network model are dynamically adjusted according to the feedback, increasing the training samples and optimizing the feature engineering to improve the generalization ability of the model, the processing ability of complex data, and the adaptability to new equipment types.

[0059] Conduct multi-dimensional analysis of users' electricity consumption behaviors in terms of time, equipment, and user groups. The time-dimensional analysis helps the power department formulate time-of-use electricity price policies, reasonably arrange power generation plans, and grid dispatching; the equipment-dimensional analysis helps users carry out energy-saving renovations and provides data support for the development of smart homes; the user group-dimensional analysis provides a basis for the power department to implement differential services, planning, and management, and realizes the optimal allocation of power resources. Description of the Drawings

[0060] Figure 1 It is a flowchart of the electricity consumption analysis method based on non-intrusive load identification technology of the present invention;

[0061] Figure 2 It is a schematic block diagram of the electricity consumption analysis system based on non-intrusive load identification technology of the present invention;

[0062] Figure 3 It is a line graph of four characteristics of voltage, current, fundamental active power, and fundamental reactive power of different household appliances. Detailed Implementation Modes

[0063] To make the technical problems, technical solutions, and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the drawings and specific embodiments.

[0064] As Figure 1 shown, an embodiment of the present invention proposes an electricity consumption analysis method based on non-intrusive load identification technology, including:

[0065] S1. Data collection, collecting load data through non-intrusive measurement switches or smart meters; wherein, the load data includes three-phase voltage signals, three-phase current signals, active power, reactive power, frequency, power factor, angular difference, and harmonic content;

[0066] S2. Using professional 3D modeling software, constructing a digital twin model of the electricity consumption load according to the load data of the user's electricity consumption equipment and the power system architecture, and establishing a two-way mapping relationship between the high-frequency real-time data and the digital twin model;

[0067] S3. Data preprocessing, using a denoising algorithm for the characteristics of high-frequency real-time data to remove noise signals, ensuring the time consistency of high-frequency data of different sensors through a high-precision time synchronization algorithm, and mapping the load data to a specific interval by a normalization method applicable to high-frequency data;

[0068] S4. Feature extraction, extracting key features from the preprocessed load data, the key features including frequency-domain features, time-domain features, and statistical features, and converting the original data into a feature vector that can reflect the load characteristics;

[0069] S5. Select the neural network algorithm to construct the load recognition model. After collecting a large number of samples of known electrical equipment types and high-frequency load data for data annotation, perform supervised training on the model, use the digital twin model to display the training process in real time, and dynamically optimize and adjust the neural network model according to the feedback of the digital twin model;

[0070] S6. Use the trained neural network model to predict and decompose new unlabeled load data, and identify the decomposition results of each load according to the output of the neural network model;

[0071] S7. Compare the decomposition results of each load with the pre-established actual load feature library, evaluate the results of load decomposition, and adjust the parameters of the load recognition model or adopt other optimization strategies according to the evaluation results;

[0072] S8. Use the load recognition model adjusted after comparison with the load feature library as the final load recognition model to perform load recognition on the collected real-time load data and conduct multi-dimensional analysis of user electricity consumption behavior.

[0073] In step S1, load data is collected through a non-intrusive measurement switch or smart meter. The low-voltage intelligent measurement switch realizes the acquisition of the grid quantity signal at the incoming line end of the meter box, and performs power load feature extraction, feature recognition, feature decomposition, and feature classification on the grid quantity signal of the incoming line of the meter box (including: three-phase voltage signal, three-phase current signal, active power, reactive power, frequency, power factor, angular difference, harmonic quantity, etc.). It collects the electricity load recognition data of all smart electricity meters in the meter box through a communication system such as a carrier communication module, RS485, or Bluetooth.

[0074] In this embodiment, in step S2, a digital twin model of the electricity load is constructed, and a two-way mapping relationship between high-frequency real-time data and the digital twin model is established. Specifically:

[0075] (1) Establish a mathematical model of the electricity load

[0076] Mathematical model of resistive load: The resistive load follows Ohm's law. In an AC circuit, assume the voltage u = U m sin(ωt), and the resistance is R, then the current Active power (U is the effective value of the voltage). Obtain the resistance value through actual measurement or by referring to the equipment parameters. When multiple resistive elements are combined, calculate the equivalent resistance according to the series-parallel relationship. The mathematical models of motor-type loads and power electronic-type loads will not be elaborated here.

[0077] (2) Establish a physical simulation model of the digital twin

[0078] Establish a numerical model based on principles and equations: Based on the basic principles of electrical loads, such as Kirchhoff's Current Law (KCL) and Kirchhoff's Voltage Law (KVL), combined with the established mathematical model of electrical loads, use numerical calculation methods to simulate the physical process in simulation software. For example, use the finite element method, finite difference method, etc. to perform numerical calculations on the electric and magnetic field distributions of transmission lines, and simulate the steady-state and transient processes of the power system by iteratively solving circuit equations. Set appropriate time steps in the simulation software to ensure calculation accuracy and efficiency, and achieve dynamic simulation of the operating state of the power system.

[0079] Model construction and parameter setting: In professional simulation software (such as MATLAB / Simulink, PSCAD, etc.), build corresponding simulation modules according to the mathematical model of electrical loads. Combine each load model according to the connection method of the actual power system, and set the parameters of the model to be consistent with the parameters of the actual electrical equipment and system. For example, for the motor model, set parameters such as the rated power, rated voltage, and rated current of the motor; for the power electronics model, set control parameters such as the switching frequency and duty cycle.

[0080] (3) Integrate data and models

[0081] Data collection and transmission: Collect the actual operating data of electrical loads, including parameters such as current, voltage, and power, through non-intrusive measurement switches or smart meters. Use wired or wireless communication technology to transmit the collected data to the data processing center in real time. Perform preprocessing on the transmitted data, such as denoising, filtering, data alignment, etc., to ensure the accuracy and reliability of the data.

[0082] Data and model fusion: Input the preprocessed actual operating data into the physical simulation model, and achieve the interaction between data and model through the data interface. In the simulation model, adjust the operating state and parameters of the model according to the input actual data, so that the model can reflect the actual operating conditions of electrical loads in real time. At the same time, compare and analyze the output results of the simulation model with the actual measurement data to verify the accuracy and reliability of the model. Through continuous adjustment of model parameters and data processing methods, achieve deep fusion of data and model, and form a complete digital twin model.

[0083] (4) Verify the established digital twin model, and conduct scenario simulation and simulation testing

[0084] Data comparison and verification: Collect the measurement data of the actual power system under different operating conditions and conduct a comparative analysis with the data output by the digital twin model. Calculate the error metrics between the two, such as the root mean square error (RMSE), mean absolute error (MAE), etc., to evaluate the accuracy of the digital twin model. If the error exceeds the allowable range, analyze the reasons and adjust and optimize the model parameters, such as recalibrating the sensor data, correcting the parameters of the mathematical model, etc.

[0085] Consistency check: Check the logical relationships and physical characteristics of the digital twin model to ensure that the connection relationships, electrical characteristics, and operating logics between the components in the model are consistent with the actual power system. For example, check whether the connections between transmission lines and transformers, generators, and loads are correct, and verify whether the operating characteristics of the power equipment in the model conform to the actual physical laws. By carefully checking the structure and parameters of the model, ensure the consistency of the digital twin model with the actual system at the physical level.

[0086] Scenario simulation and simulation test: Set various different operating conditions in the digital twin model, such as normal operation, load mutation, short circuit fault, open circuit fault, etc. By changing the input parameters and boundary conditions of the model, simulate the operation of the actual power system under various conditions. For example, when simulating a load mutation, instantaneously increase or decrease the power of a certain part of the load, and observe the changes in parameters such as current, voltage, and power in the digital twin model, as well as the response characteristics of each power equipment. According to the results of the scenario simulation, evaluate the performance of the digital twin model, analyze indicators such as the stability, accuracy, and response speed of the system under different conditions, and provide a basis for further optimizing and improving the digital twin model.

[0087] Step S3 Data preprocessing: Use a denoising algorithm for high-frequency real-time data characteristics to remove noise signals, ensure the time consistency of high-frequency data from different sensors through a high-precision time synchronization algorithm, and use a normalization method suitable for high-frequency data to map the load data to a specific interval.

[0088] First of all, high-frequency data is extremely vulnerable to interference from factors such as complex electromagnetic environments and the electronic noise of the equipment itself, resulting in a large amount of noise being mixed into the data, seriously affecting subsequent analysis and the accuracy of the model. Therefore, a denoising algorithm based on wavelet packet transform is adopted. This algorithm makes full use of the fine partitioning ability of wavelet packets in the time-frequency domain and can adaptively select appropriate wavelet packet basis functions to decompose the signal according to the frequency characteristics of high-frequency data. By accurately identifying the noise components in different frequency bands, separating the noise signal from the useful signal, and then using means such as threshold processing to remove the noise, a pure high-frequency signal is finally reconstructed, maximizing the retention of the effective information in the original high-frequency data. For example, in an actual power system, high-frequency current signals may be affected by electromagnetic interference generated by nearby communication equipment, high-power electrical appliances, etc. After denoising by wavelet packet transform, the true change characteristics of the load current can be clearly presented.

[0089] Secondly, when different sensors collect high-frequency data, due to factors such as differences in transmission paths and equipment performance, there is a time delay in the data reaching the data processing center, which will cause the problem of time inconsistency, seriously affecting the relevance of the data and the accuracy of the analysis results. The time synchronization method based on the Global Positioning System (GPS) can effectively solve this problem. The GPS system provides a high-precision time reference. The high-precision clock built into the sensor is synchronized and calibrated with the GPS time, and accurate timestamp information is added when collecting data. At the data processing center, through the analysis and comparison of timestamps, techniques such as time interpolation and alignment algorithms are used to accurately align the high-frequency data of different sensors in the time dimension, ensuring the consistency of the data in time, and the error can be controlled within an extremely small range of microseconds or even nanoseconds. For example, in a multi-node power monitoring network, after the sensors of each node are synchronized by GPS time, the operating state changes of the power system at the same moment can be accurately captured, providing a reliable data basis for subsequent collaborative analysis and fault diagnosis.

[0090] Finally, there are significant differences in the amplitude and variation range between high-frequency data and low-frequency data, and the dimensions and scales of different types of high-frequency data (such as voltage, current, power, etc.) are different, which brings great difficulties to subsequent data processing and model analysis. The normalization formula based on the data distribution characteristics can map the data to a specific interval, such as [0, 1], according to the specific distribution of high-frequency data. This formula analyzes the statistical characteristics of the data, such as the maximum value, minimum value, mean value, standard deviation, etc., and uses linear or non-linear transformation methods to convert the original data into data with a unified scale and dimension. For example, for high-frequency voltage data, according to its fluctuation range and distribution characteristics under different working conditions, through normalization processing, it is converted to the interval [0, 1], so that the voltage data can be analyzed and compared with other high-frequency data (such as current, power) on the same scale, providing a standardized data basis for subsequent feature extraction, model training and data analysis, and improving the convergence speed and accuracy of the model.

[0091] Step S4 extracts key features from the preprocessed load data. The key features include frequency-domain features, time-domain features, and statistical features. When an electrical device is put into use, by extracting the steady-state features and transient features of the load signature, it can assist in identifying the type of electrical appliance. The feature extraction technology based on steady-state features cannot effectively handle some scenarios with higher identification difficulties, such as similar features and overlap of electrical devices; while the extraction technology based on transient features has stronger adaptability because the load signature of transient features can better reflect the characteristics and functions of different devices. In addition, the transient process event is short, and the possibility of feature overlap is less. However, the extraction of transient features places higher requirements on data acquisition and processing. Therefore, the feature extraction technology of the present invention comprehensively considers the steady-state features and transient features, absorbs the advantages of both, and can further improve the load identification accuracy. Combining the steady-state features of active power, reactive power, voltage, and current, and adding the transient waveform of active power for feature extraction can obtain higher identification accuracy.

[0092] As Figure 3 shown, the four subgraphs are respectively line graphs of four features: voltage, current, fundamental active power, and fundamental reactive power of different household appliances. Through the line graph of the current feature, different electrical devices can be identified. Of course, for different electrical appliances and the same electrical appliance in different operating modes and states, their characteristic parameters are also different. Taking an air conditioner as an example, the active power of the air conditioner varies greatly between cooling and heating, and there are also obvious differences in reactive power. The differences in the characteristic parameters of each electrical appliance can be clearly seen from the line graph. Using these differences, as Figure 3 shown, they can be separated. The methods for extracting key features are as follows:

[0093] (1) Select a feature extraction algorithm

[0094] Frequency domain feature extraction algorithm: For frequency domain features, Fourier transform related algorithms are selected, such as the Fast Fourier Transform (FFT). FFT can efficiently convert time-domain signals into frequency-domain signals and is suitable for processing large-scale load data. For scenarios that require more refined frequency analysis, the Short-Time Fourier Transform (STFT) can also be considered. It can analyze the frequency characteristics of signals within different time windows and capture the time-varying frequency components in load data.

[0095] Time domain feature extraction algorithm: In terms of time domain feature extraction, calculation methods of basic statistics such as mean, variance, and crest factor are adopted. For the extraction of waveform features, algorithms based on morphology can be used, such as using operations like erosion and dilation in mathematical morphology to analyze the contour and details of the load data waveform and obtain features such as waveform factors.

[0096] Statistical feature extraction algorithm: For statistical features, in addition to calculating common maximum, minimum, and standard deviation values, statistical methods based on quantiles can also be used, such as calculating the 25% and 75% quantiles, etc., to more comprehensively describe the distribution characteristics of load data. At the same time, calculation methods of skewness and kurtosis are used to analyze the asymmetry and steepness of data distribution.

[0097] (2) Data framing and windowing (for frequency domain and some time domain features)

[0098] Data framing: The preprocessed continuous load data is segmented into frames of fixed length, and each frame contains a certain number of sampling points. The selection of the frame length needs to comprehensively consider the change frequency of the data and the calculation efficiency. Generally, it can be set to several hundred to several thousand sampling points according to experience. For example, for load data with high-frequency sampling, a shorter frame length (such as 256 sampling points) can be selected; for relatively slowly changing low-frequency data, the frame length can be appropriately increased (such as 1024 sampling points).

[0099] Windowing: To reduce problems such as spectral leakage, the framed data is windowed. Commonly used window functions include Hanning window, Hamming window, Blackman window, etc. Select an appropriate window function according to the data characteristics and analysis requirements. For example, the Hanning window performs well in reducing spectral leakage and is suitable for most load data analysis scenarios. Multiply the window function by the framed data to make the data transition smoothly at the frame boundaries and improve the accuracy of frequency domain analysis.

[0100] (3) Frequency domain feature calculation

[0101] FFT calculation: Perform FFT operations on each windowed frame of data to obtain the frequency domain representation. Through FFT calculation, the amplitude and phase information corresponding to each frequency point can be obtained. For example, for a frame of data with length N, after FFT, N / 2 + 1 independent frequency components can be obtained (assuming N is even), where the frequency range is from 0 to half of the sampling frequency.

[0102] Fourier transform coefficient extraction: Extract Fourier transform coefficients from the FFT results. These coefficients reflect the energy distribution of the signal at different frequencies. Usually, key frequency components in the low-frequency band and high-frequency band are concerned, such as the fundamental frequency (50 Hz or 60 Hz) and its harmonic frequency components in the power system. According to actual needs, extract the coefficients at specific frequency points, or calculate statistical quantities such as the sum or average value of the coefficients within a certain frequency range as frequency-domain features.

[0103] Power spectrum estimation: Use the FFT results for power spectrum estimation. Commonly used methods include the periodogram method, Welch method, etc. Power spectrum estimation can more intuitively display the distribution of the signal power at different frequencies, providing a basis for analyzing the frequency characteristics of load data. Through power spectrum estimation, the main frequency components of the signal and their corresponding power magnitudes can be obtained to further explore the frequency-domain features of load data.

[0104] (4) Time-domain feature calculation

[0105] Calculation of basic statistics: Calculate the mean value of the load data, which is the average value of all sampling points and reflects the average level of the signal; calculate the variance to measure the degree of data dispersion. The larger the variance, the greater the data fluctuation; calculate the peak factor, which is the ratio of the peak value to the effective value, and is used to evaluate the peak characteristics of the signal, which is of great significance for identifying data with impulsive loads.

[0106] Calculation of waveform factors: Analyze the waveform of the load data to calculate waveform factors such as the crest factor and waveform coefficient. The crest factor is defined as the ratio of the peak value to the root mean square value, and the waveform coefficient is the ratio of the effective value to the average value. These waveform factors can reflect the waveform shape of the signal and have a certain indicating effect on distinguishing different types of loads.

[0107] Feature extraction based on morphology: Use mathematical morphology algorithms to perform operations such as erosion and dilation on the waveform of the load data to extract the contour features and detailed information of the waveform. For example, remove small noises and burrs in the waveform through erosion operations, and highlight the main features of the waveform through dilation operations. According to the results of morphological processing, calculate relevant feature parameters such as the area, perimeter, and centroid of the waveform as a supplement to the time-domain features.

[0108] (5) Statistical feature calculation

[0109] Calculation of maximum, minimum, and standard deviation: Determine the maximum and minimum values of the load data, which reflect the value range of the data; calculate the standard deviation to measure the degree of data dispersion relative to the mean value. The smaller the standard deviation, the more concentrated the data is around the mean value. This maximum, minimum, and standard deviation information can help understand the fluctuation range and stability of the load data.

[0110] Quantile calculation: Calculate the quantiles of the load data, such as the 25% quantile (the first quartile), 75% quantile (the third quartile), etc. Quantiles can more comprehensively describe the distribution of data. By analyzing the values of different quantiles, the distribution characteristics of data in different intervals can be understood, which is of great significance for discovering outliers and extreme points in the data.

[0111] Skewness and kurtosis calculation: Calculate the skewness of the load data to measure the asymmetry of the data distribution. A positive skewness indicates that the data distribution is right-skewed, that is, there is a longer tail on the right side; a negative skewness indicates that the data distribution is left-skewed. Calculate the kurtosis to describe the steepness of the data distribution. The larger the kurtosis, the more concentrated the data distribution is around the mean and the thinner the tail. The calculation of skewness and kurtosis can further reveal the distribution characteristics of the load data and provide more dimensional information for data analysis.

[0112] In step S5, a load recognition model is constructed using a bidirectional long short-term memory network BiLSTM. After collecting a large number of samples of known electrical equipment types and high-frequency load data for data annotation, the model is supervised and trained.

[0113] A bidirectional long short-term memory network (BiLSTM) is used to extract the time-dependent relationships in the power load data. It is a special type of recurrent neural network and has significant advantages over the standard unidirectional LSTM in processing power load sequence data. In the standard LSTM, information is controlled by memory units, input gates, forget gates, and output gates to flow. Different from the standard LSTM, BiLSTM contains two independent layers that process the forward (past to future) and backward (future to past) information of the power load sequence respectively. In this way, the output at each time point contains the context information before and after, which is particularly important in power load decomposition because the current power consumption may be affected by events before and after. In BiLSTM, the forward and backward LSTMs usually operate independently. They each process the input sequence and generate their own output sequences. These outputs are then combined to form a unified representation that synthesizes the information learned from both directions. Power load data usually contains complex time dependencies, such as periodic patterns and long-term trends. BiLSTM can effectively capture this dependency through its bidirectional structure. In addition, power load data is often very complex, containing various different electricity consumption patterns and random fluctuations. The complex network structure of BiLSTM can handle this non-linearity and complexity and extract useful features. In the case of multiple electrical appliances running simultaneously, BiLSTM can more accurately decompose the load contributions of each electrical appliance by considering the electricity consumption patterns before and after, thus improving the decomposition accuracy.

[0114] Before retraining the model, collect a large number of samples containing known types of electrical equipment and high-frequency load data. The high-frequency load data can be obtained through the non-intrusive measurement switches or smart meters mentioned above. The data should cover information such as high-frequency current, voltage, power, etc. of various electrical equipment under different operating states. Then, label the collected samples to clarify the type of electrical equipment corresponding to each sample. The labeling process can be carried out manually or by automatically labeling in combination with prior knowledge to ensure the accuracy and consistency of the labeling. Next, divide the labeled data into a training set, a validation set, and a test set. Generally, the training set accounts for 70%-80% of the total data and is used for training the model; the validation set accounts for 10%-15% and is used to evaluate the performance of the model and adjust the hyperparameters of the model during the training process; the test set accounts for 10%-15% and is used to finally evaluate the generalization ability of the model.

[0115] Finally, perform supervised training on the model, use the digital twin model to display the training process in real time, and dynamically optimize and adjust the neural network model according to the feedback of the digital twin model. In the process of constructing and validating the digital twin model of the electrical load, the key to using the digital twin model to assist in neural network optimization lies in establishing the connection between the two, obtaining feedback, and adjusting accordingly. Specifically, it includes:

[0116] (1) Establish a data docking and communication mechanism

[0117] Data interface development: Develop a dedicated data interface program to achieve data interaction between the digital twin model and the neural network training environment (such as deep learning frameworks like TensorFlow and PyTorch). This interface is responsible for transmitting key data during the neural network training process, such as input features of training samples, predicted outputs, loss function values, etc., to the digital twin model; at the same time, receiving information feedback from the digital twin model, such as real-time state changes of power system components, simulated fault information, etc.

[0118] Communication protocol selection: Select a suitable communication protocol, such as the TCP / IP protocol, to ensure stable and efficient data transmission between the digital twin model and the neural network training environment. Serialize and deserialize the transmitted data to ensure the consistency and compatibility of the data format.

[0119] (2) Display the training process in the digital twin model

[0120] Visualization interface design: In the 3D visualization interface of the digital twin model, design a dedicated area to display the neural network training process. For example, create a data panel to display information such as the number of training samples, the current training epoch, and the change curve of the loss function in real time; at the same time, visually display the operation state changes of power system components at different training stages through colors, animations, etc.

[0121] Data mapping and display: Map the data during the neural network training process into the 3D scene of the digital twin model. For example, according to the prediction output results, dynamically change the color or blinking frequency of the electrical equipment model to represent the model's judgment on the equipment operation status; display the dynamic changes of current and voltage on the power transmission line model to show the learning situation of the neural network on the power transmission status.

[0122] (3) Obtain the feedback information of the digital twin model

[0123] Real-time monitoring and analysis: During the operation of the digital twin model, real-time monitor the state changes of power system components, simulate the occurrence of faults, etc. Through data analysis algorithms, extract key information, such as the location and type of faults, abnormal fluctuations in power system parameters, etc., and transmit them as feedback information to the neural network training environment.

[0124] Feedback information integration: Integrate the feedback information from different sources to form a unified feedback data set. Preprocess the feedback information, such as data cleaning, normalization, etc., so that it can be effectively received and processed by the neural network.

[0125] (4) Analyze the training effect based on the feedback information

[0126] Comparison and evaluation: Compare the actual power system operation status feedback by the digital twin model with the prediction results of the neural network, and evaluate the training effect of the neural network. Calculate prediction error metrics, such as mean square error (MSE), mean absolute error (MAE), etc., and analyze the causes and distributions of the errors.

[0127] Analysis of influencing factors: Deeply analyze the factors in the feedback information of the digital twin model that affect the training effect of the neural network. For example, inaccurate model predictions under certain complex working conditions may be caused by insufficient training samples, insufficient feature extraction, or unreasonable neural network structure.

[0128] (5) Dynamically optimize and adjust the neural network model

[0129] Parameter adjustment: Dynamically adjust the parameters of the neural network according to the feedback information and training effect analysis. For example, if it is found that the loss function converges slowly or fluctuates during training, parameters such as the learning rate and weight decay coefficient can be appropriately adjusted; update the weights and biases of the neural network through the backpropagation algorithm to reduce the prediction error.

[0130] Structure optimization: If it is found through feedback analysis that there are defects in the neural network structure, such as too many or too few layers, unreasonable number of neurons, etc., optimize the neural network structure. You can try to increase or decrease the number of hidden layers, adjust the connection method of neurons, re-initialize the network parameters, and then re-train.

[0131] Increasing training samples: If the feedback from the digital twin model shows that the neural network's prediction performance is poor under certain specific working conditions, it may be due to insufficient training samples. At this time, obtain more simulation data under relevant working conditions from the digital twin model, add it to the training sample set, and retrain the neural network to improve the model's adaptability and generalization ability to different working conditions.

[0132] Feature engineering optimization: Optimize the features input into the neural network according to the information feedback from the digital twin model. For example, if it is found that certain features do not contribute much to the model prediction, consider removing these features; or extract new features according to the actual operation of the power system. For example, in the fault diagnosis scenario, add features related to the fault type to improve the neural network's fault recognition ability.

[0133] In this embodiment, step S6 uses the trained neural network model to predict and decompose new unlabeled load data, and identifies the decomposition results of each load according to the output of the neural network model, including:

[0134] First, continuously collect new unlabeled load data through non-intrusive measurement switches or smart meters at a pre-determined high-frequency sampling frequency. Ensure the integrity and accuracy of data collection, and preliminarily organize the collected data, classifying and storing it according to time sequence and data type for convenient subsequent processing.

[0135] Next, check whether there are outliers and missing values in the data. For outliers, anomaly detection methods based on statistical methods or machine learning algorithms can be used for identification and processing. For example, replace the outliers with reasonable estimated values or mark them for subsequent analysis. For missing values, select an appropriate filling method according to the data characteristics, such as mean filling, linear interpolation, or filling methods based on model prediction.

[0136] Data normalization: Use the same normalization method as the training data to map the feature values of the new data to a specific interval, such as [0, 1] or [-1, 1].

[0137] Then, according to the feature extraction method determined during model training, extract key features from the preprocessed new data, including frequency-domain features (such as Fourier transform coefficients), time-domain features (such as mean, variance, waveform factor), and statistical features (such as maximum value, minimum value, standard deviation). Ensure the consistency and accuracy of feature extraction to provide effective data input for model prediction. Organize the extracted features into a format acceptable to the model, for example, form feature vectors or feature matrices from the features, and input them into the trained neural network model according to the input structure set during model training. The model uses the patterns and parameters learned during training to perform forward propagation calculations based on the input feature data and outputs the prediction results. The prediction results are usually represented in the form of probability distributions or numerical values, reflecting the likelihood that each load belongs to different types of electrical equipment or the relevant parameter values.

[0138] Finally, perform load decomposition result identification. If the model output is a probability distribution, set a reasonable probability threshold (such as 0.8). For each load, compare the probability predicted by the model with the threshold. If the probability of a certain type of electrical equipment is greater than the threshold, then identify this load as the electrical equipment of this type; if all probabilities are less than the threshold, it can be marked as an unknown load or further analyzed. According to the prediction results output by the model, map them to the actual types of electrical equipment and load parameters. For example, if the features corresponding to a certain load output by the model match the feature pattern of a refrigerator the best, then identify this load as a refrigerator and determine the operating state and load size of the refrigerator according to the relevant parameters (such as power, current, etc.) output by the model.

[0139] Verify the initially identified load decomposition results, which can be judged by combining the physical simulation results in the digital twin model or the operating common sense of the actual power system. If it is found that the results are unreasonable, such as the load power exceeding the rated range of the equipment or not conforming to the actual operating scenario, the results can be corrected, for example, recheck the data, adjust the model parameters, or use other auxiliary analysis methods for further judgment.

[0140] In this embodiment, step S7 compares the decomposition results of each load with the pre-established actual load feature library to evaluate the results of load decomposition. According to the evaluation results, adjust the parameters of the load identification model or adopt other optimization strategies. Specifically:

[0141] (1) Data reading and preparation

[0142] Extract the decomposition results: From the output of the load decomposition model, extract the decomposition result data of each load. These data usually include information such as the type of electrical equipment corresponding to the load, relevant electrical parameters (such as power, current, voltage, etc.), and the confidence level predicted by the model. Organize these data into a structured format for subsequent comparison with the actual load feature library.

[0143] Read the actual load characteristic library: Access the pre-established actual load characteristic library, which stores a large amount of load characteristic data of various electrical equipment under different operating conditions that have been actually measured and verified. The characteristic library can be stored using a relational database or a non-relational database. According to the types of electrical equipment involved in the decomposition results, read the load characteristic data of the corresponding equipment from the characteristic library, including the rated parameters of the equipment, the typical operating parameter ranges, and the characteristic curves under different working conditions, etc.

[0144] (2) Feature matching and comparison

[0145] Equipment type matching: First, match the types of electrical equipment in the load decomposition results with the equipment types in the actual load characteristic library. Ensure their consistency. If there are mismatches, the reasons need to be further analyzed. It may be due to model recognition errors or the lack of information on corresponding equipment types in the characteristic library. For mismatched equipment types, they can be marked as abnormal situations and recorded and analyzed separately.

[0146] Electrical parameter comparison: For the matched equipment types, compare the electrical parameters in the load decomposition results with the corresponding parameters in the characteristic library one by one. For example, compare parameters such as the active power, reactive power, root mean square current, and root mean square voltage of the load. Calculate the difference between the actual measured value and the standard value or typical value in the characteristic library, as well as the percentage of the difference to the standard value, to quantify the degree of difference between the two.

[0147] Characteristic curve comparison: For some electrical equipment with complex operating characteristics, such as the load characteristics of motors at different speeds and the loss characteristics of transformers at different load ratios, compare the characteristic curves in the load decomposition results with the standard characteristic curves in the characteristic library. Methods such as curve fitting and similarity calculation can be used to evaluate the similarity between the two and determine whether the load decomposition results conform to the operating characteristics of the actual equipment.

[0148] (3) Calculation of evaluation indicators

[0149] Error calculation: According to the results of electrical parameter comparison and characteristic curve comparison, calculate various error indicators. Commonly used error indicators include absolute error (AE), relative error (RE), root mean square error (RMSE), etc.

[0150] Accuracy evaluation: Calculate the accuracy of load decomposition according to the results of equipment type matching.

[0151] Reliability assessment: In addition to error and accuracy metrics, the load decomposition results can also be evaluated from the perspective of reliability. For example, analyze the stability of the model under different operating conditions, and observe the fluctuations of error metrics and accuracy at different times and different load levels; evaluate the model's ability to handle abnormal situations, such as when new equipment types or equipment failures not included in the feature library occur, and observe the model's performance and response capabilities.

[0152] (4) Analysis and feedback of evaluation results

[0153] Result analysis: Conduct in-depth analysis of the calculated evaluation metrics to judge the accuracy and reliability of the load decomposition results. If the error metrics exceed the pre-set allowable range, or the accuracy is low, it is necessary to analyze the causes of the problems. Possible reasons include insufficient model training, incomplete feature extraction, imperfect actual load feature library, data acquisition errors, etc. By analyzing the evaluation results in detail, identify the root causes of the problems to provide directions for subsequent optimization and improvement.

[0154] Feedback and improvement: According to the evaluation results, feedback the analyzed problems to the relevant links. For example, if the problem is caused by insufficient model training, the number of training samples can be increased, the training algorithm can be adjusted, or the model structure can be optimized; if the feature extraction is incomplete, the feature extraction method can be improved to add more effective features; if the actual load feature library is imperfect, actual measurement data can be further collected and supplemented to update and improve the feature library; if the problem is caused by data acquisition errors, the data acquisition equipment can be checked and calibrated to improve the accuracy of data acquisition. Through continuous feedback and improvement, gradually improve the performance and accuracy of the load decomposition model.

[0155] Next, according to the evaluation results, adjust the parameters of the load recognition model or adopt other optimization strategies. Adjusting the parameters of the load recognition model includes adjusting the learning rate, weight decay coefficient, and the number of neurons in the hidden layer. The adoption of other optimization strategies includes increasing training samples, data augmentation, and feature engineering optimization.

[0156] (1) Adjust the parameters of the neural network model

[0157] Learning rate adjustment: The learning rate is a key hyperparameter in neural network training, which determines the step size of parameter updates during model training. If the evaluation results show that the model converges slowly and the loss function does not decrease significantly, the learning rate can be appropriately increased to accelerate the parameter update speed; conversely, if the model shows oscillations or non-convergence, the learning rate should be decreased to make the model training more stable. For example, use a dynamic learning rate adjustment strategy, such as exponential decay learning rate. As the number of training epochs increases, the learning rate gradually decreases according to an exponential law, which can not only ensure fast convergence in the early stage but also avoid oscillations near the optimal solution in the later stage.

[0158] Weight decay coefficient adjustment: Weight decay (L2 regularization) is used to prevent the model from overfitting. By imposing a penalty term on the weight parameters, the weight values of the model will not be too large. If it is evaluated that the model performs well on the training set but has poor generalization ability on the test set and shows overfitting, the weight decay coefficient can be appropriately increased to enhance the regularization effect; if the model is underfitting and insufficiently learns complex data features, the weight decay coefficient can be decreased.

[0159] Number of neurons in the hidden layer adjustment: The number of neurons in the hidden layer affects the learning ability and expressive ability of the model. If the evaluation results show that the model has insufficient ability to extract complex load features and cannot accurately identify the load type, the number of neurons in the hidden layer can be appropriately increased to improve the fitting ability of the model; if the model shows overfitting and the computing resources are limited, the number of neurons in the hidden layer can be tried to be reduced to simplify the model structure.

[0160] (2) Other optimization strategies

[0161] Increase training samples: If the evaluation results show that the model performs poorly under certain specific load types or working conditions, it may be due to insufficient training samples. By collecting more actual operation data, especially for the scenarios where the model performs poorly, increasing the corresponding number of training samples and enriching the diversity of the samples, the model can learn more load features in different situations and improve the generalization ability and adaptability of the model.

[0162] Data augmentation: For the existing training data, data augmentation techniques are used to expand the dataset. For example, operations such as translation, scaling, and adding noise are performed on the load data to generate new training samples. In the frequency domain features, small perturbations can be made to the coefficients after Fourier transform; in the time domain features, small time offsets or amplitude scalings are made to the waveform, enabling the model to learn more possibilities of data changes and enhancing the robustness of the model.

[0163] Feature engineering optimization: Reexamine the feature extraction process and optimize the feature engineering according to the evaluation results and power system knowledge. Remove redundant features that contribute less to the model performance to avoid overfitting of the model and waste of computing resources caused by excessive features; at the same time, try to extract new effective features, such as features based on load change trends and correlation features with other relevant power parameters, to improve the representational ability of the model for load data.

[0164] In this embodiment, step S8 performs load identification on the collected real-time load data and conducts multi-dimensional analysis of users' electricity consumption behaviors, specifically including: time dimension analysis, device dimension analysis, and user group dimension analysis; among which, the time dimension analysis includes daily electricity consumption patterns, intra-week electricity consumption differences, monthly and annual electricity consumption trends; the device dimension analysis includes the proportion of electricity consumption of devices, the usage frequency and duration of devices, and the combined usage patterns of devices; the user group dimension analysis includes the electricity consumption differences among users with different occupations, the electricity consumption differences due to family size, and the electricity consumption differences in different regions. Specifically:

[0165] (1) Time dimension analysis

[0166] Daily electricity consumption patterns: Taking hours as the unit, the electricity consumption of users in 24 hours a day is refined and statistically analyzed. With the help of high-frequency data collected by smart meters or non-intrusive measurement devices, a detailed electricity consumption-time curve is plotted. For example, through analysis, it is found that from 6 to 8 am on weekdays, kitchen appliances (such as coffee makers, toasters) and lighting equipment in working-class families are intensively used, and the electricity consumption increases significantly; from 7 to 10 pm, residents' entertainment, cooking, cleaning and other activities increase, and devices such as TVs, microwaves, and washing machines run simultaneously, forming the daily electricity consumption peak. While from 2 to 5 am, most devices are in the shutdown or low-power standby state, and the electricity consumption is at a low valley. Mastering the daily electricity consumption patterns helps the power department formulate time-of-use electricity price policies, guide users to consume electricity during low-demand periods, and achieve the optimal allocation of power resources.

[0167] Intra-week electricity consumption differences: By comparing the electricity consumption data of users on different days of a week, the electricity consumption characteristics on weekdays, weekends, and holidays are analyzed. Generally speaking, on weekdays, residents go out to work or study, and the electricity consumption at home during the day is relatively low; on weekends, residents' time at home increases, and the demand for leisure, entertainment, and living electricity consumption rises, such as the usage frequency and duration of devices such as air conditioners, TVs, and computers increase significantly. For commercial users, weekends and holidays are often peak business periods, and the electricity consumption of lighting, heating, cooling, and electrical equipment in places such as shopping malls and restaurants increases significantly. Understanding the intra-week electricity consumption differences can help power companies reasonably arrange power generation plans and grid dispatching, and improve the reliability and stability of power supply.

[0168] Monthly and annual electricity consumption trends: The electricity consumption of users throughout the year is statistically analyzed month by month to analyze the impact of factors such as seasons, temperature, and holidays on electricity consumption. During the high temperature in summer, the demand for air conditioning for cooling significantly increases electricity consumption. Especially in months when the temperature continuously exceeds 30°C, the proportion of air conditioning electricity consumption in the total electricity consumption may be as high as 40%-60%; in winter, heating equipment in the northern regions (such as electric heaters and air conditioning for heating) and small household appliances for heating in the southern regions (such as hand warmers and small suns) are frequently used, resulting in an increase in electricity consumption. In addition, during long holidays such as the Spring Festival and National Day, the number of residents traveling and family reunion activities increases, and the household electricity consumption pattern will also change. By analyzing monthly and annual electricity consumption trends, the power department can predict electricity demand in advance and reasonably plan the construction and maintenance plans of power facilities.

[0169] (2) Equipment dimension analysis

[0170] Proportion of equipment electricity consumption: Accurately calculate the proportion of various types of electrical equipment in the total electricity consumption. Large household appliances such as air conditioners, electric water heaters, and refrigerators are usually big electricity consumers. For example, in summer, the electricity consumption proportion of air conditioners may be as high as 50%-70%, becoming the main factor affecting electricity consumption; although the power per hour of refrigerators is relatively low due to continuous operation for 24 hours, the cumulative electricity consumption accounts for about 10%-20% of the total electricity. By understanding the proportion of equipment electricity consumption, users can targetedly carry out energy-saving transformations on high-energy-consuming equipment, such as replacing energy-saving air conditioners and optimizing the usage habits of refrigerators, to reduce electricity costs.

[0171] Equipment usage frequency and duration: Detailed statistics of the usage frequency and duration of various types of electrical equipment. Although the power of lighting equipment is small, the usage frequency is high, and the cumulative usage duration per day can reach 6-8 hours; while some seasonal equipment, such as electric heaters and fans, have a low usage frequency but a long single usage duration, and each time they may last for 3-5 hours when used in winter or summer. In addition, although the power of some smart devices (such as smart speakers and smart cameras) is low, due to long-term standby, the cumulative energy consumption cannot be ignored. Analyzing the equipment usage frequency and duration helps to evaluate the actual energy consumption of the equipment and provide a reference basis for users to select energy-saving equipment.

[0172] Equipment combination usage mode: Deeply study the combined usage of different equipment. In special scenarios such as family gatherings and holidays, kitchen appliances (such as ovens, microwaves, and induction cookers), lighting equipment, and entertainment equipment (such as TVs and stereos) may be used simultaneously, forming a specific electricity consumption combination mode. By analyzing the equipment combination usage mode, the power department can more accurately predict the changes in power load and optimize the power grid dispatching strategy; at the same time, it can also provide data support for the development of smart home systems to achieve intelligent linkage and energy-saving control of equipment.

[0173] (3) User group dimension analysis

[0174] Differences in electricity consumption among users of different occupations: Due to different work and life patterns, there are significant differences in the electricity consumption behaviors of users in different occupations. Office workers are not at home for most of the day, and electricity consumption at home mainly concentrates in the morning and evening; while freelancers or retirees spend more time at home, and the electricity consumption time is more evenly distributed. For example, the electricity consumption of office worker families during the day on weekdays only accounts for 20%-30% of the whole day, while the electricity consumption of freelancer families during the day can account for 40%-50%. By analyzing the electricity consumption differences among users of different occupations, the power department can formulate differentiated power services and marketing strategies, such as providing night-time off-peak electricity price packages for office workers and flexible electricity consumption packages for freelancers.

[0175] Relationship between household size and electricity consumption: The size of a household will affect the electricity consumption demand. In large families, due to the large number of people, the usage frequency and duration of various electrical appliances will increase, and the electricity consumption is relatively high; the situation is opposite in small families. In addition, members with special electricity consumption needs in the family, such as patients using medical equipment, will also affect the household electricity consumption pattern. For example, the monthly electricity consumption of a five-person family may be 50%-100% higher than that of a two-person family. Understanding the relationship between household size and electricity consumption helps the power department provide suitable power packages and energy-saving suggestions for different households to meet the personalized needs of users.

[0176] Regional electricity consumption differences: Due to different economic development levels, climate conditions and living habits in different regions, there are regional differences in electricity consumption behaviors. For example, in the northern regions, it is cold in winter, and the electricity consumption demand for heating is large; in the southern regions, it is hot in summer, and the electricity consumption demand for cooling is high. Users in economically developed regions, due to their relatively high living standards, use various electrical equipment more frequently, and the electricity consumption is also relatively high. By analyzing the impact of regional differences on electricity consumption, the power department can implement differentiated power planning and management measures in different regions, such as strengthening the power guarantee for winter heating in the northern regions and optimizing the summer cooling power supply plan in the southern regions.

[0177] As Figure 2 shown, the present invention also proposes an electricity consumption analysis system based on non-intrusive load identification technology, including:

[0178] A data acquisition module 101 that collects load data through non-intrusive measurement switches or smart meters; wherein, the load data includes three-phase voltage signals, three-phase current signals, active power, reactive power, frequency, power factor, angular difference, and harmonic content;

[0179] A data twin module 102 that constructs a digital twin model of the electricity load according to the load data of the user's electrical equipment and the power system architecture, and establishes a two-way mapping relationship between high-frequency real-time data and the digital twin model;

[0180] The data preprocessing module 103 uses a denoising algorithm for high-frequency real-time data characteristics to remove noise signals, ensures the time consistency of high-frequency data from different sensors through a high-precision time synchronization algorithm, and maps the load data to a specific interval using a normalization method applicable to high-frequency data;

[0181] The feature extraction module 104 extracts key features from the preprocessed load data. The key features include frequency-domain features, time-domain features, and statistical features, and converts the original data into a feature vector that can reflect the load characteristics;

[0182] The load recognition model construction module 105 collects a large number of samples of known electrical equipment types and high-frequency load data for data annotation, then conducts supervised training on the model, uses a digital twin model to display the training process in real time, and dynamically optimizes and adjusts the neural network model according to the feedback of the digital twin model;

[0183] The load decomposition and recognition module 106 uses the trained neural network model to predict and decompose new unlabeled load data, and identifies the decomposition results of each load according to the output of the neural network model;

[0184] The load recognition evaluation module 107 compares the decomposition results of each load with a pre-established actual load feature library to evaluate the results of load decomposition, and adjusts the parameters of the load recognition model or adopts other optimization strategies according to the evaluation results;

[0185] The electricity consumption behavior analysis module 108 uses the load recognition model adjusted after comparison with the load feature library as the final load recognition model to perform load recognition on the collected real-time load data and conduct multi-dimensional analysis of the user's electricity consumption behavior.

[0186] The implementation steps of the above-mentioned each module correspond to the steps of the electricity consumption analysis method based on the non-intrusive load recognition technology, which will not be elaborated here.

[0187] In summary, the present invention combines the non-intrusive load recognition technology with the digital twin technology, uses high-frequency data to improve the electricity consumption analysis ability, optimizes the recognition model, improves the recognition accuracy, and can deeply analyze the electricity consumption behavior, specifically reflected in:

[0188] (1) By adopting the non-intrusive load recognition technology, only a small number of sensors are installed at the user's power inlet to collect data such as total current and total voltage to achieve load decomposition and recognition. Compared with the traditional method that requires monitoring devices to be installed on each electrical equipment, the monitoring cost is greatly reduced, and the installation and maintenance processes are simple, reducing the interference to the user's original power consumption system.

[0189] (2) Load identification is carried out based on high-frequency data. High-frequency data carries rich information, and the algorithm can more effectively extract electrical appliance feature information. Combining with the bidirectional mapping relationship of the digital twin model, the operation status feedback of the actual power system is obtained in real time to optimize the neural network model, so as to achieve more accurate electrical appliance identification and improve the accuracy of load identification.

[0190] (3) By establishing a digital twin model, verifying and conducting scenario simulation tests on it, ensuring that the model accurately reflects the actual power system. During the model training process, the digital twin model is used to display the training situation in real time, and the parameters and structure of the neural network model are dynamically adjusted according to the feedback, increasing the training samples and optimizing the feature engineering to improve the generalization ability of the model, the processing ability of complex data and the adaptability to new equipment types.

[0191] (4) Conduct multi-dimensional analysis of users' electricity consumption behavior in terms of time, equipment, and user groups. The time-dimensional analysis helps the power department formulate time-of-use electricity price policies, reasonably arrange power generation plans and grid dispatching; the equipment-dimensional analysis helps users carry out energy-saving renovations and provides data support for the development of smart homes; the user-group-dimensional analysis provides a basis for the power department to implement differential services, planning and management, and realizes the optimal allocation of power resources.

[0192] The above is the preferred implementation manner of the present invention. It should be noted that for those of ordinary skill in the art of this technology, without departing from the principle described in the present invention, several improvements and refinements can still be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A power consumption analysis method based on non-intrusive load identification technology, characterized in that, Including: Data collection, collecting load data through non-invasive measurement switches or smart meters; Using professional 3D modeling software, constructing a digital twin model of the electrical load based on the load data of the user's electrical equipment and the power system architecture, and establishing a two-way mapping relationship between the high-frequency real-time data and the digital twin model; Data preprocessing, using a denoising algorithm for the characteristics of high-frequency real-time data to remove noise signals, ensuring the time consistency of high-frequency data from different sensors through a high-precision time synchronization algorithm, and mapping the load data to a specific interval using a normalization method suitable for high-frequency data; Feature extraction, extracting key features from the preprocessed load data, where the key features include frequency-domain features, time-domain features, and statistical features, and converting the original data into a feature vector that can reflect the load characteristics; Selecting a neural network algorithm to construct a load recognition model. After collecting a large number of samples of known electrical equipment types and high-frequency load data for data annotation, training the model in a supervised manner, using the digital twin model to display the training process in real time, and dynamically optimizing and adjusting the neural network model according to the feedback of the digital twin model; Using the trained neural network model to predict and decompose new unlabeled load data, and identifying the decomposition results of each load according to the output of the neural network model; Comparing the decomposition results of each load with the pre-established actual load feature library, evaluating the results of the load decomposition, and adjusting the parameters of the load recognition model or adopting other optimization strategies according to the evaluation results; Taking the load recognition model adjusted after comparison with the load feature library as the final load recognition model, performing load recognition on the collected real-time load data, and conducting multi-dimensional analysis of the user's electricity consumption behavior.

2. The power consumption analysis method based on non-intrusive load identification technology according to claim 1, characterized in that, The construction of the digital twin model of the electrical load and the establishment of a two-way mapping relationship between the high-frequency real-time data and the digital twin model are specifically as follows: Establishing a mathematical model of the electrical load, where the electrical load includes resistive loads, motor loads, and power electronic loads; Establishing a physical simulation model of the digital twin, and establishing a numerical model based on the principles and equations of the electrical load; Integrating data and models, integrating the actual operation data of the electrical load with the physical simulation model to form a complete digital twin model; Verifying the established digital twin model, and conducting scenario simulation and simulation testing.

3. The power consumption analysis method based on non-intrusive load identification technology according to claim 1, characterized in that The extraction of key features from the preprocessed load data, where the key features include frequency-domain features, time-domain features, and statistical features, is specifically to combine the steady-state features of active power, reactive power, voltage, and current, and then add the active power transient waveform for feature extraction, where: For frequency-domain features, the Fast Fourier Transform (FFT) captures the time-varying frequency components in the load data; in terms of time-domain feature extraction, calculation methods of basic statistics such as mean, variance, or crest factor are used; for the extraction of waveform features, a morphology-based algorithm is used; for statistical features, algorithms such as maximum value, minimum value, or standard deviation are used; Dividing the preprocessed continuous load data into frames of a fixed length, where each frame contains a certain number of sampling points. The frame length is selected according to the change frequency of the data and the calculation efficiency to set the sampling points, and window processing is performed on the framed data; The corresponding algorithms are used to calculate the frequency-domain features, time-domain features, and statistical features respectively.

4. The power consumption analysis method based on non-intrusive load identification technology according to claim 1, wherein The neural network algorithm is selected to construct the load recognition model. Specifically, a bidirectional long short-term memory network (BiLSTM) is used to construct the load recognition model.

5. The power consumption analysis method based on non-intrusive load identification technology according to claim 1, wherein The digital twin model is used to display the training process in real time. According to the feedback of the digital twin model, the neural network model is dynamically optimized and adjusted. Specifically: A data docking and communication mechanism is established, and a special data interface program is developed to realize the data interaction between the digital twin model and the neural network training environment. The training process is displayed in the digital twin model. In the 3D visualization interface of the digital twin model, a special area is designed to display the neural network training process, and the data during the neural network training process is mapped to the 3D scene of the digital twin model. During the operation of the digital twin model, the state changes of power system components and the occurrence of simulated faults are monitored in real time, and the feedback information from different sources is integrated to form a unified feedback data set. The actual operating state of the power system feedback by the digital twin model is compared with the prediction results of the neural network to evaluate the training effect of the neural network. According to the analysis of the feedback information and training effect, the parameters of the neural network are dynamically adjusted, the structure is optimized, the training samples are increased, and the feature engineering is optimized.

6. The power consumption analysis method based on non-intrusive load identification technology according to claim 1, characterized in that, The trained neural network model is used to predict and decompose new unlabeled load data. According to the output of the neural network model, the decomposition results of each load are identified. Specifically: New unlabeled load data is continuously collected through non-intrusive measurement switches or smart meters at a predetermined high-frequency sampling rate. Preprocessing is performed on the collected new unlabeled load data, including but not limited to data cleaning and data normalization. According to the feature extraction method determined during model training, key features are extracted from the preprocessed new data, including frequency-domain features, time-domain features, and statistical features, and the extracted features are organized into a format acceptable to the model. Based on the input feature data, the model uses the patterns and parameters learned during the training process to perform forward propagation calculations and outputs prediction results represented in the form of probability distributions or numerical values. According to the prediction results output by the model, they are mapped to the actual electrical equipment types and load parameters, and the initially identified load decomposition results are verified, and judgments are made in combination with the physical simulation results or the operating common sense of the actual power system in the digital twin model.

7. The power consumption analysis method based on non-intrusive load identification technology according to claim 1, characterized in that, The decomposition results of each load are compared with the pre-established actual load feature library to evaluate the results of load decomposition. Specifically: From the output of the load decomposition model, the decomposition result data of each load is extracted. The decomposition result data includes information such as the electrical equipment type corresponding to the load, relevant electrical parameters, and the confidence level predicted by the model. The pre-established actual load feature library is accessed. The feature library is stored using a relational database or a non-relational database. According to the electrical equipment type involved in the decomposition results, the load feature data of the corresponding equipment is read from the feature library, including the rated parameters of the equipment, the typical operating parameter range, and the characteristic curves under different working conditions. Match the types of electrical equipment in the load decomposition result with the equipment types in the actual load characteristic library, including electrical parameter comparison and characteristic curve comparison; Calculate various error indicators based on the results of electrical parameter comparison and characteristic curve comparison, calculate the accuracy rate of load decomposition according to the result of equipment type matching, and evaluate the load decomposition result in combination with operation reliability.

8. The power consumption analysis method based on non-intrusive load identification technology according to claim 1, wherein According to the evaluation result, adjust the parameters of the load recognition model or adopt other optimization strategies. Specifically, adjusting the parameters of the load recognition model includes adjusting the learning rate, weight decay coefficient, and the number of neurons in the hidden layer, and adopting other optimization strategies includes increasing training samples, data augmentation, and feature engineering optimization.

9. The power consumption analysis method based on non-intrusive load identification technology according to claim 1, wherein Perform load recognition on the collected real-time load data and conduct multi-dimensional analysis of user electricity consumption behavior, specifically including: time dimension analysis, equipment dimension analysis, and user group dimension analysis; among them, time dimension analysis includes daily electricity consumption patterns, intra-week electricity consumption differences, monthly and annual electricity consumption trends; equipment dimension analysis includes equipment electricity consumption ratio, equipment usage frequency and duration, and equipment combined usage patterns; user group dimension analysis includes electricity consumption differences among users of different occupations, household size and electricity consumption differences, and regional electricity consumption differences.

10. An electricity consumption analysis system based on non-intrusive load identification technology, according to the electricity consumption analysis method based on non-intrusive load identification technology described in any one of claims 1-9, characterized in that, The system includes: A data acquisition module that collects load data through non-intrusive measurement switches or smart meters; A data twin module that constructs a digital twin model of the electricity load based on the load data of the user's electrical equipment and the power system architecture, and establishes a two-way mapping relationship between the high-frequency real-time data and the digital twin model; A data preprocessing module that uses a denoising algorithm for high-frequency real-time data characteristics to remove noise signals, ensures the time consistency of high-frequency data from different sensors through a high-precision time synchronization algorithm, and maps the load data to a specific interval using a normalization method suitable for high-frequency data; A feature extraction module that extracts key features from the preprocessed load data. The key features include frequency domain features, time domain features, and statistical features, and converts the original data into a feature vector that can reflect the load characteristics; A load recognition model construction module that collects a large number of samples of known electrical equipment types and high-frequency load data for data annotation, then conducts supervised training on the model, uses the digital twin model to display the training process in real time, and dynamically optimizes and adjusts the neural network model according to the feedback of the digital twin model; A load decomposition and recognition module that uses the trained neural network model to predict and decompose new unlabeled load data, and identifies the decomposition result of each load according to the output of the neural network model; A load recognition evaluation module that compares the decomposition result of each load with the pre-established actual load characteristic library, evaluates the result of load decomposition, and adjusts the parameters of the load recognition model or adopts other optimization strategies according to the evaluation result; An electricity consumption behavior analysis module that uses the adjusted load recognition model after comparison with the load characteristic library as the final load recognition model, performs load recognition on the collected real-time load data, and conducts multi-dimensional analysis of user electricity consumption behavior.

Images